phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-31 07:39:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge-related changes.

Browse source
This commit is contained in:
ilya-kolchinsky 2025-04-02 19:56:44 +02:00
commit 60e9f46856
456 changed files with 38636 additions and 10892 deletions
Download patch file Download diff file Expand all files Collapse all files

606
llama_stack/providers/inline/agents/meta_reference/agent_instance.py
View file

@ -6,14 +6,12 @@
import copy
import json
import os
import re
import secrets
import string
import uuid
from datetime import datetime
from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple, Union
from urllib.parse import urlparse
from datetime import datetime, timezone
from typing import AsyncGenerator, List, Optional, Union
import httpx
@ -59,12 +57,7 @@ from llama_stack.apis.inference import (
UserMessage,
)
from llama_stack.apis.safety import Safety
from llama_stack.apis.tools import (
RAGDocument,
ToolGroups,
ToolInvocationResult,
ToolRuntime,
)
from llama_stack.apis.tools import ToolGroups, ToolInvocationResult, ToolRuntime
from llama_stack.apis.vector_io import VectorIO
from llama_stack.log import get_logger
from llama_stack.models.llama.datatypes import (
@ -153,7 +146,6 @@ class ChatAgent(ShieldRunnerMixin):
messages.append(
ToolResponseMessage(
call_id=response.call_id,
tool_name=response.tool_name,
content=response.content,
)
)
@ -181,23 +173,29 @@ class ChatAgent(ShieldRunnerMixin):
return messages
async def create_and_execute_turn(self, request: AgentTurnCreateRequest) -> AsyncGenerator:
async with tracing.span("create_and_execute_turn") as span:
span = tracing.get_current_span()
if span:
span.set_attribute("session_id", request.session_id)
span.set_attribute("agent_id", self.agent_id)
span.set_attribute("request", request.model_dump_json())
turn_id = str(uuid.uuid4())
span.set_attribute("turn_id", turn_id)
async for chunk in self._run_turn(request, turn_id):
yield chunk
await self._initialize_tools(request.toolgroups)
async for chunk in self._run_turn(request, turn_id):
yield chunk
async def resume_turn(self, request: AgentTurnResumeRequest) -> AsyncGenerator:
async with tracing.span("resume_turn") as span:
span = tracing.get_current_span()
if span:
span.set_attribute("agent_id", self.agent_id)
span.set_attribute("session_id", request.session_id)
span.set_attribute("turn_id", request.turn_id)
span.set_attribute("request", request.model_dump_json())
async for chunk in self._run_turn(request):
yield chunk
span.set_attribute("turn_id", request.turn_id)
await self._initialize_tools()
async for chunk in self._run_turn(request):
yield chunk
async def _run_turn(
self,
@ -218,18 +216,9 @@ class ChatAgent(ShieldRunnerMixin):
steps = []
messages = await self.get_messages_from_turns(turns)
if is_resume:
if isinstance(request.tool_responses[0], ToolResponseMessage):
tool_response_messages = request.tool_responses
tool_responses = [
ToolResponse(call_id=x.call_id, tool_name=x.tool_name, content=x.content)
for x in request.tool_responses
]
else:
tool_response_messages = [
ToolResponseMessage(call_id=x.call_id, tool_name=x.tool_name, content=x.content)
for x in request.tool_responses
]
tool_responses = request.tool_responses
tool_response_messages = [
ToolResponseMessage(call_id=x.call_id, content=x.content) for x in request.tool_responses
]
messages.extend(tool_response_messages)
last_turn = turns[-1]
last_turn_messages = self.turn_to_messages(last_turn)
@ -247,12 +236,12 @@ class ChatAgent(ShieldRunnerMixin):
in_progress_tool_call_step = await self.storage.get_in_progress_tool_call_step(
request.session_id, request.turn_id
)
now = datetime.now().astimezone().isoformat()
now = datetime.now(timezone.utc).isoformat()
tool_execution_step = ToolExecutionStep(
step_id=(in_progress_tool_call_step.step_id if in_progress_tool_call_step else str(uuid.uuid4())),
turn_id=request.turn_id,
tool_calls=(in_progress_tool_call_step.tool_calls if in_progress_tool_call_step else []),
tool_responses=tool_responses,
tool_responses=request.tool_responses,
completed_at=now,
started_at=(in_progress_tool_call_step.started_at if in_progress_tool_call_step else now),
)
@ -272,7 +261,7 @@ class ChatAgent(ShieldRunnerMixin):
start_time = last_turn.started_at
else:
messages.extend(request.messages)
start_time = datetime.now().astimezone().isoformat()
start_time = datetime.now(timezone.utc).isoformat()
input_messages = request.messages
output_message = None
@ -283,7 +272,6 @@ class ChatAgent(ShieldRunnerMixin):
sampling_params=self.agent_config.sampling_params,
stream=request.stream,
documents=request.documents if not is_resume else None,
toolgroups_for_turn=request.toolgroups if not is_resume else None,
):
if isinstance(chunk, CompletionMessage):
output_message = chunk
@ -304,7 +292,7 @@ class ChatAgent(ShieldRunnerMixin):
input_messages=input_messages,
output_message=output_message,
started_at=start_time,
completed_at=datetime.now().astimezone().isoformat(),
completed_at=datetime.now(timezone.utc).isoformat(),
steps=steps,
)
await self.storage.add_turn_to_session(request.session_id, turn)
@ -335,7 +323,6 @@ class ChatAgent(ShieldRunnerMixin):
sampling_params: SamplingParams,
stream: bool = False,
documents: Optional[List[Document]] = None,
toolgroups_for_turn: Optional[List[AgentToolGroup]] = None,
) -> AsyncGenerator:
# Doing async generators makes downstream code much simpler and everything amenable to
# streaming. However, it also makes things complicated here because AsyncGenerators cannot
@ -358,7 +345,6 @@ class ChatAgent(ShieldRunnerMixin):
sampling_params,
stream,
documents,
toolgroups_for_turn,
):
if isinstance(res, bool):
return
@ -397,7 +383,7 @@ class ChatAgent(ShieldRunnerMixin):
return
step_id = str(uuid.uuid4())
shield_call_start_time = datetime.now().astimezone().isoformat()
shield_call_start_time = datetime.now(timezone.utc).isoformat()
try:
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
@ -421,7 +407,7 @@ class ChatAgent(ShieldRunnerMixin):
turn_id=turn_id,
violation=e.violation,
started_at=shield_call_start_time,
completed_at=datetime.now().astimezone().isoformat(),
completed_at=datetime.now(timezone.utc).isoformat(),
),
)
)
@ -444,7 +430,7 @@ class ChatAgent(ShieldRunnerMixin):
turn_id=turn_id,
violation=None,
started_at=shield_call_start_time,
completed_at=datetime.now().astimezone().isoformat(),
completed_at=datetime.now(timezone.utc).isoformat(),
),
)
)
@ -459,30 +445,35 @@ class ChatAgent(ShieldRunnerMixin):
sampling_params: SamplingParams,
stream: bool = False,
documents: Optional[List[Document]] = None,
toolgroups_for_turn: Optional[List[AgentToolGroup]] = None,
) -> AsyncGenerator:
# TODO: simplify all of this code, it can be simpler
toolgroup_args = {}
toolgroups = set()
for toolgroup in self.agent_config.toolgroups + (toolgroups_for_turn or []):
if isinstance(toolgroup, AgentToolGroupWithArgs):
tool_group_name, tool_name = self._parse_toolgroup_name(toolgroup.name)
toolgroups.add(tool_group_name)
toolgroup_args[tool_group_name] = toolgroup.args
else:
toolgroups.add(toolgroup)
tool_defs, tool_to_group = await self._get_tool_defs(toolgroups_for_turn)
# if document is passed in a turn, we parse the raw text of the document
# and sent it as a user message
if documents:
await self.handle_documents(session_id, documents, input_messages, tool_defs)
contexts = []
for document in documents:
raw_document_text = await get_raw_document_text(document)
contexts.append(raw_document_text)
attached_context = "\n".join(contexts)
if isinstance(input_messages[-1].content, str):
input_messages[-1].content += attached_context
elif isinstance(input_messages[-1].content, list):
input_messages[-1].content.append(TextContentItem(text=attached_context))
else:
input_messages[-1].content = [
input_messages[-1].content,
TextContentItem(text=attached_context),
]
session_info = await self.storage.get_session_info(session_id)
# if the session has a memory bank id, let the memory tool use it
if session_info and session_info.vector_db_id:
if RAG_TOOL_GROUP not in toolgroup_args:
toolgroup_args[RAG_TOOL_GROUP] = {"vector_db_ids": [session_info.vector_db_id]}
else:
toolgroup_args[RAG_TOOL_GROUP]["vector_db_ids"].append(session_info.vector_db_id)
for tool_name in self.tool_name_to_args.keys():
if tool_name == MEMORY_QUERY_TOOL:
if "vector_db_ids" not in self.tool_name_to_args[tool_name]:
self.tool_name_to_args[tool_name]["vector_db_ids"] = [session_info.vector_db_id]
else:
self.tool_name_to_args[tool_name]["vector_db_ids"].append(session_info.vector_db_id)
output_attachments = []
@ -494,7 +485,7 @@ class ChatAgent(ShieldRunnerMixin):
client_tools[tool.name] = tool
while True:
step_id = str(uuid.uuid4())
inference_start_time = datetime.now().astimezone().isoformat()
inference_start_time = datetime.now(timezone.utc).isoformat()
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepStartPayload(
@ -512,7 +503,7 @@ class ChatAgent(ShieldRunnerMixin):
async for chunk in await self.inference_api.chat_completion(
self.agent_config.model,
input_messages,
tools=tool_defs,
tools=self.tool_defs,
tool_prompt_format=self.agent_config.tool_config.tool_prompt_format,
response_format=self.agent_config.response_format,
stream=True,
@ -604,7 +595,7 @@ class ChatAgent(ShieldRunnerMixin):
turn_id=turn_id,
model_response=copy.deepcopy(message),
started_at=inference_start_time,
completed_at=datetime.now().astimezone().isoformat(),
completed_at=datetime.now(timezone.utc).isoformat(),
),
)
)
@ -636,125 +627,143 @@ class ChatAgent(ShieldRunnerMixin):
logger.debug(f"completion message with EOM (iter: {n_iter}): {str(message)}")
input_messages = input_messages + [message]
else:
logger.debug(f"completion message (iter: {n_iter}) from the model: {str(message)}")
# 1. Start the tool execution step and progress
step_id = str(uuid.uuid4())
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepStartPayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
)
)
)
tool_call = message.tool_calls[0]
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepProgressPayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
tool_call=tool_call,
delta=ToolCallDelta(
parse_status=ToolCallParseStatus.in_progress,
tool_call=tool_call,
),
)
)
)
input_messages = input_messages + [message]
# If tool is a client tool, yield CompletionMessage and return
if tool_call.tool_name in client_tools:
# NOTE: mark end_of_message to indicate to client that it may
# call the tool and continue the conversation with the tool's response.
message.stop_reason = StopReason.end_of_message
# Process tool calls in the message
client_tool_calls = []
non_client_tool_calls = []
# Separate client and non-client tool calls
for tool_call in message.tool_calls:
if tool_call.tool_name in client_tools:
client_tool_calls.append(tool_call)
else:
non_client_tool_calls.append(tool_call)
# Process non-client tool calls first
for tool_call in non_client_tool_calls:
step_id = str(uuid.uuid4())
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepStartPayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
)
)
)
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepProgressPayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
delta=ToolCallDelta(
parse_status=ToolCallParseStatus.in_progress,
tool_call=tool_call,
),
)
)
)
# Execute the tool call
async with tracing.span(
"tool_execution",
{
"tool_name": tool_call.tool_name,
"input": message.model_dump_json(),
},
) as span:
tool_execution_start_time = datetime.now(timezone.utc).isoformat()
tool_result = await self.execute_tool_call_maybe(
session_id,
tool_call,
)
if tool_result.content is None:
raise ValueError(
f"Tool call result (id: {tool_call.call_id}, name: {tool_call.tool_name}) does not have any content"
)
result_message = ToolResponseMessage(
call_id=tool_call.call_id,
content=tool_result.content,
)
span.set_attribute("output", result_message.model_dump_json())
# Store tool execution step
tool_execution_step = ToolExecutionStep(
step_id=step_id,
turn_id=turn_id,
tool_calls=[tool_call],
tool_responses=[
ToolResponse(
call_id=tool_call.call_id,
tool_name=tool_call.tool_name,
content=tool_result.content,
metadata=tool_result.metadata,
)
],
started_at=tool_execution_start_time,
completed_at=datetime.now(timezone.utc).isoformat(),
)
# Yield the step completion event
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepCompletePayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
step_details=tool_execution_step,
)
)
)
# Add the result message to input_messages for the next iteration
input_messages.append(result_message)
# TODO: add tool-input touchpoint and a "start" event for this step also
# but that needs a lot more refactoring of Tool code potentially
if (type(result_message.content) is str) and (
out_attachment := _interpret_content_as_attachment(result_message.content)
):
# NOTE: when we push this message back to the model, the model may ignore the
# attached file path etc. since the model is trained to only provide a user message
# with the summary. We keep all generated attachments and then attach them to final message
output_attachments.append(out_attachment)
# If there are client tool calls, yield a message with only those tool calls
if client_tool_calls:
await self.storage.set_in_progress_tool_call_step(
session_id,
turn_id,
ToolExecutionStep(
step_id=step_id,
turn_id=turn_id,
tool_calls=[tool_call],
tool_calls=client_tool_calls,
tool_responses=[],
started_at=datetime.now().astimezone().isoformat(),
started_at=datetime.now(timezone.utc).isoformat(),
),
)
yield message
# Create a copy of the message with only client tool calls
client_message = message.model_copy(deep=True)
client_message.tool_calls = client_tool_calls
# NOTE: mark end_of_message to indicate to client that it may
# call the tool and continue the conversation with the tool's response.
client_message.stop_reason = StopReason.end_of_message
# Yield the message with client tool calls
yield client_message
return
# If tool is a builtin server tool, execute it
tool_name = tool_call.tool_name
if isinstance(tool_name, BuiltinTool):
tool_name = tool_name.value
async with tracing.span(
"tool_execution",
{
"tool_name": tool_name,
"input": message.model_dump_json(),
},
) as span:
tool_execution_start_time = datetime.now().astimezone().isoformat()
tool_call = message.tool_calls[0]
tool_result = await execute_tool_call_maybe(
self.tool_runtime_api,
session_id,
tool_call,
toolgroup_args,
tool_to_group,
)
if tool_result.content is None:
raise ValueError(
f"Tool call result (id: {tool_call.call_id}, name: {tool_call.tool_name}) does not have any content"
)
result_messages = [
ToolResponseMessage(
call_id=tool_call.call_id,
tool_name=tool_call.tool_name,
content=tool_result.content,
)
]
assert len(result_messages) == 1, "Currently not supporting multiple messages"
result_message = result_messages[0]
span.set_attribute("output", result_message.model_dump_json())
async def _initialize_tools(
self,
toolgroups_for_turn: Optional[List[AgentToolGroup]] = None,
) -> None:
toolgroup_to_args = {}
for toolgroup in (self.agent_config.toolgroups or []) + (toolgroups_for_turn or []):
if isinstance(toolgroup, AgentToolGroupWithArgs):
tool_group_name, _ = self._parse_toolgroup_name(toolgroup.name)
toolgroup_to_args[tool_group_name] = toolgroup.args
yield AgentTurnResponseStreamChunk(
event=AgentTurnResponseEvent(
payload=AgentTurnResponseStepCompletePayload(
step_type=StepType.tool_execution.value,
step_id=step_id,
step_details=ToolExecutionStep(
step_id=step_id,
turn_id=turn_id,
tool_calls=[tool_call],
tool_responses=[
ToolResponse(
call_id=result_message.call_id,
tool_name=result_message.tool_name,
content=result_message.content,
metadata=tool_result.metadata,
)
],
started_at=tool_execution_start_time,
completed_at=datetime.now().astimezone().isoformat(),
),
)
)
)
# TODO: add tool-input touchpoint and a "start" event for this step also
# but that needs a lot more refactoring of Tool code potentially
if (type(result_message.content) is str) and (
out_attachment := _interpret_content_as_attachment(result_message.content)
):
# NOTE: when we push this message back to the model, the model may ignore the
# attached file path etc. since the model is trained to only provide a user message
# with the summary. We keep all generated attachments and then attach them to final message
output_attachments.append(out_attachment)
input_messages = input_messages + [message, result_message]
async def _get_tool_defs(
self, toolgroups_for_turn: Optional[List[AgentToolGroup]] = None
) -> Tuple[List[ToolDefinition], Dict[str, str]]:
# Determine which tools to include
tool_groups_to_include = toolgroups_for_turn or self.agent_config.toolgroups or []
agent_config_toolgroups = []
@ -763,8 +772,10 @@ class ChatAgent(ShieldRunnerMixin):
if name not in agent_config_toolgroups:
agent_config_toolgroups.append(name)
toolgroup_to_args = toolgroup_to_args or {}
tool_name_to_def = {}
tool_to_group = {}
tool_name_to_args = {}
for tool_def in self.agent_config.client_tools:
if tool_name_to_def.get(tool_def.name, None):
@ -782,53 +793,38 @@ class ChatAgent(ShieldRunnerMixin):
for param in tool_def.parameters
},
)
tool_to_group[tool_def.name] = "__client_tools__"
for toolgroup_name_with_maybe_tool_name in agent_config_toolgroups:
toolgroup_name, tool_name = self._parse_toolgroup_name(toolgroup_name_with_maybe_tool_name)
toolgroup_name, input_tool_name = self._parse_toolgroup_name(toolgroup_name_with_maybe_tool_name)
tools = await self.tool_groups_api.list_tools(toolgroup_id=toolgroup_name)
if not tools.data:
available_tool_groups = ", ".join(
[t.identifier for t in (await self.tool_groups_api.list_tool_groups()).data]
)
raise ValueError(f"Toolgroup {toolgroup_name} not found, available toolgroups: {available_tool_groups}")
if tool_name is not None and not any(tool.identifier == tool_name for tool in tools.data):
if input_tool_name is not None and not any(tool.identifier == input_tool_name for tool in tools.data):
raise ValueError(
f"Tool {tool_name} not found in toolgroup {toolgroup_name}. Available tools: {', '.join([tool.identifier for tool in tools.data])}"
f"Tool {input_tool_name} not found in toolgroup {toolgroup_name}. Available tools: {', '.join([tool.identifier for tool in tools.data])}"
)
for tool_def in tools.data:
if toolgroup_name.startswith("builtin") and toolgroup_name != RAG_TOOL_GROUP:
tool_name = tool_def.identifier
built_in_type = BuiltinTool.brave_search
if tool_name == "web_search":
built_in_type = BuiltinTool.brave_search
identifier: str | BuiltinTool | None = tool_def.identifier
if identifier == "web_search":
identifier = BuiltinTool.brave_search
else:
built_in_type = BuiltinTool(tool_name)
identifier = BuiltinTool(identifier)
else:
# add if tool_name is unspecified or the tool_def identifier is the same as the tool_name
if input_tool_name in (None, tool_def.identifier):
identifier = tool_def.identifier
else:
identifier = None
if tool_name_to_def.get(built_in_type, None):
raise ValueError(f"Tool {built_in_type} already exists")
tool_name_to_def[built_in_type] = ToolDefinition(
tool_name=built_in_type,
description=tool_def.description,
parameters={
param.name: ToolParamDefinition(
param_type=param.parameter_type,
description=param.description,
required=param.required,
default=param.default,
)
for param in tool_def.parameters
},
)
tool_to_group[built_in_type] = tool_def.toolgroup_id
continue
if tool_name_to_def.get(tool_def.identifier, None):
raise ValueError(f"Tool {tool_def.identifier} already exists")
if tool_name in (None, tool_def.identifier):
if tool_name_to_def.get(identifier, None):
raise ValueError(f"Tool {identifier} already exists")
if identifier:
tool_name_to_def[tool_def.identifier] = ToolDefinition(
tool_name=tool_def.identifier,
tool_name=identifier,
description=tool_def.description,
parameters={
param.name: ToolParamDefinition(
@ -840,9 +836,12 @@ class ChatAgent(ShieldRunnerMixin):
for param in tool_def.parameters
},
)
tool_to_group[tool_def.identifier] = tool_def.toolgroup_id
tool_name_to_args[tool_def.identifier] = toolgroup_to_args.get(toolgroup_name, {})
return list(tool_name_to_def.values()), tool_to_group
self.tool_defs, self.tool_name_to_args = (
list(tool_name_to_def.values()),
tool_name_to_args,
)
def _parse_toolgroup_name(self, toolgroup_name_with_maybe_tool_name: str) -> tuple[str, Optional[str]]:
"""Parse a toolgroup name into its components.
@ -861,176 +860,59 @@ class ChatAgent(ShieldRunnerMixin):
tool_group, tool_name = split_names[0], None
return tool_group, tool_name
async def handle_documents(
async def execute_tool_call_maybe(
self,
session_id: str,
documents: List[Document],
input_messages: List[Message],
tool_defs: Dict[str, ToolDefinition],
) -> None:
memory_tool = any(tool_def.tool_name == MEMORY_QUERY_TOOL for tool_def in tool_defs)
code_interpreter_tool = any(tool_def.tool_name == BuiltinTool.code_interpreter for tool_def in tool_defs)
content_items = []
url_items = []
pattern = re.compile("^(https?://|file://|data:)")
for d in documents:
if isinstance(d.content, URL):
url_items.append(d.content)
elif pattern.match(d.content):
url_items.append(URL(uri=d.content))
tool_call: ToolCall,
) -> ToolInvocationResult:
tool_name = tool_call.tool_name
registered_tool_names = [tool_def.tool_name for tool_def in self.tool_defs]
if tool_name not in registered_tool_names:
raise ValueError(
f"Tool {tool_name} not found in provided tools, registered tools: {', '.join([str(x) for x in registered_tool_names])}"
)
if isinstance(tool_name, BuiltinTool):
if tool_name == BuiltinTool.brave_search:
tool_name_str = WEB_SEARCH_TOOL
else:
content_items.append(d)
# Save the contents to a tempdir and use its path as a URL if code interpreter is present
if code_interpreter_tool:
for c in content_items:
temp_file_path = os.path.join(self.tempdir, f"{make_random_string()}.txt")
with open(temp_file_path, "w") as temp_file:
temp_file.write(c.content)
url_items.append(URL(uri=f"file://{temp_file_path}"))
if memory_tool and code_interpreter_tool:
# if both memory and code_interpreter are available, we download the URLs
# and attach the data to the last message.
msg = await attachment_message(self.tempdir, url_items)
input_messages.append(msg)
# Since memory is present, add all the data to the memory bank
await self.add_to_session_vector_db(session_id, documents)
elif code_interpreter_tool:
# if only code_interpreter is available, we download the URLs to a tempdir
# and attach the path to them as a message to inference with the
# assumption that the model invokes the code_interpreter tool with the path
msg = await attachment_message(self.tempdir, url_items)
input_messages.append(msg)
elif memory_tool:
# if only memory is available, we load the data from the URLs and content items to the memory bank
await self.add_to_session_vector_db(session_id, documents)
tool_name_str = tool_name.value
else:
# if no memory or code_interpreter tool is available,
# we try to load the data from the URLs and content items as a message to inference
# and add it to the last message's context
input_messages[-1].context = "\n".join(
[doc.content for doc in content_items] + await load_data_from_urls(url_items)
)
tool_name_str = tool_name
async def _ensure_vector_db(self, session_id: str) -> str:
session_info = await self.storage.get_session_info(session_id)
if session_info is None:
raise ValueError(f"Session {session_id} not found")
if session_info.vector_db_id is None:
vector_db_id = f"vector_db_{session_id}"
# TODO: the semantic for registration is definitely not "creation"
# so we need to fix it if we expect the agent to create a new vector db
# for each session
await self.vector_io_api.register_vector_db(
vector_db_id=vector_db_id,
embedding_model="all-MiniLM-L6-v2",
)
await self.storage.add_vector_db_to_session(session_id, vector_db_id)
else:
vector_db_id = session_info.vector_db_id
return vector_db_id
async def add_to_session_vector_db(self, session_id: str, data: List[Document]) -> None:
vector_db_id = await self._ensure_vector_db(session_id)
documents = [
RAGDocument(
document_id=str(uuid.uuid4()),
content=a.content,
mime_type=a.mime_type,
metadata={},
)
for a in data
]
await self.tool_runtime_api.rag_tool.insert(
documents=documents,
vector_db_id=vector_db_id,
chunk_size_in_tokens=512,
logger.info(f"executing tool call: {tool_name_str} with args: {tool_call.arguments}")
result = await self.tool_runtime_api.invoke_tool(
tool_name=tool_name_str,
kwargs={
"session_id": session_id,
# get the arguments generated by the model and augment with toolgroup arg overrides for the agent
**tool_call.arguments,
**self.tool_name_to_args.get(tool_name_str, {}),
},
)
logger.debug(f"tool call {tool_name_str} completed with result: {result}")
return result
async def load_data_from_urls(urls: List[URL]) -> List[str]:
data = []
for url in urls:
uri = url.uri
if uri.startswith("file://"):
filepath = uri[len("file://") :]
with open(filepath, "r") as f:
data.append(f.read())
elif uri.startswith("http"):
async with httpx.AsyncClient() as client:
r = await client.get(uri)
resp = r.text
data.append(resp)
return data
async def load_data_from_url(url: str) -> str:
if url.startswith("http"):
async with httpx.AsyncClient() as client:
r = await client.get(url)
resp = r.text
return resp
raise ValueError(f"Unexpected URL: {type(url)}")
async def attachment_message(tempdir: str, urls: List[URL]) -> ToolResponseMessage:
content = []
for url in urls:
uri = url.uri
if uri.startswith("file://"):
filepath = uri[len("file://") :]
elif uri.startswith("http"):
path = urlparse(uri).path
basename = os.path.basename(path)
filepath = f"{tempdir}/{make_random_string() + basename}"
logger.info(f"Downloading {url} -> {filepath}")
async with httpx.AsyncClient() as client:
r = await client.get(uri)
resp = r.text
with open(filepath, "w") as fp:
fp.write(resp)
else:
raise ValueError(f"Unsupported URL {url}")
content.append(
TextContentItem(
text=f'# User provided a file accessible to you at "{filepath}"\nYou can use code_interpreter to load and inspect it.'
)
)
return ToolResponseMessage(
call_id="",
tool_name=BuiltinTool.code_interpreter,
content=content,
)
async def execute_tool_call_maybe(
tool_runtime_api: ToolRuntime,
session_id: str,
tool_call: ToolCall,
toolgroup_args: Dict[str, Dict[str, Any]],
tool_to_group: Dict[str, str],
) -> ToolInvocationResult:
name = tool_call.tool_name
group_name = tool_to_group.get(name, None)
if group_name is None:
raise ValueError(f"Tool {name} not found in any tool group")
if isinstance(name, BuiltinTool):
if name == BuiltinTool.brave_search:
name = WEB_SEARCH_TOOL
else:
name = name.value
logger.info(f"executing tool call: {name} with args: {tool_call.arguments}")
result = await tool_runtime_api.invoke_tool(
tool_name=name,
kwargs={
"session_id": session_id,
# get the arguments generated by the model and augment with toolgroup arg overrides for the agent
**tool_call.arguments,
**toolgroup_args.get(group_name, {}),
},
)
logger.info(f"tool call {name} completed with result: {result}")
return result
async def get_raw_document_text(document: Document) -> str:
if not document.mime_type.startswith("text/"):
raise ValueError(f"Unexpected document mime type: {document.mime_type}")
if isinstance(document.content, URL):
return await load_data_from_url(document.content.uri)
elif isinstance(document.content, str):
return document.content
elif isinstance(document.content, TextContentItem):
return document.content.text
else:
raise ValueError(f"Unexpected document content type: {type(document.content)}")
def _interpret_content_as_attachment(

2
llama_stack/providers/inline/agents/meta_reference/agents.py
View file

@ -172,7 +172,7 @@ class MetaReferenceAgentsImpl(Agents):
agent_id: str,
session_id: str,
turn_id: str,
tool_responses: Union[List[ToolResponse], List[ToolResponseMessage]],
tool_responses: List[ToolResponse],
stream: Optional[bool] = False,
) -> AsyncGenerator:
request = AgentTurnResumeRequest(

64
llama_stack/providers/inline/agents/meta_reference/persistence.py
View file

@ -7,12 +7,15 @@
import json
import logging
import uuid
from datetime import datetime
from datetime import datetime, timezone
from typing import List, Optional
from pydantic import BaseModel
from llama_stack.apis.agents import ToolExecutionStep, Turn
from llama_stack.distribution.access_control import check_access
from llama_stack.distribution.datatypes import AccessAttributes
from llama_stack.distribution.request_headers import get_auth_attributes
from llama_stack.providers.utils.kvstore import KVStore
log = logging.getLogger(__name__)
@ -24,6 +27,7 @@ class AgentSessionInfo(BaseModel):
# TODO: is this used anywhere?
vector_db_id: Optional[str] = None
started_at: datetime
access_attributes: Optional[AccessAttributes] = None
class AgentPersistence:
@ -33,11 +37,18 @@ class AgentPersistence:
async def create_session(self, name: str) -> str:
session_id = str(uuid.uuid4())
# Get current user's auth attributes for new sessions
auth_attributes = get_auth_attributes()
access_attributes = AccessAttributes(**auth_attributes) if auth_attributes else None
session_info = AgentSessionInfo(
session_id=session_id,
session_name=name,
started_at=datetime.now(),
started_at=datetime.now(timezone.utc),
access_attributes=access_attributes,
)
await self.kvstore.set(
key=f"session:{self.agent_id}:{session_id}",
value=session_info.model_dump_json(),
@ -51,12 +62,34 @@ class AgentPersistence:
if not value:
return None
return AgentSessionInfo(**json.loads(value))
session_info = AgentSessionInfo(**json.loads(value))
# Check access to session
if not self._check_session_access(session_info):
return None
return session_info
def _check_session_access(self, session_info: AgentSessionInfo) -> bool:
"""Check if current user has access to the session."""
# Handle backward compatibility for old sessions without access control
if not hasattr(session_info, "access_attributes"):
return True
return check_access(session_info.session_id, session_info.access_attributes, get_auth_attributes())
async def get_session_if_accessible(self, session_id: str) -> Optional[AgentSessionInfo]:
"""Get session info if the user has access to it. For internal use by sub-session methods."""
session_info = await self.get_session_info(session_id)
if not session_info:
return None
return session_info
async def add_vector_db_to_session(self, session_id: str, vector_db_id: str):
session_info = await self.get_session_info(session_id)
session_info = await self.get_session_if_accessible(session_id)
if session_info is None:
raise ValueError(f"Session {session_id} not found")
raise ValueError(f"Session {session_id} not found or access denied")
session_info.vector_db_id = vector_db_id
await self.kvstore.set(
@ -65,12 +98,18 @@ class AgentPersistence:
)
async def add_turn_to_session(self, session_id: str, turn: Turn):
if not await self.get_session_if_accessible(session_id):
raise ValueError(f"Session {session_id} not found or access denied")
await self.kvstore.set(
key=f"session:{self.agent_id}:{session_id}:{turn.turn_id}",
value=turn.model_dump_json(),
)
async def get_session_turns(self, session_id: str) -> List[Turn]:
if not await self.get_session_if_accessible(session_id):
raise ValueError(f"Session {session_id} not found or access denied")
values = await self.kvstore.range(
start_key=f"session:{self.agent_id}:{session_id}:",
end_key=f"session:{self.agent_id}:{session_id}:\xff\xff\xff\xff",
@ -87,6 +126,9 @@ class AgentPersistence:
return turns
async def get_session_turn(self, session_id: str, turn_id: str) -> Optional[Turn]:
if not await self.get_session_if_accessible(session_id):
raise ValueError(f"Session {session_id} not found or access denied")
value = await self.kvstore.get(
key=f"session:{self.agent_id}:{session_id}:{turn_id}",
)
@ -95,24 +137,36 @@ class AgentPersistence:
return Turn(**json.loads(value))
async def set_in_progress_tool_call_step(self, session_id: str, turn_id: str, step: ToolExecutionStep):
if not await self.get_session_if_accessible(session_id):
raise ValueError(f"Session {session_id} not found or access denied")
await self.kvstore.set(
key=f"in_progress_tool_call_step:{self.agent_id}:{session_id}:{turn_id}",
value=step.model_dump_json(),
)
async def get_in_progress_tool_call_step(self, session_id: str, turn_id: str) -> Optional[ToolExecutionStep]:
if not await self.get_session_if_accessible(session_id):
return None
value = await self.kvstore.get(
key=f"in_progress_tool_call_step:{self.agent_id}:{session_id}:{turn_id}",
)
return ToolExecutionStep(**json.loads(value)) if value else None
async def set_num_infer_iters_in_turn(self, session_id: str, turn_id: str, num_infer_iters: int):
if not await self.get_session_if_accessible(session_id):
raise ValueError(f"Session {session_id} not found or access denied")
await self.kvstore.set(
key=f"num_infer_iters_in_turn:{self.agent_id}:{session_id}:{turn_id}",
value=str(num_infer_iters),
)
async def get_num_infer_iters_in_turn(self, session_id: str, turn_id: str) -> Optional[int]:
if not await self.get_session_if_accessible(session_id):
return None
value = await self.kvstore.get(
key=f"num_infer_iters_in_turn:{self.agent_id}:{session_id}:{turn_id}",
)

16
llama_stack/providers/inline/datasetio/localfs/config.py
View file

@ -3,9 +3,10 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
from llama_stack.providers.utils.kvstore.config import (
KVStoreConfig,
SqliteKVStoreConfig,
@ -13,6 +14,13 @@ from llama_stack.providers.utils.kvstore.config import (
class LocalFSDatasetIOConfig(BaseModel):
kvstore: KVStoreConfig = SqliteKVStoreConfig(
db_path=(RUNTIME_BASE_DIR / "localfs_datasetio.db").as_posix()
) # Uses SQLite config specific to localfs storage
kvstore: KVStoreConfig
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="localfs_datasetio.db",
)
}

144
llama_stack/providers/inline/datasetio/localfs/datasetio.py
View file

@ -3,20 +3,16 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import base64
import os
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Any, Dict, List, Optional
from urllib.parse import urlparse
import pandas
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.datasetio import DatasetIO, PaginatedRowsResult
from llama_stack.apis.common.responses import PaginatedResponse
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Dataset
from llama_stack.providers.datatypes import DatasetsProtocolPrivate
from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_url
from llama_stack.providers.utils.datasetio.pagination import paginate_records
from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_uri
from llama_stack.providers.utils.kvstore import kvstore_impl
from .config import LocalFSDatasetIOConfig
@ -24,30 +20,7 @@ from .config import LocalFSDatasetIOConfig
DATASETS_PREFIX = "localfs_datasets:"
class BaseDataset(ABC):
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
@abstractmethod
def __len__(self) -> int:
raise NotImplementedError()
@abstractmethod
def __getitem__(self, idx):
raise NotImplementedError()
@abstractmethod
def load(self):
raise NotImplementedError()
@dataclass
class DatasetInfo:
dataset_def: Dataset
dataset_impl: BaseDataset
class PandasDataframeDataset(BaseDataset):
class PandasDataframeDataset:
def __init__(self, dataset_def: Dataset, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.dataset_def = dataset_def
@ -64,23 +37,19 @@ class PandasDataframeDataset(BaseDataset):
else:
return self.df.iloc[idx].to_dict()
def _validate_dataset_schema(self, df) -> pandas.DataFrame:
# note that we will drop any columns in dataset that are not in the schema
df = df[self.dataset_def.dataset_schema.keys()]
# check all columns in dataset schema are present
assert len(df.columns) == len(self.dataset_def.dataset_schema)
# TODO: type checking against column types in dataset schema
return df
def load(self) -> None:
async def load(self) -> None:
if self.df is not None:
return
df = get_dataframe_from_url(self.dataset_def.url)
if df is None:
raise ValueError(f"Failed to load dataset from {self.dataset_def.url}")
if self.dataset_def.source.type == "uri":
self.df = await get_dataframe_from_uri(self.dataset_def.source.uri)
elif self.dataset_def.source.type == "rows":
self.df = pandas.DataFrame(self.dataset_def.source.rows)
else:
raise ValueError(f"Unsupported dataset source type: {self.dataset_def.source.type}")
self.df = self._validate_dataset_schema(df)
if self.df is None:
raise ValueError(f"Failed to load dataset from {self.dataset_def.url}")
class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
@ -99,95 +68,44 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
for dataset in stored_datasets:
dataset = Dataset.model_validate_json(dataset)
dataset_impl = PandasDataframeDataset(dataset)
self.dataset_infos[dataset.identifier] = DatasetInfo(
dataset_def=dataset,
dataset_impl=dataset_impl,
)
self.dataset_infos[dataset.identifier] = dataset
async def shutdown(self) -> None: ...
async def register_dataset(
self,
dataset: Dataset,
dataset_def: Dataset,
) -> None:
# Store in kvstore
key = f"{DATASETS_PREFIX}{dataset.identifier}"
key = f"{DATASETS_PREFIX}{dataset_def.identifier}"
await self.kvstore.set(
key=key,
value=dataset.json(),
)
dataset_impl = PandasDataframeDataset(dataset)
self.dataset_infos[dataset.identifier] = DatasetInfo(
dataset_def=dataset,
dataset_impl=dataset_impl,
value=dataset_def.model_dump_json(),
)
self.dataset_infos[dataset_def.identifier] = dataset_def
async def unregister_dataset(self, dataset_id: str) -> None:
key = f"{DATASETS_PREFIX}{dataset_id}"
await self.kvstore.delete(key=key)
del self.dataset_infos[dataset_id]
async def get_rows_paginated(
async def iterrows(
self,
dataset_id: str,
rows_in_page: int,
page_token: Optional[str] = None,
filter_condition: Optional[str] = None,
) -> PaginatedRowsResult:
dataset_info = self.dataset_infos.get(dataset_id)
dataset_info.dataset_impl.load()
start_index: Optional[int] = None,
limit: Optional[int] = None,
) -> PaginatedResponse:
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
await dataset_impl.load()
if page_token and not page_token.isnumeric():
raise ValueError("Invalid page_token")
if page_token is None or len(page_token) == 0:
next_page_token = 0
else:
next_page_token = int(page_token)
start = next_page_token
if rows_in_page == -1:
end = len(dataset_info.dataset_impl)
else:
end = min(start + rows_in_page, len(dataset_info.dataset_impl))
rows = dataset_info.dataset_impl[start:end]
return PaginatedRowsResult(
rows=rows,
total_count=len(rows),
next_page_token=str(end),
)
records = dataset_impl.df.to_dict("records")
return paginate_records(records, start_index, limit)
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
dataset_info = self.dataset_infos.get(dataset_id)
if dataset_info is None:
raise ValueError(f"Dataset with id {dataset_id} not found")
dataset_impl = dataset_info.dataset_impl
dataset_impl.load()
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
await dataset_impl.load()
new_rows_df = pandas.DataFrame(rows)
new_rows_df = dataset_impl._validate_dataset_schema(new_rows_df)
dataset_impl.df = pandas.concat([dataset_impl.df, new_rows_df], ignore_index=True)
url = str(dataset_info.dataset_def.url)
parsed_url = urlparse(url)
if parsed_url.scheme == "file" or not parsed_url.scheme:
file_path = parsed_url.path
os.makedirs(os.path.dirname(file_path), exist_ok=True)
dataset_impl.df.to_csv(file_path, index=False)
elif parsed_url.scheme == "data":
# For data URLs, we need to update the base64-encoded content
if not parsed_url.path.startswith("text/csv;base64,"):
raise ValueError("Data URL must be a base64-encoded CSV")
csv_buffer = dataset_impl.df.to_csv(index=False)
base64_content = base64.b64encode(csv_buffer.encode("utf-8")).decode("utf-8")
dataset_info.dataset_def.url = URL(uri=f"data:text/csv;base64,{base64_content}")
else:
raise ValueError(
f"Unsupported URL scheme: {parsed_url.scheme}. Only file:// and data: URLs are supported for writing."
)

16
llama_stack/providers/inline/eval/meta_reference/config.py
View file

@ -3,9 +3,10 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
from llama_stack.providers.utils.kvstore.config import (
KVStoreConfig,
SqliteKVStoreConfig,
@ -13,6 +14,13 @@ from llama_stack.providers.utils.kvstore.config import (
class MetaReferenceEvalConfig(BaseModel):
kvstore: KVStoreConfig = SqliteKVStoreConfig(
db_path=(RUNTIME_BASE_DIR / "meta_reference_eval.db").as_posix()
) # Uses SQLite config specific to Meta Reference Eval storage
kvstore: KVStoreConfig
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="meta_reference_eval.db",
)
}

43
llama_stack/providers/inline/eval/meta_reference/eval.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import json
from typing import Any, Dict, List, Optional
from typing import Any, Dict, List
from tqdm import tqdm
@ -12,22 +12,17 @@ from llama_stack.apis.agents import Agents, StepType
from llama_stack.apis.benchmarks import Benchmark
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.inference import Inference, UserMessage
from llama_stack.apis.inference import Inference, SystemMessage, UserMessage
from llama_stack.apis.scoring import Scoring
from llama_stack.distribution.datatypes import Api
from llama_stack.providers.datatypes import BenchmarksProtocolPrivate
from llama_stack.providers.inline.agents.meta_reference.agent_instance import (
MEMORY_QUERY_TOOL,
)
from llama_stack.providers.utils.common.data_schema_validator import (
ColumnName,
get_valid_schemas,
validate_dataset_schema,
)
from llama_stack.providers.utils.common.data_schema_validator import ColumnName
from llama_stack.providers.utils.kvstore import kvstore_impl
from .....apis.common.job_types import Job
from .....apis.eval.eval import BenchmarkConfig, Eval, EvaluateResponse, JobStatus
from .....apis.common.job_types import Job, JobStatus
from .....apis.eval.eval import BenchmarkConfig, Eval, EvaluateResponse
from .config import MetaReferenceEvalConfig
EVAL_TASKS_PREFIX = "benchmarks:"
@ -88,15 +83,17 @@ class MetaReferenceEvalImpl(
task_def = self.benchmarks[benchmark_id]
dataset_id = task_def.dataset_id
scoring_functions = task_def.scoring_functions
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.eval.value))
all_rows = await self.datasetio_api.get_rows_paginated(
# TODO (xiyan): validate dataset schema
# dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=(-1 if benchmark_config.num_examples is None else benchmark_config.num_examples),
limit=(-1 if benchmark_config.num_examples is None else benchmark_config.num_examples),
)
res = await self.evaluate_rows(
benchmark_id=benchmark_id,
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
benchmark_config=benchmark_config,
)
@ -105,7 +102,7 @@ class MetaReferenceEvalImpl(
# need job scheduler queue (ray/celery) w/ jobs api
job_id = str(len(self.jobs))
self.jobs[job_id] = res
return Job(job_id=job_id)
return Job(job_id=job_id, status=JobStatus.completed)
async def _run_agent_generation(
self, input_rows: List[Dict[str, Any]], benchmark_config: BenchmarkConfig
@ -118,7 +115,7 @@ class MetaReferenceEvalImpl(
for i, x in tqdm(enumerate(input_rows)):
assert ColumnName.chat_completion_input.value in x, "Invalid input row"
input_messages = json.loads(x[ColumnName.chat_completion_input.value])
input_messages = [UserMessage(**x) for x in input_messages]
input_messages = [UserMessage(**x) for x in input_messages if x["role"] == "user"]
# NOTE: only single-turn agent generation is supported. Create a new session for each input row
session_create_response = await self.agents_api.create_agent_session(agent_id, f"session-{i}")
@ -168,10 +165,11 @@ class MetaReferenceEvalImpl(
generations.append({ColumnName.generated_answer.value: response.completion_message.content})
elif ColumnName.chat_completion_input.value in x:
chat_completion_input_json = json.loads(x[ColumnName.chat_completion_input.value])
input_messages = [UserMessage(**x) for x in chat_completion_input_json]
input_messages = [UserMessage(**x) for x in chat_completion_input_json if x["role"] == "user"]
messages = []
if candidate.system_message:
messages.append(candidate.system_message)
messages += [SystemMessage(**x) for x in chat_completion_input_json if x["role"] == "system"]
messages += input_messages
response = await self.inference_api.chat_completion(
model_id=candidate.model,
@ -218,17 +216,18 @@ class MetaReferenceEvalImpl(
return EvaluateResponse(generations=generations, scores=score_response.results)
async def job_status(self, benchmark_id: str, job_id: str) -> Optional[JobStatus]:
async def job_status(self, benchmark_id: str, job_id: str) -> Job:
if job_id in self.jobs:
return JobStatus.completed
return Job(job_id=job_id, status=JobStatus.completed)
return None
raise ValueError(f"Job {job_id} not found")
async def job_cancel(self, benchmark_id: str, job_id: str) -> None:
raise NotImplementedError("Job cancel is not implemented yet")
async def job_result(self, benchmark_id: str, job_id: str) -> EvaluateResponse:
status = await self.job_status(benchmark_id, job_id)
job = await self.job_status(benchmark_id, job_id)
status = job.status
if not status or status != JobStatus.completed:
raise ValueError(f"Job is not completed, Status: {status.value}")

3
llama_stack/providers/inline/inference/meta_reference/parallel_utils.py
View file

@ -10,6 +10,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import copy
import json
import logging
import multiprocessing
@ -213,7 +214,7 @@ def maybe_parse_message(maybe_json: Optional[str]) -> Optional[ProcessingMessage
def parse_message(json_str: str) -> ProcessingMessage:
data = json.loads(json_str)
return ProcessingMessageWrapper(**data).payload
return copy.deepcopy(ProcessingMessageWrapper(**data).payload)
def worker_process_entrypoint(

2
llama_stack/providers/inline/inference/sentence_transformers/sentence_transformers.py
View file

@ -43,7 +43,7 @@ class SentenceTransformersInferenceImpl(
async def shutdown(self) -> None:
pass
async def register_model(self, model: Model) -> None:
async def register_model(self, model: Model) -> Model:
return model
async def unregister_model(self, model_id: str) -> None:

4
llama_stack/providers/inline/inference/vllm/config.py
View file

@ -4,6 +4,8 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel, Field
from llama_stack.schema_utils import json_schema_type
@ -40,7 +42,7 @@ class VLLMConfig(BaseModel):
)
@classmethod
def sample_run_config(cls):
def sample_run_config(cls, **kwargs: Any) -> Dict[str, Any]:
return {
"tensor_parallel_size": "${env.TENSOR_PARALLEL_SIZE:1}",
"max_tokens": "${env.MAX_TOKENS:4096}",

1
llama_stack/providers/inline/inference/vllm/vllm.py
View file

@ -582,6 +582,7 @@ class VLLMInferenceImpl(Inference, ModelsProtocolPrivate):
tool_name=t.function.name,
# vLLM function args come back as a string. Llama Stack expects JSON.
arguments=json.loads(t.function.arguments),
arguments_json=t.function.arguments,
)
for t in vllm_message.tool_calls
],

8
llama_stack/providers/inline/post_training/common/validator.py
View file

@ -9,6 +9,9 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any
from llama_stack.apis.common.type_system import (
ChatCompletionInputType,
DialogType,
@ -20,7 +23,7 @@ from llama_stack.providers.utils.common.data_schema_validator import (
validate_dataset_schema,
)
EXPECTED_DATASET_SCHEMA = {
EXPECTED_DATASET_SCHEMA: dict[str, list[dict[str, Any]]] = {
"instruct": [
{
ColumnName.chat_completion_input.value: ChatCompletionInputType(),
@ -41,6 +44,9 @@ async def validate_input_dataset_schema(
dataset_type: str,
) -> None:
dataset_def = await datasets_api.get_dataset(dataset_id=dataset_id)
if not dataset_def:
raise ValueError(f"Dataset {dataset_id} does not exist.")
if not dataset_def.dataset_schema or len(dataset_def.dataset_schema) == 0:
raise ValueError(f"Dataset {dataset_id} does not have a schema defined.")

8
llama_stack/providers/inline/post_training/torchtune/common/checkpointer.py
View file

@ -37,7 +37,7 @@ class TorchtuneCheckpointer:
checkpoint_files: List[str],
output_dir: str,
model_type: str,
) -> None:
):
# Fail fast if ``checkpoint_files`` is invalid
# TODO: support loading more than one file
if len(checkpoint_files) != 1:
@ -58,7 +58,7 @@ class TorchtuneCheckpointer:
"""
Load Meta checkpoint from file. Currently only loading from a single file is supported.
"""
state_dict: Dict[str:Any] = {}
state_dict: Dict[str, Any] = {}
model_state_dict = safe_torch_load(self._checkpoint_path)
if self._model_type == ModelType.LLAMA3_VISION:
from torchtune.models.llama3_2_vision._convert_weights import (
@ -85,10 +85,10 @@ class TorchtuneCheckpointer:
state_dict: Dict[str, Any],
epoch: int,
adapter_only: bool = False,
checkpoint_format: str = "meta",
checkpoint_format: str | None = None,
) -> str:
model_file_path = Path(self._output_dir) / f"{self._model_id}-{self._training_algorithm}-{epoch}"
if checkpoint_format == "meta":
if checkpoint_format == "meta" or checkpoint_format is None:
self._save_meta_format_checkpoint(model_file_path, state_dict, adapter_only)
elif checkpoint_format == "huggingface":
# Note: for saving hugging face format checkpoints, we only suppport saving adapter weights now

13
llama_stack/providers/inline/post_training/torchtune/common/utils.py
View file

@ -10,7 +10,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Callable, Dict
from typing import Callable, Dict
import torch
from pydantic import BaseModel
@ -25,10 +25,13 @@ from llama_stack.apis.post_training import DatasetFormat
from llama_stack.models.llama.datatypes import Model
from llama_stack.models.llama.sku_list import resolve_model
BuildLoraModelCallable = Callable[..., torch.nn.Module]
BuildTokenizerCallable = Callable[..., Llama3Tokenizer]
class ModelConfig(BaseModel):
model_definition: Any
tokenizer_type: Any
model_definition: BuildLoraModelCallable
tokenizer_type: BuildTokenizerCallable
checkpoint_type: str
@ -51,10 +54,6 @@ DATA_FORMATS: Dict[str, Transform] = {
}
BuildLoraModelCallable = Callable[..., torch.nn.Module]
BuildTokenizerCallable = Callable[..., Llama3Tokenizer]
def _validate_model_id(model_id: str) -> Model:
model = resolve_model(model_id)
if model is None or model.core_model_id.value not in MODEL_CONFIGS:

8
llama_stack/providers/inline/post_training/torchtune/config.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Literal, Optional
from typing import Any, Dict, Literal, Optional
from pydantic import BaseModel
@ -12,3 +12,9 @@ from pydantic import BaseModel
class TorchtunePostTrainingConfig(BaseModel):
torch_seed: Optional[int] = None
checkpoint_format: Optional[Literal["meta", "huggingface"]] = "meta"
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {
"checkpoint_format": "meta",
}

2
llama_stack/providers/inline/post_training/torchtune/datasets/sft.py
View file

@ -55,7 +55,7 @@ class SFTDataset(Dataset):
if "messages" in transformed_sample:
validate_messages(transformed_sample["messages"])
tokenized_dict = self._model_transform(transformed_sample)
tokenized_dict: dict[str, Any] = self._model_transform(transformed_sample)
if not ("tokens" in tokenized_dict and "mask" in tokenized_dict):
keys_str = ", ".join(tokenized_dict.keys())

8
llama_stack/providers/inline/post_training/torchtune/post_training.py
View file

@ -3,7 +3,7 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from datetime import datetime
from datetime import datetime, timezone
from typing import Any, Dict, Optional
from llama_stack.apis.datasetio import DatasetIO
@ -64,7 +64,7 @@ class TorchtunePostTrainingImpl:
job_status_response = PostTrainingJobStatusResponse(
job_uuid=job_uuid,
status=JobStatus.scheduled,
scheduled_at=datetime.now(),
scheduled_at=datetime.now(timezone.utc),
)
self.jobs[job_uuid] = job_status_response
@ -84,7 +84,7 @@ class TorchtunePostTrainingImpl:
)
job_status_response.status = JobStatus.in_progress
job_status_response.started_at = datetime.now()
job_status_response.started_at = datetime.now(timezone.utc)
await recipe.setup()
resources_allocated, checkpoints = await recipe.train()
@ -93,7 +93,7 @@ class TorchtunePostTrainingImpl:
job_status_response.resources_allocated = resources_allocated
job_status_response.checkpoints = checkpoints
job_status_response.status = JobStatus.completed
job_status_response.completed_at = datetime.now()
job_status_response.completed_at = datetime.now(timezone.utc)
except Exception:
job_status_response.status = JobStatus.failed

55
llama_stack/providers/inline/post_training/torchtune/recipes/lora_finetuning_single_device.py
View file

@ -8,7 +8,7 @@ import gc
import logging
import os
import time
from datetime import datetime
from datetime import datetime, timezone
from functools import partial
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
@ -37,10 +37,10 @@ from llama_stack.apis.common.training_types import PostTrainingMetric
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.post_training import (
AlgorithmConfig,
Checkpoint,
LoraFinetuningConfig,
OptimizerConfig,
QATFinetuningConfig,
TrainingConfig,
)
from llama_stack.distribution.utils.config_dirs import DEFAULT_CHECKPOINT_DIR
@ -73,6 +73,9 @@ class LoraFinetuningSingleDevice:
# Currently logging only logs limited training metrics to local disk
# will figure out more loggings and how it works with telemetry in future PRs
_checkpointer: TorchtuneCheckpointer
def __init__(
self,
config: TorchtunePostTrainingConfig,
@ -82,7 +85,7 @@ class LoraFinetuningSingleDevice:
logger_config: Dict[str, Any],
model: str,
checkpoint_dir: Optional[str],
algorithm_config: Optional[AlgorithmConfig],
algorithm_config: LoraFinetuningConfig | QATFinetuningConfig | None,
datasetio_api: DatasetIO,
datasets_api: Datasets,
) -> None:
@ -109,12 +112,12 @@ class LoraFinetuningSingleDevice:
return str(checkpoint_dir)
if checkpoint_dir and checkpoint_dir != "null":
self.checkpoint_dir = config.checkpoint_dir
self.checkpoint_dir = checkpoint_dir
else:
model = resolve_model(self.model_id)
if model is None:
model_obj = resolve_model(self.model_id)
if model_obj is None:
raise ValueError(f"{self.model_id} not found. Your model id should be in the llama models SKU list")
self.checkpoint_dir = model_checkpoint_dir(model)
self.checkpoint_dir = model_checkpoint_dir(model_obj)
self._output_dir = str(DEFAULT_CHECKPOINT_DIR)
self._checkpoint_format = config.checkpoint_format
@ -135,16 +138,16 @@ class LoraFinetuningSingleDevice:
self.max_validation_steps = training_config.max_validation_steps
self._clip_grad_norm = 1.0
self._enable_activation_checkpointing = (
(training_config.efficiency_config.enable_activation_checkpointing)
if training_config.efficiency_config
else False
)
self._enable_activation_offloading = (
(training_config.efficiency_config.enable_activation_offloading)
if training_config.efficiency_config
else False
)
self._enable_activation_checkpointing = False
self._enable_activation_offloading = False
if training_config.efficiency_config:
if training_config.efficiency_config.enable_activation_checkpointing:
self._enable_activation_checkpointing = (
training_config.efficiency_config.enable_activation_checkpointing
)
if training_config.efficiency_config.enable_activation_offloading:
self._enable_activation_offloading = training_config.efficiency_config.enable_activation_offloading
self.datasetio_api = datasetio_api
self.datasets_api = datasets_api
@ -328,13 +331,13 @@ class LoraFinetuningSingleDevice:
batch_size: int,
) -> Tuple[DistributedSampler, DataLoader]:
async def fetch_rows(dataset_id: str):
return await self.datasetio_api.get_rows_paginated(
return await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
all_rows = await fetch_rows(dataset_id)
rows = all_rows.rows
rows = all_rows.data
await validate_input_dataset_schema(
datasets_api=self.datasets_api,
@ -451,12 +454,12 @@ class LoraFinetuningSingleDevice:
"""
# Initialize tokens count and running loss (for grad accumulation)
t0 = time.perf_counter()
running_loss = 0
running_loss: float = 0.0
num_tokens = 0
# training artifacts
checkpoints = []
memory_stats = {}
memory_stats: Dict[str, Any] = {}
# self.epochs_run should be non-zero when we're resuming from a checkpoint
for curr_epoch in range(self.epochs_run, self.total_epochs):
@ -484,7 +487,7 @@ class LoraFinetuningSingleDevice:
# Loss is normalized by default so we multiply by the number of tokens
# This way we can normalize by the total number of tokens if we're accumulating gradients
current_loss = await self._loss_step(batch) * current_num_tokens
running_loss += current_loss
running_loss += current_loss.detach().item()
current_loss.backward()
# Step with optimizer
@ -500,7 +503,7 @@ class LoraFinetuningSingleDevice:
# Update the number of steps when the weights are updated
self.global_step += 1
loss_to_log = running_loss.item() / num_tokens
loss_to_log = running_loss / num_tokens
pbar.update(1)
pbar.set_description(f"{curr_epoch + 1}|{self.global_step}|Loss: {loss_to_log}")
@ -523,7 +526,7 @@ class LoraFinetuningSingleDevice:
)
# Reset running stats for the next step
running_loss = 0
running_loss = 0.0
num_tokens = 0
t0 = time.perf_counter()
@ -532,7 +535,7 @@ class LoraFinetuningSingleDevice:
checkpoint_path = await self.save_checkpoint(epoch=curr_epoch)
checkpoint = Checkpoint(
identifier=f"{self.model_id}-sft-{curr_epoch}",
created_at=datetime.now(),
created_at=datetime.now(timezone.utc),
epoch=curr_epoch,
post_training_job_id=self.job_uuid,
path=checkpoint_path,

24
llama_stack/providers/inline/preprocessing/basic/basic.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
import logging
import re
from typing import List, Optional
from typing import Any, List, Optional
import httpx
@ -39,18 +39,24 @@ class InclineBasicPreprocessorImpl(Preprocessing, PreprocessorsProtocolPrivate):
# this preprocessor optionally retrieves the documents and converts them into plain text
output_types = [PreprocessingDataType.raw_text_document]
preprocessor_store = None
URL_VALIDATION_PATTERN = re.compile("^(https?://|file://|data:)")
def __init__(self, config: InlineBasicPreprocessorConfig) -> None:
self.config = config
async def initialize(self) -> None: ...
async def initialize(self) -> None:
pass
async def shutdown(self) -> None: ...
async def shutdown(self) -> None:
pass
async def register_preprocessor(self, preprocessor: Preprocessor) -> None: ...
async def register_preprocessor(self, preprocessor: Preprocessor) -> None:
pass
async def unregister_preprocessor(self, preprocessor_id: str) -> None: ...
async def unregister_preprocessor(self, preprocessor_id: str) -> None:
pass
async def do_preprocess(
self,
@ -78,7 +84,7 @@ class InclineBasicPreprocessorImpl(Preprocessing, PreprocessorsProtocolPrivate):
)
continue
elif input_type == PreprocessingDataType.raw_text_document:
document = interleaved_content_as_str(inp.data_element_path_or_content)
document = interleaved_content_as_str(inp.data_element_path_or_content) # type: ignore
else:
log.error(f"Unexpected preprocessor input type: {input_type}")
continue
@ -112,7 +118,9 @@ class InclineBasicPreprocessorImpl(Preprocessing, PreprocessorsProtocolPrivate):
if isinstance(preprocessor_input.data_element_path_or_content, URL):
return PreprocessingDataType.document_uri
if InclineBasicPreprocessorImpl.URL_VALIDATION_PATTERN.match(preprocessor_input.data_element_path_or_content):
if InclineBasicPreprocessorImpl.URL_VALIDATION_PATTERN.match(
str(preprocessor_input.data_element_path_or_content)
):
return PreprocessingDataType.document_uri
if preprocessor_input.data_element_format == PreprocessingDataFormat.pdf:
return PreprocessingDataType.binary_document
@ -120,7 +128,7 @@ class InclineBasicPreprocessorImpl(Preprocessing, PreprocessorsProtocolPrivate):
return PreprocessingDataType.raw_text_document
@staticmethod
async def _fetch_document(preprocessor_input: PreprocessingDataElement) -> str | None:
async def _fetch_document(preprocessor_input: PreprocessingDataElement) -> Any:
if isinstance(preprocessor_input.data_element_path_or_content, str):
url = preprocessor_input.data_element_path_or_content
if not InclineBasicPreprocessorImpl.URL_VALIDATION_PATTERN.match(url):

6
llama_stack/providers/inline/preprocessing/simple_chunking/simple_chunking.py
View file

@ -36,6 +36,8 @@ class InclineSimpleChunkingImpl(Preprocessing, PreprocessorsProtocolPrivate):
input_types = [PreprocessingDataType.raw_text_document]
output_types = [PreprocessingDataType.chunks]
preprocessor_store = None
def __init__(self, config: InclineSimpleChunkingConfig) -> None:
self.config = config
@ -59,7 +61,7 @@ class InclineSimpleChunkingImpl(Preprocessing, PreprocessorsProtocolPrivate):
for inp in preprocessor_inputs:
new_chunks = self.make_overlapped_chunks(
inp.data_element_id, inp.data_element_path_or_content, window_len, overlap_len
inp.data_element_id, str(inp.data_element_path_or_content), window_len, overlap_len
)
for i, chunk in enumerate(new_chunks):
new_chunk_data_element = PreprocessingDataElement(
@ -79,7 +81,7 @@ class InclineSimpleChunkingImpl(Preprocessing, PreprocessorsProtocolPrivate):
) -> PreprocessorResponse:
return await self.do_preprocess(preprocessor_id="", preprocessor_inputs=preprocessor_inputs)
def _resolve_chunk_size_params(self, options: PreprocessorOptions) -> Tuple[int, int]:
def _resolve_chunk_size_params(self, options: PreprocessorOptions | None) -> Tuple[int, int]:
window_len = (options or {}).get(
str(SimpleChunkingOptions.chunk_size_in_tokens), self.config.chunk_size_in_tokens
)

6
llama_stack/providers/inline/safety/code_scanner/config.py
View file

@ -4,8 +4,12 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
class CodeScannerConfig(BaseModel):
pass
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {}

8
llama_stack/providers/inline/safety/llama_guard/config.py
View file

@ -4,10 +4,16 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List
from typing import Any, Dict, List
from pydantic import BaseModel
class LlamaGuardConfig(BaseModel):
excluded_categories: List[str] = []
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {
"excluded_categories": [],
}

7
llama_stack/providers/inline/safety/llama_guard/llama_guard.py
View file

@ -227,13 +227,6 @@ class LlamaGuardShield:
if len(messages) >= 2 and (messages[0].role == Role.user.value and messages[1].role == Role.user.value):
messages = messages[1:]
for i in range(1, len(messages)):
if messages[i].role == messages[i - 1].role:
for i, m in enumerate(messages):
print(f"{i}: {m.role}: {m.content}")
raise ValueError(
f"Messages must alternate between user and assistant. Message {i} has the same role as message {i - 1}"
)
return messages
async def run(self, messages: List[Message]) -> RunShieldResponse:

7
llama_stack/providers/inline/safety/prompt_guard/config.py
View file

@ -5,6 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict
from pydantic import BaseModel, field_validator
@ -23,3 +24,9 @@ class PromptGuardConfig(BaseModel):
if v not in [t.value for t in PromptGuardType]:
raise ValueError(f"Unknown prompt guard type: {v}")
return v
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {
"guard_type": "injection",
}

7
llama_stack/providers/inline/scoring/basic/config.py
View file

@ -3,7 +3,12 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
class BasicScoringConfig(BaseModel): ...
class BasicScoringConfig(BaseModel):
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {}

23
llama_stack/providers/inline/scoring/basic/scoring.py
View file

@ -22,12 +22,25 @@ from llama_stack.providers.utils.common.data_schema_validator import (
)
from .config import BasicScoringConfig
from .scoring_fn.bfcl_scoring_fn import BFCLScoringFn
from .scoring_fn.docvqa_scoring_fn import DocVQAScoringFn
from .scoring_fn.equality_scoring_fn import EqualityScoringFn
from .scoring_fn.regex_parser_math_response_scoring_fn import RegexParserMathResponseScoringFn
from .scoring_fn.ifeval_scoring_fn import IfEvalScoringFn
from .scoring_fn.regex_parser_math_response_scoring_fn import (
RegexParserMathResponseScoringFn,
)
from .scoring_fn.regex_parser_scoring_fn import RegexParserScoringFn
from .scoring_fn.subset_of_scoring_fn import SubsetOfScoringFn
FIXED_FNS = [EqualityScoringFn, SubsetOfScoringFn, RegexParserScoringFn, RegexParserMathResponseScoringFn]
FIXED_FNS = [
EqualityScoringFn,
SubsetOfScoringFn,
RegexParserScoringFn,
RegexParserMathResponseScoringFn,
BFCLScoringFn,
IfEvalScoringFn,
DocVQAScoringFn,
]
class BasicScoringImpl(
@ -75,12 +88,12 @@ class BasicScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
)
if save_results_dataset:

93
llama_stack/providers/inline/scoring/basic/scoring_fn/bfcl_scoring_fn.py Normal file
View file

@ -0,0 +1,93 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import json
import re
from typing import Any, Dict, Optional
from llama_stack.apis.scoring import ScoringResultRow
from llama_stack.apis.scoring_functions import ScoringFnParams
from llama_stack.providers.utils.scoring.base_scoring_fn import RegisteredBaseScoringFn
from ..utils.bfcl.ast_parser import decode_ast
from ..utils.bfcl.checker import ast_checker, is_empty_output
from .fn_defs.bfcl import bfcl
def postprocess(x: Dict[str, Any], test_category: str) -> Dict[str, Any]:
contain_func_call = False
error = None
error_type = None
checker_result = {}
try:
prediction = decode_ast(x["generated_answer"], x["language"]) or ""
contain_func_call = True
# if not is_function_calling_format_output(prediction):
if is_empty_output(prediction):
contain_func_call = False
error = "Did not output in the specified format. Note: the model_result is wrapped in a string to ensure json serializability."
error_type = "ast_decoder:decoder_wrong_output_format"
else:
checker_result = ast_checker(
json.loads(x["function"]),
prediction,
json.loads(x["ground_truth"]),
x["language"],
test_category=test_category,
model_name="",
)
except Exception as e:
prediction = ""
error = f"Invalid syntax. Failed to decode AST. {str(e)}"
error_type = "ast_decoder:decoder_failed"
return {
"prediction": prediction,
"contain_func_call": contain_func_call,
"valid": checker_result.get("valid", False),
"error": error or checker_result.get("error", ""),
"error_type": error_type or checker_result.get("error_type", ""),
}
def gen_valid(x: Dict[str, Any]) -> Dict[str, float]:
return {"valid": x["valid"]}
def gen_relevance_acc(x: Dict[str, Any]) -> Dict[str, float]:
# This function serves for both relevance and irrelevance tests, which share the exact opposite logic.
# If `test_category` is "irrelevance", the model is expected to output no function call.
# No function call means either the AST decoding fails (a error message is generated) or the decoded AST does not contain any function call (such as a empty list, `[]`).
# If `test_category` is "relevance", the model is expected to output to a function call, and empty list doesn't count as a function call.
acc = not x["contain_func_call"] if "irrelevance" in x["id"] else x["contain_func_call"]
return {"valid": float(acc)}
class BFCLScoringFn(RegisteredBaseScoringFn):
"""
A scoring_fn for BFCL
"""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.supported_fn_defs_registry = {
bfcl.identifier: bfcl,
}
async def score_row(
self,
input_row: Dict[str, Any],
scoring_fn_identifier: Optional[str] = "bfcl",
scoring_params: Optional[ScoringFnParams] = None,
) -> ScoringResultRow:
test_category = re.sub(r"_[0-9_-]+$", "", input_row["id"])
score_result = postprocess(input_row, test_category)
if test_category in {"irrelevance", "live_relevance", "live_irrelevance"}:
score = gen_relevance_acc(score_result)["valid"]
else:
score = gen_valid(score_result)["valid"]
return {
"score": float(score),
}

240
llama_stack/providers/inline/scoring/basic/scoring_fn/docvqa_scoring_fn.py Normal file
View file

@ -0,0 +1,240 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import json
import re
from typing import Any, Dict, Optional
from llama_stack.apis.scoring import ScoringResultRow
from llama_stack.apis.scoring_functions import ScoringFnParams
from llama_stack.providers.utils.scoring.base_scoring_fn import RegisteredBaseScoringFn
from .fn_defs.docvqa import docvqa
CONTRACTIONS = {
"aint": "ain't",
"arent": "aren't",
"cant": "can't",
"couldve": "could've",
"couldnt": "couldn't",
"couldn'tve": "couldn't've",
"couldnt've": "couldn't've",
"didnt": "didn't",
"doesnt": "doesn't",
"dont": "don't",
"hadnt": "hadn't",
"hadnt've": "hadn't've",
"hadn'tve": "hadn't've",
"hasnt": "hasn't",
"havent": "haven't",
"hed": "he'd",
"hed've": "he'd've",
"he'dve": "he'd've",
"hes": "he's",
"howd": "how'd",
"howll": "how'll",
"hows": "how's",
"Id've": "I'd've",
"I'dve": "I'd've",
"Im": "I'm",
"Ive": "I've",
"isnt": "isn't",
"itd": "it'd",
"itd've": "it'd've",
"it'dve": "it'd've",
"itll": "it'll",
"let's": "let's",
"maam": "ma'am",
"mightnt": "mightn't",
"mightnt've": "mightn't've",
"mightn'tve": "mightn't've",
"mightve": "might've",
"mustnt": "mustn't",
"mustve": "must've",
"neednt": "needn't",
"notve": "not've",
"oclock": "o'clock",
"oughtnt": "oughtn't",
"ow's'at": "'ow's'at",
"'ows'at": "'ow's'at",
"'ow'sat": "'ow's'at",
"shant": "shan't",
"shed've": "she'd've",
"she'dve": "she'd've",
"she's": "she's",
"shouldve": "should've",
"shouldnt": "shouldn't",
"shouldnt've": "shouldn't've",
"shouldn'tve": "shouldn't've",
"somebody'd": "somebodyd",
"somebodyd've": "somebody'd've",
"somebody'dve": "somebody'd've",
"somebodyll": "somebody'll",
"somebodys": "somebody's",
"someoned": "someone'd",
"someoned've": "someone'd've",
"someone'dve": "someone'd've",
"someonell": "someone'll",
"someones": "someone's",
"somethingd": "something'd",
"somethingd've": "something'd've",
"something'dve": "something'd've",
"somethingll": "something'll",
"thats": "that's",
"thered": "there'd",
"thered've": "there'd've",
"there'dve": "there'd've",
"therere": "there're",
"theres": "there's",
"theyd": "they'd",
"theyd've": "they'd've",
"they'dve": "they'd've",
"theyll": "they'll",
"theyre": "they're",
"theyve": "they've",
"twas": "'twas",
"wasnt": "wasn't",
"wed've": "we'd've",
"we'dve": "we'd've",
"weve": "we've",
"werent": "weren't",
"whatll": "what'll",
"whatre": "what're",
"whats": "what's",
"whatve": "what've",
"whens": "when's",
"whered": "where'd",
"wheres": "where's",
"whereve": "where've",
"whod": "who'd",
"whod've": "who'd've",
"who'dve": "who'd've",
"wholl": "who'll",
"whos": "who's",
"whove": "who've",
"whyll": "why'll",
"whyre": "why're",
"whys": "why's",
"wont": "won't",
"wouldve": "would've",
"wouldnt": "wouldn't",
"wouldnt've": "wouldn't've",
"wouldn'tve": "wouldn't've",
"yall": "y'all",
"yall'll": "y'all'll",
"y'allll": "y'all'll",
"yall'd've": "y'all'd've",
"y'alld've": "y'all'd've",
"y'all'dve": "y'all'd've",
"youd": "you'd",
"youd've": "you'd've",
"you'dve": "you'd've",
"youll": "you'll",
"youre": "you're",
"youve": "you've",
"1st": "first",
"2nd": "second",
"3rd": "third",
}
NUMBERS = {
"none": "0",
"zero": "0",
"one": "1",
"two": "2",
"three": "3",
"four": "4",
"five": "5",
"six": "6",
"seven": "7",
"eight": "8",
"nine": "9",
"ten": "10",
}
ARTICLES = [
"a",
"an",
"the",
"to",
"in",
"from",
"by",
] # Contains a bit more than just articles, but we want to get rid of these elements influencing the accuracy
PERIOD_STRIP = re.compile(r"(?!<=\d)(\.)(?!\d)")
COMMA_STRIP = re.compile(r"(\d)(\,)(\d)")
PUNCTUATION = [
";",
r"/",
"[",
"]",
'"',
"{",
"}",
"(",
")",
"=",
"+",
"\\",
"_",
"-",
">",
"<",
"@",
"`",
",",
"?",
"!",
]
def normalize_answer(s: str) -> str:
# process punctuation
for p in PUNCTUATION:
if (p + " " in s or " " + p in s) or (re.search(COMMA_STRIP, s) is not None):
s = s.replace(p, "")
else:
s = s.replace(p, " ")
s = PERIOD_STRIP.sub("", s, re.UNICODE)
# process digits and articles
temp_text = s.lower().split()
out_text = []
for word in temp_text:
word = NUMBERS.setdefault(word, word)
if word not in ARTICLES:
out_text.append(word)
# standardize contractions
for word_id, word in enumerate(out_text):
if word in CONTRACTIONS:
out_text[word_id] = CONTRACTIONS[word]
return " ".join(out_text)
class DocVQAScoringFn(RegisteredBaseScoringFn):
"""
docvqa basically matches the generated answer against several allowed
choices, but we need to normalize the answer to avoid penalizing
trivial differences
"""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.supported_fn_defs_registry = {
docvqa.identifier: docvqa,
}
async def score_row(
self,
input_row: Dict[str, Any],
scoring_fn_identifier: Optional[str] = "docvqa",
scoring_params: Optional[ScoringFnParams] = None,
) -> ScoringResultRow:
expected_answers = json.loads(input_row["expected_answer"])
generated_answer = input_row["generated_answer"]
score = 1.0 if normalize_answer(generated_answer) in [normalize_answer(s) for s in expected_answers] else 0.0
return {
"score": score,
}

21
llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/bfcl.py Normal file
View file

@ -0,0 +1,21 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.apis.common.type_system import NumberType
from llama_stack.apis.scoring_functions import (
AggregationFunctionType,
BasicScoringFnParams,
ScoringFn,
)
bfcl = ScoringFn(
identifier="basic::bfcl",
description="BFCL complex scoring",
return_type=NumberType(),
provider_id="basic",
provider_resource_id="bfcl",
params=BasicScoringFnParams(aggregation_functions=[AggregationFunctionType.accuracy]),
)

21
llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/docvqa.py Normal file
View file

@ -0,0 +1,21 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.apis.common.type_system import NumberType
from llama_stack.apis.scoring_functions import (
AggregationFunctionType,
BasicScoringFnParams,
ScoringFn,
)
docvqa = ScoringFn(
identifier="basic::docvqa",
description="DocVQA Visual Question & Answer scoring function",
return_type=NumberType(),
provider_id="basic",
provider_resource_id="docvqa",
params=BasicScoringFnParams(aggregation_functions=[AggregationFunctionType.accuracy]),
)

23
llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/ifeval.py Normal file
View file

@ -0,0 +1,23 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.apis.common.type_system import NumberType
from llama_stack.apis.scoring_functions import (
AggregationFunctionType,
BasicScoringFnParams,
ScoringFn,
)
ifeval = ScoringFn(
identifier="basic::ifeval",
description="Eval intruction follow capacity by checkping how many instructions can be followed in each example",
return_type=NumberType(),
provider_id="basic",
provider_resource_id="ifeval",
params=BasicScoringFnParams(
aggregation_functions=[AggregationFunctionType.weighted_average],
),
)

80
llama_stack/providers/inline/scoring/basic/scoring_fn/ifeval_scoring_fn.py Normal file
View file

@ -0,0 +1,80 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, Optional
from llama_stack.apis.scoring import ScoringResultRow
from llama_stack.apis.scoring_functions import ScoringFnParams
from llama_stack.providers.utils.scoring.base_scoring_fn import RegisteredBaseScoringFn
from .fn_defs.ifeval import (
ifeval,
)
class IfEvalScoringFn(RegisteredBaseScoringFn):
"""
A scoring_fn Instruction-Following Eval (IFEval) benchmark
"""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.supported_fn_defs_registry = {
ifeval.identifier: ifeval,
}
async def score_row(
self,
input_row: Dict[str, Any],
scoring_fn_identifier: Optional[str] = None,
scoring_params: Optional[ScoringFnParams] = None,
) -> ScoringResultRow:
from ..utils.ifeval_utils import INSTRUCTION_DICT, INSTRUCTION_LIST
assert scoring_fn_identifier is not None, "Scoring function identifier not found."
fn_def = self.supported_fn_defs_registry[scoring_fn_identifier]
if scoring_params is not None:
fn_def.params = scoring_params
instruction_list = input_row["instruction_id_list"]
generated_answer = input_row["generated_answer"].strip()
is_following_list = []
results = dict(
{k + "_correct": 0.0 for k in INSTRUCTION_LIST},
**{k + "_total": 0.0 for k in INSTRUCTION_LIST},
)
for index, instruction_id in enumerate(instruction_list):
instruction_cls = INSTRUCTION_DICT[instruction_id]
instruction = instruction_cls(instruction_id)
results[instruction_id + "_total"] += 1.0
results[instruction_id.split(":")[0] + "_total"] += 1.0
clean_input_row = {k: v for k, v in input_row["kwargs"][index].items() if v is not None}
print(clean_input_row)
instruction.build_description(**clean_input_row)
args = instruction.get_instruction_args()
if args and "prompt" in args:
instruction.build_description(prompt=input_row["prompt"])
if generated_answer and instruction.check_following(generated_answer):
is_following_list.append(True)
results[instruction_id + "_correct"] += 1.0
results[instruction_id.split(":")[0] + "_correct"] += 1.0
else:
is_following_list.append(False)
if len(is_following_list) == 0:
return {
"score": 0.0,
"weight": 0.0,
}
return {
"score": float(sum(is_following_list)) / float(len(is_following_list)),
"weight": float(len(is_following_list)),
}

7
llama_stack/providers/inline/telemetry/sample/config.py → llama_stack/providers/inline/scoring/basic/utils/bfcl/__init__.py
View file

@ -3,10 +3,3 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from pydantic import BaseModel
class SampleConfig(BaseModel):
host: str = "localhost"
port: int = 9999

296
llama_stack/providers/inline/scoring/basic/utils/bfcl/ast_parser.py Normal file
View file

@ -0,0 +1,296 @@
# ruff: noqa
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import ast
from .tree_sitter import get_parser
def parse_java_function_call(source_code):
if not source_code.endswith(";"):
source_code += ";" # Necessary for the parser not to register an error
parser = get_parser("java")
tree = parser.parse(bytes(source_code, "utf8"))
root_node = tree.root_node
if root_node.has_error:
raise Exception("Error parsing java the source code.")
def get_text(node):
"""Returns the text represented by the node."""
return source_code[node.start_byte : node.end_byte]
def traverse_node(node, nested=False):
if node.type == "string_literal":
if nested:
return get_text(node)
# Strip surrounding quotes from string literals
return get_text(node)[1:-1]
elif node.type == "character_literal":
if nested:
return get_text(node)
# Strip surrounding single quotes from character literals
return get_text(node)[1:-1]
"""Traverse the node to collect texts for complex structures."""
if node.type in [
"identifier",
"class_literal",
"type_identifier",
"method_invocation",
]:
return get_text(node)
elif node.type == "array_creation_expression":
# Handle array creation expression specifically
type_node = node.child_by_field_name("type")
value_node = node.child_by_field_name("value")
type_text = traverse_node(type_node, True)
value_text = traverse_node(value_node, True)
return f"new {type_text}[]{value_text}"
elif node.type == "object_creation_expression":
# Handle object creation expression specifically
type_node = node.child_by_field_name("type")
arguments_node = node.child_by_field_name("arguments")
type_text = traverse_node(type_node, True)
if arguments_node:
# Process each argument carefully, avoiding unnecessary punctuation
argument_texts = []
for child in arguments_node.children:
if child.type not in [
",",
"(",
")",
]: # Exclude commas and parentheses
argument_text = traverse_node(child, True)
argument_texts.append(argument_text)
arguments_text = ", ".join(argument_texts)
return f"new {type_text}({arguments_text})"
else:
return f"new {type_text}()"
elif node.type == "set":
# Handling sets specifically
items = [traverse_node(n, True) for n in node.children if n.type not in [",", "set"]]
return "{" + ", ".join(items) + "}"
elif node.child_count > 0:
return "".join(traverse_node(child, True) for child in node.children)
else:
return get_text(node)
def extract_arguments(args_node):
arguments = {}
for child in args_node.children:
if child.type == "assignment_expression":
# For named parameters
name_node, value_node = child.children[0], child.children[2]
name = get_text(name_node)
value = traverse_node(value_node)
if name in arguments:
if not isinstance(arguments[name], list):
arguments[name] = [arguments[name]]
arguments[name].append(value)
else:
arguments[name] = value
# arguments.append({'name': name, 'value': value})
elif child.type in ["identifier", "class_literal", "set"]:
# For unnamed parameters and handling sets
value = traverse_node(child)
if None in arguments:
if not isinstance(arguments[None], list):
arguments[None] = [arguments[None]]
arguments[None].append(value)
else:
arguments[None] = value
return arguments
def traverse(node):
if node.type == "method_invocation":
# Extract the function name and its arguments
method_name = get_text(node.child_by_field_name("name"))
class_name_node = node.child_by_field_name("object")
if class_name_node:
class_name = get_text(class_name_node)
function_name = f"{class_name}.{method_name}"
else:
function_name = method_name
arguments_node = node.child_by_field_name("arguments")
if arguments_node:
arguments = extract_arguments(arguments_node)
for key, value in arguments.items():
if isinstance(value, list):
raise Exception("Error: Multiple arguments with the same name are not supported.")
return [{function_name: arguments}]
else:
for child in node.children:
result = traverse(child)
if result:
return result
result = traverse(root_node)
return result if result else {}
def parse_javascript_function_call(source_code):
if not source_code.endswith(";"):
source_code += ";" # Necessary for the parser not to register an error
parser = get_parser("javascript")
# Parse the source code
tree = parser.parse(bytes(source_code, "utf8"))
root_node = tree.root_node
if root_node.has_error:
raise Exception("Error js parsing the source code.")
# Function to recursively extract argument details
def extract_arguments(node):
args = {}
for child in node.children:
if child.type == "assignment_expression":
# Extract left (name) and right (value) parts of the assignment
name = child.children[0].text.decode("utf-8")
value = child.children[2].text.decode("utf-8")
if (value.startswith('"') and value.endswith('"')) or (value.startswith("'") and value.endswith("'")):
value = value[1:-1] # Trim the quotation marks
if name in args:
if not isinstance(args[name], list):
args[name] = [args[name]]
args[name].append(value)
else:
args[name] = value
elif child.type == "identifier" or child.type == "true":
# Handle non-named arguments and boolean values
value = child.text.decode("utf-8")
if None in args:
if not isinstance(args[None], list):
args[None] = [args[None]]
args[None].append(value)
else:
args[None] = value
return args
# Find the function call and extract its name and arguments
if root_node.type == "program":
for child in root_node.children:
if child.type == "expression_statement":
for sub_child in child.children:
if sub_child.type == "call_expression":
function_name = sub_child.children[0].text.decode("utf8")
arguments_node = sub_child.children[1]
parameters = extract_arguments(arguments_node)
for key, value in parameters.items():
if isinstance(value, list):
raise Exception("Error: Multiple arguments with the same name are not supported.")
result = [{function_name: parameters}]
return result
def ast_parse(input_str, language="Python"):
if language == "Python":
cleaned_input = input_str.strip("[]'")
parsed = ast.parse(cleaned_input, mode="eval")
extracted = []
if isinstance(parsed.body, ast.Call):
extracted.append(resolve_ast_call(parsed.body))
else:
for elem in parsed.body.elts:
extracted.append(resolve_ast_call(elem))
return extracted
elif language == "Java":
return parse_java_function_call(input_str[1:-1]) # Remove the [ and ] from the string
elif language == "JavaScript":
return parse_javascript_function_call(input_str[1:-1])
else:
raise NotImplementedError(f"Unsupported language: {language}")
def resolve_ast_call(elem):
# Handle nested attributes for deeply nested module paths
func_parts = []
func_part = elem.func
while isinstance(func_part, ast.Attribute):
func_parts.append(func_part.attr)
func_part = func_part.value
if isinstance(func_part, ast.Name):
func_parts.append(func_part.id)
func_name = ".".join(reversed(func_parts))
args_dict = {}
# Parse when args are simply passed as an unnamed dictionary arg
for arg in elem.args:
if isinstance(arg, ast.Dict):
for key, value in zip(arg.keys, arg.values):
if isinstance(key, ast.Constant):
arg_name = key.value
output = resolve_ast_by_type(value)
args_dict[arg_name] = output
for arg in elem.keywords:
output = resolve_ast_by_type(arg.value)
args_dict[arg.arg] = output
return {func_name: args_dict}
def resolve_ast_by_type(value):
if isinstance(value, ast.Constant):
if value.value is Ellipsis:
output = "..."
else:
output = value.value
elif isinstance(value, ast.UnaryOp):
output = -value.operand.value
elif isinstance(value, ast.List):
output = [resolve_ast_by_type(v) for v in value.elts]
elif isinstance(value, ast.Dict):
output = {resolve_ast_by_type(k): resolve_ast_by_type(v) for k, v in zip(value.keys, value.values)}
elif isinstance(value, ast.NameConstant): # Added this condition to handle boolean values
output = value.value
elif isinstance(value, ast.BinOp): # Added this condition to handle function calls as arguments
output = eval(ast.unparse(value))
elif isinstance(value, ast.Name):
output = value.id
elif isinstance(value, ast.Call):
if len(value.keywords) == 0:
output = ast.unparse(value)
else:
output = resolve_ast_call(value)
elif isinstance(value, ast.Tuple):
output = tuple(resolve_ast_by_type(v) for v in value.elts)
elif isinstance(value, ast.Lambda):
output = eval(ast.unparse(value.body[0].value))
elif isinstance(value, ast.Ellipsis):
output = "..."
elif isinstance(value, ast.Subscript):
try:
output = ast.unparse(value.body[0].value)
except:
output = ast.unparse(value.value) + "[" + ast.unparse(value.slice) + "]"
else:
raise Exception(f"Unsupported AST type: {type(value)}")
return output
def decode_ast(result, language="Python"):
func = result
func = func.replace("\n", "") # remove new line characters
if not func.startswith("["):
func = "[" + func
if not func.endswith("]"):
func = func + "]"
decoded_output = ast_parse(func, language)
return decoded_output
def decode_execute(result):
func = result
func = func.replace("\n", "") # remove new line characters
if not func.startswith("["):
func = "[" + func
if not func.endswith("]"):
func = func + "]"
decode_output = ast_parse(func)
execution_list = []
for function_call in decode_output:
for key, value in function_call.items():
execution_list.append(f"{key}({','.join([f'{k}={repr(v)}' for k, v in value.items()])})")
return execution_list

989
llama_stack/providers/inline/scoring/basic/utils/bfcl/checker.py Normal file
View file

@ -0,0 +1,989 @@
# ruff: noqa
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import json
import re
import time
from typing import Any
# Comment out for now until we actually use the rest checker in evals
# import requests # Do not remove this import even though it seems to be unused. It's used in the executable_checker_rest function.
class NoAPIKeyError(Exception):
def __init__(self):
self.message = "❗️Please fill in the API keys in the function_credential_config.json file. If you do not provide the API keys, the executable test category results will be inaccurate."
super().__init__(self.message)
REAL_TIME_MATCH_ALLOWED_DIFFERENCE = 0.2
JAVA_TYPE_CONVERSION = {
"byte": int,
"short": int,
"integer": int,
"float": float,
"double": float,
"long": int,
"boolean": bool,
"char": str,
"Array": list,
"ArrayList": list,
"Set": set,
"HashMap": dict,
"Hashtable": dict,
"Queue": list, # this can be `queue.Queue` as well, for simplicity we check with list
"Stack": list,
"String": str,
"any": str,
}
JS_TYPE_CONVERSION = {
"String": str,
"integer": int,
"float": float,
"Bigint": int,
"Boolean": bool,
"dict": dict,
"array": list,
"any": str,
}
# We switch to conditional import for the following two imports to avoid unnecessary installations.
# User doesn't need to setup the tree-sitter packages if they are not running the test for that language.
# from js_type_converter import js_type_converter
# from java_type_converter import java_type_converter
PYTHON_TYPE_MAPPING = {
"string": str,
"integer": int,
"float": float,
"boolean": bool,
"array": list,
"tuple": list,
"dict": dict,
"any": str,
}
# This is the list of types that we need to recursively check its values
PYTHON_NESTED_TYPE_CHECK_LIST = ["array", "tuple"]
NESTED_CONVERSION_TYPE_LIST = ["Array", "ArrayList", "array"]
#### Helper functions for AST ####
def find_description(func_descriptions, name):
if type(func_descriptions) == list:
for func_description in func_descriptions:
if func_description["name"] == name:
return func_description
return None
else:
# it is a dict, there is only one function
return func_descriptions
def get_possible_answer_type(possible_answer: list):
for answer in possible_answer:
if answer != "": # Optional parameter
return type(answer)
return None
def type_checker(
param: str,
value,
possible_answer: list,
expected_type_description: str,
expected_type_converted,
nested_type_converted,
):
# NOTE: This type checker only supports nested type checking for one level deep.
# We didn't implement recursive type checking for nested types, as it's not needed for the current use case and it's very complex.
result: Any = {
"valid": True,
"error": [],
"is_variable": False,
"error_type": "type_error:simple",
}
is_variable = False
# check for the case where a variable is used instead of a actual value.
# use the type in possible_answer as the expected type
possible_answer_type = get_possible_answer_type(possible_answer)
# if possible_answer only contains optional parameters, we can't determine the type
if possible_answer_type != None:
# we are being precise here.
# in fact, possible_answer_type should always be string, as that's how we treat varibale in possible_answer
if possible_answer_type != expected_type_converted:
is_variable = True
# value is the same type as in function description
if type(value) == expected_type_converted:
# We don't need to do recursive check for simple types
if nested_type_converted == None:
result["is_variable"] = is_variable
return result
else:
for possible_answer_item in possible_answer:
flag = True # Each parameter should match to at least one possible answer type.
# Here, we assume that each item should be the same type. We could also relax it.
if type(possible_answer_item) == list:
for value_item in value:
checker_result = type_checker(
param,
value_item,
possible_answer_item,
str(nested_type_converted),
nested_type_converted,
None,
)
if not checker_result["valid"]:
flag = False
break
if flag:
return {"valid": True, "error": [], "is_variable": is_variable}
result["valid"] = False
result["error"] = [
f"Nested type checking failed for parameter {repr(param)}. Expected outer type {expected_type_description} with inner type {str(nested_type_converted)}. Parameter value: {repr(value)}."
]
result["error_type"] = "type_error:nested"
# value is not as expected, check for the case where a variable is used instead of a actual value
# use the type in possible_answer as the expected type
possible_answer_type = get_possible_answer_type(possible_answer)
# if possible_answer only contains optional parameters, we can't determine the type
if possible_answer_type != None:
# we are being precise here.
# in fact, possible_answer_type should always be string, as that's how we treat varibale in possible_answer
if type(value) == possible_answer_type:
result["is_variable"] = True
return result
result["valid"] = False
result["error"].append(
f"Incorrect type for parameter {repr(param)}. Expected type {expected_type_description}, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:simple"
return result
def standardize_string(input_string: str):
# This function standardizes the string by removing all the spaces, ",./-_*^" punctuation, and converting it to lowercase
# It will also convert all the single quotes to double quotes
# This is used to compare the model output with the possible answers
# We don't want to punish model for answer like April 1, 2024 vs April 1,2024, vs April 1 2024
regex_string = r"[ \,\.\/\-\_\*\^]"
return re.sub(regex_string, "", input_string).lower().replace("'", '"')
def string_checker(param: str, model_output: str, possible_answer: list):
standardize_possible_answer = []
standardize_model_output = standardize_string(model_output)
for i in range(len(possible_answer)):
if type(possible_answer[i]) == str:
standardize_possible_answer.append(standardize_string(possible_answer[i]))
if standardize_model_output not in standardize_possible_answer:
return {
"valid": False,
"error": [
f"Invalid value for parameter {repr(param)}: {repr(model_output)}. Expected one of {possible_answer}. Case insensitive."
],
"error_type": "value_error:string",
}
return {"valid": True, "error": []}
def list_checker(param: str, model_output: list, possible_answer: list):
# Convert the tuple to a list
standardize_model_output = list(model_output)
# If the element in the list is a string, we need to standardize it
for i in range(len(standardize_model_output)):
if type(standardize_model_output[i]) == str:
standardize_model_output[i] = standardize_string(model_output[i])
standardize_possible_answer: Any = []
# We also need to standardize the possible answers
for i in range(len(possible_answer)):
standardize_possible_answer.append([])
for j in range(len(possible_answer[i])):
if type(possible_answer[i][j]) == str:
standardize_possible_answer[i].append(standardize_string(possible_answer[i][j]))
else:
standardize_possible_answer[i].append(possible_answer[i][j])
if standardize_model_output not in standardize_possible_answer:
return {
"valid": False,
"error": [
f"Invalid value for parameter {repr(param)}: {repr(model_output)}. Expected one of {possible_answer}."
],
"error_type": "value_error:list/tuple",
}
return {"valid": True, "error": []}
def dict_checker(param: str, model_output: dict, possible_answers: list):
# This function works for simple dictionaries, but not dictionaries with nested dictionaries.
# The current dataset only contains simple dictionaries, so this is sufficient.
result = {"valid": False, "error": [], "error_type": "dict_checker:unclear"}
for i in range(len(possible_answers)):
if possible_answers[i] == "":
continue
result = {"valid": False, "error": [], "error_type": "dict_checker:unclear"}
flag = True
possible_answer = possible_answers[i]
# possible_anwer is a single dictionary
for key, value in model_output.items():
if key not in possible_answer:
result["valid"] = False
result["error"].append(f"Unexpected dict key parameter: '{key}'.") # type: ignore[attr-defined]
result["error_type"] = "value_error:dict_key"
flag = False
break
standardize_value = value
# If the value is a string, we need to standardize it
if type(value) == str:
standardize_value = standardize_string(value)
# We also need to standardize the possible answers if they are string
standardize_possible_answer = []
for i in range(len(possible_answer[key])):
if type(possible_answer[key][i]) == str:
standardize_possible_answer.append(standardize_string(possible_answer[key][i]))
else:
standardize_possible_answer.append(possible_answer[key][i])
if standardize_value not in standardize_possible_answer:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Invalid value for parameter {repr(key)}: {repr(value)}. Expected one of {standardize_possible_answer}."
)
result["error_type"] = "value_error:dict_value"
flag = False
break
for key, value in possible_answer.items():
if key not in model_output and "" not in value:
result["valid"] = False
result["error"].append(f"Missing dict key parameter: '{key}'.") # type: ignore[attr-defined]
result["error_type"] = "value_error:dict_key"
flag = False
break
if flag:
return {"valid": True, "error": []}
return result
def list_dict_checker(param: str, model_output: list, possible_answers: list):
# This function takes in a list of dictionaries and checks if each dictionary is valid
# The order of the dictionaries in the list must match the order of the possible answers
result = {"valid": False, "error": [], "error_type": "list_dict_checker:unclear"}
for answer_index in range(len(possible_answers)):
flag = True # True means so far, all dictionaries are valid
# Only proceed if the number of dictionaries in the list matches the number of dictionaries in the possible answers
if len(model_output) != len(possible_answers[answer_index]):
result["valid"] = False
result["error"] = ["Wrong number of dictionaries in the list."]
result["error_type"] = "value_error:list_dict_count"
flag = False
continue
for dict_index in range(len(model_output)):
result = dict_checker(
param,
model_output[dict_index],
[possible_answers[answer_index][dict_index]],
)
if not result["valid"]:
flag = False
break
if flag:
return {"valid": True, "error": []}
return result
def simple_function_checker(
func_description: dict,
model_output: dict,
possible_answer: dict,
language: str,
model_name: str,
):
possible_answer = list(possible_answer.values())[0]
# Extract function name and parameters details
func_name = func_description["name"]
param_details = func_description["parameters"]["properties"]
required_params = func_description["parameters"]["required"]
# Initialize a result dictionary
result = {
"valid": True,
"error": [],
"error_type": "simple_function_checker:unclear",
}
# Check if function name matches
if func_name not in model_output:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Function name {repr(func_name)} not found in model output."
)
result["error_type"] = "simple_function_checker:wrong_func_name"
return result
model_params = model_output[func_name]
# Check for required parameters in model output
for param in required_params:
if param not in model_params:
result["valid"] = False
result["error"].append(f"Missing required parameter: {repr(param)}.") # type: ignore[attr-defined]
result["error_type"] = "simple_function_checker:missing_required"
return result
# Validate types and values for each parameter in model output
for param, value in model_params.items():
if param not in param_details or param not in possible_answer:
result["valid"] = False
result["error"].append(f"Unexpected parameter: {repr(param)}.") # type: ignore[attr-defined]
result["error_type"] = "simple_function_checker:unexpected_param"
return result
full_param_details = param_details[param]
expected_type_description = full_param_details["type"] # This is a string
is_variable = False
nested_type_converted = None
if language == "Java":
from evals.utils.bfcl.java_type_converter import java_type_converter
expected_type_converted = JAVA_TYPE_CONVERSION[expected_type_description]
if expected_type_description in JAVA_TYPE_CONVERSION:
if type(value) != str:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Incorrect type for parameter {repr(param)}. Expected type String, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:java"
return result
if expected_type_description in NESTED_CONVERSION_TYPE_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = JAVA_TYPE_CONVERSION[nested_type]
value = java_type_converter(value, expected_type_description, nested_type)
else:
value = java_type_converter(value, expected_type_description)
elif language == "JavaScript":
from evals.utils.bfcl.js_type_converter import js_type_converter
expected_type_converted = JS_TYPE_CONVERSION[expected_type_description]
if expected_type_description in JS_TYPE_CONVERSION:
if type(value) != str:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Incorrect type for parameter {repr(param)}. Expected type String, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:js"
return result
if expected_type_description in NESTED_CONVERSION_TYPE_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = JS_TYPE_CONVERSION[nested_type]
value = js_type_converter(value, expected_type_description, nested_type)
else:
value = js_type_converter(value, expected_type_description)
elif language == "Python":
expected_type_converted = PYTHON_TYPE_MAPPING[expected_type_description]
if expected_type_description in PYTHON_NESTED_TYPE_CHECK_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = PYTHON_TYPE_MAPPING[nested_type]
# We convert all tuple value to list when the expected type is tuple.
# The conversion is necessary because any tuple in the possible answer would become a list after being processed through json.dump() and json.load().
# This does introduce some false positive (eg, when the model provides a list value instead of tuple). We hope to find a better solution in the future.
if expected_type_description == "tuple" and type(value) == tuple:
value = list(value)
# Allow python auto conversion from int to float
if language == "Python" and expected_type_description == "float" and type(value) == int:
value = float(value)
# Type checking
# In fact, we only check for Python here.
# Type check for other languages are handled by the type converter, and so their value (after conversion) is always correct.
type_check_result = type_checker(
param,
value,
possible_answer[param],
expected_type_description,
expected_type_converted,
nested_type_converted,
)
is_variable = type_check_result["is_variable"]
if not type_check_result["valid"]:
return type_check_result
# It doesn't make sense to special handle dictionaries and list of dictionaries if the value is a variable.
# We can just treat the variable as a string and use the normal flow.
if not is_variable:
# Special handle for dictionaries
if expected_type_converted == dict:
result = dict_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Special handle for list of dictionaries
elif expected_type_converted == list and nested_type_converted == dict:
result = list_dict_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Special handle for strings
elif expected_type_converted == str:
# We don't check for case sensitivity for string, as long as it's not a variable
result = string_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
elif expected_type_converted == list:
result = list_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Check if the value is within the possible answers
if value not in possible_answer[param]:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Invalid value for parameter {repr(param)}: {repr(value)}. Expected one of {possible_answer[param]}."
)
result["error_type"] = "value_error:others"
return result
# Check for optional parameters not provided but allowed
for param in possible_answer:
if param not in model_params and "" not in possible_answer[param]:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Optional parameter {repr(param)} not provided and not marked as optional."
)
result["error_type"] = "simple_function_checker:missing_optional"
return result
return result
def parallel_function_checker_enforce_order(
func_descriptions: list,
model_output: list,
possible_answers: dict,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "parallel_function_checker_enforce_order:wrong_count",
}
func_name_list = list(possible_answers.keys())
possible_answers_list = []
for key, value in possible_answers.items():
possible_answers_list.append({key: value})
for i in range(len(possible_answers_list)):
func_description = find_description(func_descriptions, func_name_list[i])
result = simple_function_checker(
func_description,
model_output[i],
possible_answers_list[i],
language,
model_name,
)
if not result["valid"]:
return result
return {"valid": True, "error": []}
def parallel_function_checker_no_order(
func_descriptions: list,
model_output: list,
possible_answers: list,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "parallel_function_checker_no_order:wrong_count",
}
matched_indices = []
# We go throught the possible answers one by one, and eliminate the model output that matches the possible answer
# It must be this way because we need ground truth to fetch the correct function description
for i in range(len(possible_answers)):
# possible_answers[i] is a dictionary with only one key
func_name_expected = list(possible_answers[i].keys())[0]
func_description = find_description(func_descriptions, func_name_expected)
all_errors = []
for index in range(len(model_output)):
if index in matched_indices:
continue
result = simple_function_checker(
func_description,
model_output[index],
possible_answers[i],
language,
model_name,
)
if result["valid"]:
matched_indices.append(index)
break
else:
all_errors.append(
{
f"Model Result Index {index}": {
"sub_error": result["error"],
"sub_error_type": result["error_type"],
"model_output_item": model_output[index],
"possible_answer_item": possible_answers[i],
}
}
)
if not result["valid"]:
considered_indices = [i for i in range(len(model_output)) if i not in matched_indices]
all_errors.insert(
0,
f"Could not find a matching function among index {considered_indices} of model output for index {i} of possible answers.", # type: ignore[arg-type]
)
return {
"valid": False,
"error": all_errors,
"error_type": "parallel_function_checker_no_order:cannot_find_match",
}
return {"valid": True, "error": []}
def multiple_function_checker(
func_descriptions: list,
model_output: list,
possible_answers: list,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "multiple_function_checker:wrong_count",
}
# possible_answers is a list of only one dictionary with only one key
func_name_expected = list(possible_answers[0].keys())[0]
func_description = find_description(func_descriptions, func_name_expected)
return simple_function_checker(
func_description,
model_output[0],
possible_answers[0],
language,
model_name,
)
def patten_matcher(exec_output, expected_result, function_call, is_sanity_check):
result = {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
if type(exec_output) != type(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result type for {repr(function_call)}. Expected type: {type(expected_result)}, but got: {type(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type",
"model_executed_output": exec_output,
}
if type(exec_output) == dict:
# We loose the requirement for the sanity check as the expected result used in the sanity check might not be the most up-to-date one.
# This happens when the key is a timestamp or a random number.
if is_sanity_check:
if len(exec_output) != len(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but wrong number of elements in the output. Expected length: {len(expected_result)}, but got: {len(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type:dict_length",
"model_executed_output": exec_output,
}
else:
return result
for key, value in expected_result.items():
if key not in exec_output:
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but key {repr(key)} not found in the model output."
],
"error_type": "executable_checker:wrong_result_type:dict_key_not_found",
"model_executed_output": exec_output,
}
for key, value in exec_output.items():
if key not in expected_result:
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but key {repr(key)} not expected in the model output."
],
"error_type": "executable_checker:wrong_result_type:dict_extra_key",
"model_executed_output": exec_output,
}
if type(exec_output) == list:
if len(exec_output) != len(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type list, but wrong number of elements in the output. Expected length: {len(expected_result)}, but got: {len(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type:list_length",
"model_executed_output": exec_output,
}
return result
#### Helper functions for Exec ####
def executable_checker_simple(
function_call: str,
expected_result,
expected_result_type: str,
is_sanity_check=False,
):
result = {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
exec_dict: Any = {}
try:
exec(
"from executable_python_function import *" + "\nresult=" + function_call,
exec_dict,
)
exec_output = exec_dict["result"]
except NoAPIKeyError as e:
raise e
except Exception as e:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Error in execution: {repr(function_call)}. Error: {str(e)}"
)
result["error_type"] = "executable_checker:execution_error"
return result
# We need to special handle the case where the execution result is a tuple and convert it to a list
# Because when json is stored, the tuple is converted to a list, and so the expected result is a list when loaded from json
if isinstance(exec_output, tuple):
exec_output = list(exec_output)
if expected_result_type == "exact_match":
if exec_output != expected_result:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}."
)
result["error_type"] = "executable_checker:wrong_result"
result["model_executed_output"] = exec_output
return result
elif expected_result_type == "real_time_match":
# Allow for 5% difference
if (type(expected_result) == float or type(expected_result) == int) and (
type(exec_output) == float or type(exec_output) == int
):
if not (
expected_result * (1 - REAL_TIME_MATCH_ALLOWED_DIFFERENCE)
<= exec_output
<= expected_result * (1 + REAL_TIME_MATCH_ALLOWED_DIFFERENCE)
):
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}. {REAL_TIME_MATCH_ALLOWED_DIFFERENCE * 100}% difference allowed."
)
result["error_type"] = "executable_checker:wrong_result_real_time"
result["model_executed_output"] = exec_output
return result
else:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}. Type needs to be float or int for real time match criteria."
)
result["error_type"] = "executable_checker:wrong_result_real_time"
result["model_executed_output"] = exec_output
return result
else:
# structural match
pattern_match_result = patten_matcher(exec_output, expected_result, function_call, is_sanity_check)
if not pattern_match_result["valid"]:
return pattern_match_result
return result
def executable_checker_parallel_no_order(
decoded_result: list, expected_exec_result: list, expected_exec_result_type: list
):
if len(decoded_result) != len(expected_exec_result):
return {
"valid": False,
"error": [
f"Wrong number of functions provided. Expected {len(expected_exec_result)}, but got {len(decoded_result)}."
],
"error_type": "value_error:exec_result_count",
}
matched_indices = []
for i in range(len(expected_exec_result)):
all_errors = []
for index in range(len(decoded_result)):
if index in matched_indices:
continue
result = executable_checker_simple(
decoded_result[index],
expected_exec_result[i],
expected_exec_result_type[i],
False,
)
if result["valid"]:
matched_indices.append(index)
break
else:
all_errors.append(
{
f"Model Result Index {index}": {
"sub_error": result["error"],
"sub_error_type": result["error_type"],
"model_executed_output": (
result["model_executed_output"] if "model_executed_output" in result else None
),
}
}
)
if not result["valid"]:
considered_indices = [i for i in range(len(decoded_result)) if i not in matched_indices]
all_errors.insert(
0,
f"Could not find a matching function among index {considered_indices} of model output for index {i} of possible answers.", # type: ignore[arg-type]
)
return {
"valid": False,
"error": all_errors,
"error_type": "executable_checker:cannot_find_match",
}
return {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
#### Main function ####
def executable_checker_rest(func_call, idx):
# Move this here for now to avoid needing to read this file / fix paths to be relative to dataset_dir. Fix when it's actually needed / used.
EVAL_GROUND_TRUTH_PATH = "/mnt/wsfuse/fair_llm_v2/datasets/eval/bfcl/rest-eval-response_v5.jsonl" # Ground truth file for v5 for rest execution
with open(EVAL_GROUND_TRUTH_PATH, "r") as f:
EVAL_GROUND_TRUTH = f.readlines()
if "https://geocode.maps.co" in func_call:
time.sleep(2)
if "requests_get" in func_call:
func_call = func_call.replace("requests_get", "requests.get")
try:
response = eval(func_call)
except Exception as e:
return {
"valid": False,
"error": [f"Execution failed. {str(e)}"],
"error_type": "executable_checker_rest:execution_error",
}
try:
if response.status_code == 200:
eval_GT_json = json.loads(EVAL_GROUND_TRUTH[idx])
try:
if isinstance(eval_GT_json, dict):
if isinstance(response.json(), dict):
if set(eval_GT_json.keys()) == set(response.json().keys()):
return {"valid": True, "error": [], "error_type": ""}
return {
"valid": False,
"error": ["Key inconsistency"],
"error_type": "executable_checker_rest:wrong_key",
}
return {
"valid": False,
"error": [f"Expected dictionary, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
elif isinstance(eval_GT_json, list):
if isinstance(response.json(), list):
if len(eval_GT_json) != len(response.json()):
return {
"valid": False,
"error": [f"Response list length inconsistency."],
"error_type": "value_error:exec_result_rest_count",
}
else:
for i in range(len(eval_GT_json)):
if set(eval_GT_json[i].keys()) != set(response.json()[i].keys()):
return {
"valid": False,
"error": [f"Key inconsistency"],
"error_type": "executable_checker_rest:wrong_key",
}
return {"valid": True, "error": []}
else:
return {
"valid": False,
"error": [f"Expected list, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
return {
"valid": False,
"error": [f"Expected dict or list, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
except Exception as e:
return {
"valid": False,
"error": [
f"Error in execution and type checking. Status code: {response.status_code}. Error: {str(e)}"
],
"error_type": "executable_checker_rest:response_format_error",
}
else:
return {
"valid": False,
"error": [f"Execution result status code is not 200, got {response.status_code}"],
"error_type": "executable_checker_rest:wrong_status_code",
}
except Exception as e:
return {
"valid": False,
"error": [f"Cannot get status code of the response. Error: {str(e)}"],
"error_type": "executable_checker_rest:cannot_get_status_code",
}
def ast_checker(func_description, model_output, possible_answer, language, test_category, model_name):
if "parallel" in test_category:
return parallel_function_checker_no_order(func_description, model_output, possible_answer, language, model_name)
elif "multiple" in test_category:
return multiple_function_checker(func_description, model_output, possible_answer, language, model_name)
else:
if len(model_output) != 1:
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "simple_function_checker:wrong_count",
}
return simple_function_checker(
func_description[0],
model_output[0],
possible_answer[0],
language,
model_name,
)
def exec_checker(decoded_result: list, func_description: dict, test_category: str):
if "multiple" in test_category or "parallel" in test_category:
return executable_checker_parallel_no_order(
decoded_result,
func_description["execution_result"],
func_description["execution_result_type"],
)
else:
if len(decoded_result) != 1:
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "simple_exec_checker:wrong_count",
}
return executable_checker_simple(
decoded_result[0],
func_description["execution_result"][0],
func_description["execution_result_type"][0],
False,
)
def is_empty_output(decoded_output):
# This function is a patch to the ast decoder for relevance detection
# Sometimes the ast decoder will parse successfully, but the input doens't really have a function call
# [], [{}], and anything that is not in function calling format is considered empty (and thus should be marked as correct)
if not is_function_calling_format_output(decoded_output):
return True
if len(decoded_output) == 0:
return True
if len(decoded_output) == 1 and len(decoded_output[0]) == 0:
return True
def is_function_calling_format_output(decoded_output):
# Ensure the output is a list of dictionaries
if type(decoded_output) == list:
for item in decoded_output:
if type(item) != dict:
return False
return True
return False

40
llama_stack/providers/inline/scoring/basic/utils/bfcl/tree_sitter.py Normal file
View file

@ -0,0 +1,40 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Tree-sitter changes its API with unfortunate frequency. Modules that need it should
import it from here so that we can centrally manage things as necessary.
"""
# These currently work with tree-sitter 0.23.0
# NOTE: Don't import tree-sitter or any of the language modules in the main module
# because not all environments have them. Import lazily inside functions where needed.
import importlib
import typing
if typing.TYPE_CHECKING:
import tree_sitter
def get_language(language: str) -> "tree_sitter.Language":
import tree_sitter
language_module_name = f"tree_sitter_{language}"
try:
language_module = importlib.import_module(language_module_name)
except ModuleNotFoundError as exc:
raise ValueError(
f"Language {language} is not found. Please install the tree-sitter-{language} package."
) from exc
return tree_sitter.Language(language_module.language())
def get_parser(language: str, **kwargs) -> "tree_sitter.Parser":
import tree_sitter
lang = get_language(language)
return tree_sitter.Parser(lang, **kwargs)

3319
llama_stack/providers/inline/scoring/basic/utils/ifeval_utils.py Normal file
View file

File diff suppressed because it is too large Load diff

6
llama_stack/providers/inline/scoring/braintrust/braintrust.py
View file

@ -167,11 +167,11 @@ class BraintrustScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(input_rows=all_rows.rows, scoring_functions=scoring_functions)
res = await self.score(input_rows=all_rows.data, scoring_functions=scoring_functions)
if save_results_dataset:
# TODO: persist and register dataset on to server for reading
# self.datasets_api.register_dataset()

7
llama_stack/providers/inline/scoring/llm_as_judge/config.py
View file

@ -3,7 +3,12 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
class LlmAsJudgeScoringConfig(BaseModel): ...
class LlmAsJudgeScoringConfig(BaseModel):
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {}

6
llama_stack/providers/inline/scoring/llm_as_judge/scoring.py
View file

@ -72,12 +72,12 @@ class LlmAsJudgeScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
)
if save_results_dataset:

4
llama_stack/providers/inline/telemetry/meta_reference/__init__.py
View file

@ -6,12 +6,14 @@
from typing import Any, Dict
from llama_stack.distribution.datatypes import Api
from .config import TelemetryConfig, TelemetrySink
__all__ = ["TelemetryConfig", "TelemetrySink"]
async def get_provider_impl(config: TelemetryConfig, deps: Dict[str, Any]):
async def get_provider_impl(config: TelemetryConfig, deps: Dict[Api, Any]):
from .telemetry import TelemetryAdapter
impl = TelemetryAdapter(config, deps)

16
llama_stack/providers/inline/telemetry/meta_reference/config.py
View file

@ -13,18 +13,24 @@ from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
class TelemetrySink(str, Enum):
OTEL = "otel"
OTEL_TRACE = "otel_trace"
OTEL_METRIC = "otel_metric"
SQLITE = "sqlite"
CONSOLE = "console"
class TelemetryConfig(BaseModel):
otel_endpoint: str = Field(
otel_trace_endpoint: str = Field(
default="http://localhost:4318/v1/traces",
description="The OpenTelemetry collector endpoint URL",
description="The OpenTelemetry collector endpoint URL for traces",
)
otel_metric_endpoint: str = Field(
default="http://localhost:4318/v1/metrics",
description="The OpenTelemetry collector endpoint URL for metrics",
)
service_name: str = Field(
default="llama-stack",
# service name is always the same, use zero-width space to avoid clutter
default="​",
description="The service name to use for telemetry",
)
sinks: List[TelemetrySink] = Field(
@ -46,7 +52,7 @@ class TelemetryConfig(BaseModel):
@classmethod
def sample_run_config(cls, __distro_dir__: str, db_name: str = "trace_store.db") -> Dict[str, Any]:
return {
"service_name": "${env.OTEL_SERVICE_NAME:llama-stack}",
"service_name": "${env.OTEL_SERVICE_NAME:​}",
"sinks": "${env.TELEMETRY_SINKS:console,sqlite}",
"sqlite_db_path": "${env.SQLITE_DB_PATH:" + __distro_dir__ + "/" + db_name + "}",
}

10
llama_stack/providers/inline/telemetry/meta_reference/console_span_processor.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
import json
from datetime import datetime
from datetime import datetime, timezone
from opentelemetry.sdk.trace import ReadableSpan
from opentelemetry.sdk.trace.export import SpanProcessor
@ -34,7 +34,7 @@ class ConsoleSpanProcessor(SpanProcessor):
if span.attributes and span.attributes.get("__autotraced__"):
return
timestamp = datetime.utcfromtimestamp(span.start_time / 1e9).strftime("%H:%M:%S.%f")[:-3]
timestamp = datetime.fromtimestamp(span.start_time / 1e9, tz=timezone.utc).strftime("%H:%M:%S.%f")[:-3]
print(
f"{COLORS['dim']}{timestamp}{COLORS['reset']} "
@ -46,7 +46,7 @@ class ConsoleSpanProcessor(SpanProcessor):
if span.attributes and span.attributes.get("__autotraced__"):
return
timestamp = datetime.utcfromtimestamp(span.end_time / 1e9).strftime("%H:%M:%S.%f")[:-3]
timestamp = datetime.fromtimestamp(span.end_time / 1e9, tz=timezone.utc).strftime("%H:%M:%S.%f")[:-3]
span_context = (
f"{COLORS['dim']}{timestamp}{COLORS['reset']} "
@ -74,7 +74,7 @@ class ConsoleSpanProcessor(SpanProcessor):
print(f" {COLORS['dim']}{key}: {str_value}{COLORS['reset']}")
for event in span.events:
event_time = datetime.utcfromtimestamp(event.timestamp / 1e9).strftime("%H:%M:%S.%f")[:-3]
event_time = datetime.fromtimestamp(event.timestamp / 1e9, tz=timezone.utc).strftime("%H:%M:%S.%f")[:-3]
severity = event.attributes.get("severity", "info")
message = event.attributes.get("message", event.name)
@ -101,6 +101,6 @@ class ConsoleSpanProcessor(SpanProcessor):
"""Shutdown the processor."""
pass
def force_flush(self, timeout_millis: float = None) -> bool:
def force_flush(self, timeout_millis: float | None = None) -> bool:
"""Force flush any pending spans."""
return True

21
llama_stack/providers/inline/telemetry/meta_reference/sqlite_span_processor.py
View file

@ -8,10 +8,11 @@ import json
import os
import sqlite3
import threading
from datetime import datetime
from datetime import datetime, timezone
from opentelemetry.sdk.trace import SpanProcessor
from opentelemetry.trace import Span
from opentelemetry.trace.span import format_span_id, format_trace_id
class SQLiteSpanProcessor(SpanProcessor):
@ -100,14 +101,14 @@ class SQLiteSpanProcessor(SpanProcessor):
conn = self._get_connection()
cursor = conn.cursor()
trace_id = format(span.get_span_context().trace_id, "032x")
span_id = format(span.get_span_context().span_id, "016x")
trace_id = format_trace_id(span.get_span_context().trace_id)
span_id = format_span_id(span.get_span_context().span_id)
service_name = span.resource.attributes.get("service.name", "unknown")
parent_span_id = None
parent_context = span.parent
if parent_context:
parent_span_id = format(parent_context.span_id, "016x")
parent_span_id = format_span_id(parent_context.span_id)
# Insert into traces
cursor.execute(
@ -123,9 +124,9 @@ class SQLiteSpanProcessor(SpanProcessor):
(
trace_id,
service_name,
(span_id if not parent_span_id else None),
datetime.fromtimestamp(span.start_time / 1e9).isoformat(),
datetime.fromtimestamp(span.end_time / 1e9).isoformat(),
(span_id if span.attributes.get("__root_span__") == "true" else None),
datetime.fromtimestamp(span.start_time / 1e9, timezone.utc).isoformat(),
datetime.fromtimestamp(span.end_time / 1e9, timezone.utc).isoformat(),
),
)
@ -143,8 +144,8 @@ class SQLiteSpanProcessor(SpanProcessor):
trace_id,
parent_span_id,
span.name,
datetime.fromtimestamp(span.start_time / 1e9).isoformat(),
datetime.fromtimestamp(span.end_time / 1e9).isoformat(),
datetime.fromtimestamp(span.start_time / 1e9, timezone.utc).isoformat(),
datetime.fromtimestamp(span.end_time / 1e9, timezone.utc).isoformat(),
json.dumps(dict(span.attributes)),
span.status.status_code.name,
span.kind.name,
@ -161,7 +162,7 @@ class SQLiteSpanProcessor(SpanProcessor):
(
span_id,
event.name,
datetime.fromtimestamp(event.timestamp / 1e9).isoformat(),
datetime.fromtimestamp(event.timestamp / 1e9, timezone.utc).isoformat(),
json.dumps(dict(event.attributes)),
),
)

51
llama_stack/providers/inline/telemetry/meta_reference/telemetry.py
View file

@ -44,7 +44,7 @@ from llama_stack.providers.utils.telemetry.sqlite_trace_store import SQLiteTrace
from .config import TelemetryConfig, TelemetrySink
_GLOBAL_STORAGE = {
_GLOBAL_STORAGE: dict[str, dict[str | int, Any]] = {
"active_spans": {},
"counters": {},
"gauges": {},
@ -54,25 +54,16 @@ _global_lock = threading.Lock()
_TRACER_PROVIDER = None
def string_to_trace_id(s: str) -> int:
# Convert the string to bytes and then to an integer
return int.from_bytes(s.encode(), byteorder="big", signed=False)
def string_to_span_id(s: str) -> int:
# Use only the first 8 bytes (64 bits) for span ID
return int.from_bytes(s.encode()[:8], byteorder="big", signed=False)
def is_tracing_enabled(tracer):
with tracer.start_as_current_span("check_tracing") as span:
return span.is_recording()
class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
def __init__(self, config: TelemetryConfig, deps: Dict[str, Any]) -> None:
def __init__(self, config: TelemetryConfig, deps: Dict[Api, Any]) -> None:
self.config = config
self.datasetio_api = deps.get(Api.datasetio)
self.meter = None
resource = Resource.create(
{
@ -90,15 +81,16 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
provider = TracerProvider(resource=resource)
trace.set_tracer_provider(provider)
_TRACER_PROVIDER = provider
if TelemetrySink.OTEL in self.config.sinks:
otlp_exporter = OTLPSpanExporter(
endpoint=self.config.otel_endpoint,
if TelemetrySink.OTEL_TRACE in self.config.sinks:
span_exporter = OTLPSpanExporter(
endpoint=self.config.otel_trace_endpoint,
)
span_processor = BatchSpanProcessor(otlp_exporter)
span_processor = BatchSpanProcessor(span_exporter)
trace.get_tracer_provider().add_span_processor(span_processor)
if TelemetrySink.OTEL_METRIC in self.config.sinks:
metric_reader = PeriodicExportingMetricReader(
OTLPMetricExporter(
endpoint=self.config.otel_endpoint,
endpoint=self.config.otel_metric_endpoint,
)
)
metric_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
@ -108,7 +100,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
if TelemetrySink.CONSOLE in self.config.sinks:
trace.get_tracer_provider().add_span_processor(ConsoleSpanProcessor())
if TelemetrySink.OTEL in self.config.sinks:
if TelemetrySink.OTEL_METRIC in self.config.sinks:
self.meter = metrics.get_meter(__name__)
if TelemetrySink.SQLITE in self.config.sinks:
self.trace_store = SQLiteTraceStore(self.config.sqlite_db_path)
@ -134,7 +126,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
def _log_unstructured(self, event: UnstructuredLogEvent, ttl_seconds: int) -> None:
with self._lock:
# Use global storage instead of instance storage
span_id = string_to_span_id(event.span_id)
span_id = event.span_id
span = _GLOBAL_STORAGE["active_spans"].get(span_id)
if span:
@ -145,7 +137,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
"message": event.message,
"severity": event.severity.value,
"__ttl__": ttl_seconds,
**event.attributes,
**(event.attributes or {}),
},
timestamp=timestamp_ns,
)
@ -153,6 +145,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
print(f"Warning: No active span found for span_id {span_id}. Dropping event: {event}")
def _get_or_create_counter(self, name: str, unit: str) -> metrics.Counter:
assert self.meter is not None
if name not in _GLOBAL_STORAGE["counters"]:
_GLOBAL_STORAGE["counters"][name] = self.meter.create_counter(
name=name,
@ -162,6 +155,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
return _GLOBAL_STORAGE["counters"][name]
def _get_or_create_gauge(self, name: str, unit: str) -> metrics.ObservableGauge:
assert self.meter is not None
if name not in _GLOBAL_STORAGE["gauges"]:
_GLOBAL_STORAGE["gauges"][name] = self.meter.create_gauge(
name=name,
@ -171,6 +165,8 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
return _GLOBAL_STORAGE["gauges"][name]
def _log_metric(self, event: MetricEvent) -> None:
if self.meter is None:
return
if isinstance(event.value, int):
counter = self._get_or_create_counter(event.metric, event.unit)
counter.add(event.value, attributes=event.attributes)
@ -179,6 +175,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
up_down_counter.add(event.value, attributes=event.attributes)
def _get_or_create_up_down_counter(self, name: str, unit: str) -> metrics.UpDownCounter:
assert self.meter is not None
if name not in _GLOBAL_STORAGE["up_down_counters"]:
_GLOBAL_STORAGE["up_down_counters"][name] = self.meter.create_up_down_counter(
name=name,
@ -189,8 +186,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
def _log_structured(self, event: StructuredLogEvent, ttl_seconds: int) -> None:
with self._lock:
span_id = string_to_span_id(event.span_id)
trace_id = string_to_trace_id(event.trace_id)
span_id = int(event.span_id, 16)
tracer = trace.get_tracer(__name__)
if event.attributes is None:
event.attributes = {}
@ -201,14 +197,13 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
if span_id in _GLOBAL_STORAGE["active_spans"]:
return
parent_span = None
context = None
if event.payload.parent_span_id:
parent_span_id = string_to_span_id(event.payload.parent_span_id)
parent_span_id = int(event.payload.parent_span_id, 16)
parent_span = _GLOBAL_STORAGE["active_spans"].get(parent_span_id)
context = trace.Context(trace_id=trace_id)
if parent_span:
context = trace.set_span_in_context(parent_span, context)
context = trace.set_span_in_context(parent_span)
else:
event.attributes["__root_span__"] = "true"
span = tracer.start_span(
name=event.payload.name,

17
llama_stack/providers/inline/telemetry/sample/__init__.py
View file

@ -1,17 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any
from .config import SampleConfig
async def get_adapter_impl(config: SampleConfig, _deps) -> Any:
from .sample import SampleTelemetryImpl
impl = SampleTelemetryImpl(config)
await impl.initialize()
return impl

17
llama_stack/providers/inline/telemetry/sample/sample.py
View file

@ -1,17 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.apis.telemetry import Telemetry
from .config import SampleConfig
class SampleTelemetryImpl(Telemetry):
def __init__(self, config: SampleConfig):
self.config = config
async def initialize(self):
pass

4
llama_stack/providers/inline/tool_runtime/code_interpreter/code_env_prefix.py
View file

@ -69,7 +69,7 @@ def popen_not_allowed(*args, **kwargs):
)
_subprocess.Popen = popen_not_allowed
_subprocess.Popen = popen_not_allowed # type: ignore
import atexit as _atexit
@ -104,7 +104,7 @@ def _open_connections():
return _NETWORK_CONNECTIONS
_builtins._open_connections = _open_connections
_builtins._open_connections = _open_connections # type: ignore
@_atexit.register

16
llama_stack/providers/inline/tool_runtime/code_interpreter/code_execution.py
View file

@ -76,6 +76,7 @@ class CodeExecutionRequest:
only_last_cell_fail: bool = True
seed: int = 0
strip_fpaths_in_stderr: bool = True
use_bwrap: bool = True
class CodeExecutor:
@ -103,8 +104,6 @@ _set_seeds()\
script = "\n\n".join([seeds_prefix] + [CODE_ENV_PREFIX] + scripts)
with tempfile.TemporaryDirectory() as dpath:
bwrap_prefix = "bwrap " + generate_bwrap_command(bind_dirs=[dpath])
cmd = [*bwrap_prefix.split(), sys.executable, "-c", script]
code_fpath = os.path.join(dpath, "code.py")
with open(code_fpath, "w") as f:
f.write(script)
@ -118,6 +117,13 @@ _set_seeds()\
MPLBACKEND="module://matplotlib_custom_backend",
PYTHONPATH=f"{DIRNAME}:{python_path}",
)
if req.use_bwrap:
bwrap_prefix = "bwrap " + generate_bwrap_command(bind_dirs=[dpath])
cmd = [*bwrap_prefix.split(), sys.executable, "-c", script]
else:
cmd = [sys.executable, "-c", script]
stdout, stderr, returncode = do_subprocess(
cmd=cmd,
env=env,
@ -155,14 +161,14 @@ _set_seeds()\
def process_matplotlib_response(response, matplotlib_dump_dir: str):
image_data = response["image_data"]
# Convert the base64 string to a bytes object
images = [base64.b64decode(d["image_base64"]) for d in image_data]
images_raw = [base64.b64decode(d["image_base64"]) for d in image_data]
# Create a list of PIL images from the bytes objects
images = [Image.open(BytesIO(img)) for img in images]
images = [Image.open(BytesIO(img)) for img in images_raw]
# Create a list of image paths
image_paths = []
for i, img in enumerate(images):
# create new directory for each day to better organize data:
dump_dname = datetime.today().strftime("%Y-%m-%d")
dump_dname = datetime.today().strftime("%Y-%m-%d") # noqa: DTZ002 - we don't care about timezones here since we are displaying the date
dump_dpath = Path(matplotlib_dump_dir, dump_dname)
dump_dpath.mkdir(parents=True, exist_ok=True)
# save image into a file

43
llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
View file

@ -5,12 +5,15 @@
# the root directory of this source tree.
import asyncio
import logging
import os
import tempfile
from typing import Any, Dict, List, Optional
from typing import Any, Dict, Optional
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.tools import (
ListToolDefsResponse,
Tool,
ToolDef,
ToolInvocationResult,
@ -36,7 +39,7 @@ class CodeInterpreterToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
async def initialize(self):
pass
async def register_tool(self, tool: Tool):
async def register_tool(self, tool: Tool) -> None:
pass
async def unregister_tool(self, tool_id: str) -> None:
@ -44,25 +47,29 @@ class CodeInterpreterToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
async def list_runtime_tools(
self, tool_group_id: Optional[str] = None, mcp_endpoint: Optional[URL] = None
) -> List[ToolDef]:
return [
ToolDef(
name="code_interpreter",
description="Execute code",
parameters=[
ToolParameter(
name="code",
description="The code to execute",
parameter_type="string",
),
],
)
]
) -> ListToolDefsResponse:
return ListToolDefsResponse(
data=[
ToolDef(
name="code_interpreter",
description="Execute code",
parameters=[
ToolParameter(
name="code",
description="The code to execute",
parameter_type="string",
),
],
)
]
)
async def invoke_tool(self, tool_name: str, kwargs: Dict[str, Any]) -> ToolInvocationResult:
script = kwargs["code"]
req = CodeExecutionRequest(scripts=[script])
res = self.code_executor.execute(req)
# Use environment variable to control bwrap usage
force_disable_bwrap = os.environ.get("DISABLE_CODE_SANDBOX", "").lower() in ("1", "true", "yes")
req = CodeExecutionRequest(scripts=[script], use_bwrap=not force_disable_bwrap)
res = await asyncio.to_thread(self.code_executor.execute, req)
pieces = [res["process_status"]]
for out_type in ["stdout", "stderr"]:
res_out = res[out_type]

6
llama_stack/providers/inline/tool_runtime/code_interpreter/config.py
View file

@ -4,8 +4,12 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
class CodeInterpreterToolConfig(BaseModel):
pass
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {}

2
llama_stack/providers/inline/tool_runtime/rag/__init__.py
View file

@ -11,7 +11,7 @@ from llama_stack.providers.datatypes import Api
from .config import RagToolRuntimeConfig
async def get_provider_impl(config: RagToolRuntimeConfig, deps: Dict[str, Any]):
async def get_provider_impl(config: RagToolRuntimeConfig, deps: Dict[Api, Any]):
from .memory import MemoryToolRuntimeImpl
impl = MemoryToolRuntimeImpl(config, deps[Api.vector_io], deps[Api.inference], deps[Api.preprocessing])

6
llama_stack/providers/inline/tool_runtime/rag/config.py
View file

@ -4,8 +4,12 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
class RagToolRuntimeConfig(BaseModel):
pass
@classmethod
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> Dict[str, Any]:
return {}

53
llama_stack/providers/inline/tool_runtime/rag/memory.py
View file

@ -15,6 +15,7 @@ from pydantic import TypeAdapter
from llama_stack.apis.common.content_types import (
URL,
InterleavedContent,
InterleavedContentItem,
TextContentItem,
)
from llama_stack.apis.inference import Inference
@ -27,10 +28,12 @@ from llama_stack.apis.preprocessing import (
PreprocessorChainElement,
)
from llama_stack.apis.tools import (
ListToolDefsResponse,
RAGDocument,
RAGQueryConfig,
RAGQueryResult,
RAGToolRuntime,
Tool,
ToolDef,
ToolInvocationResult,
ToolParameter,
@ -73,6 +76,12 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
async def shutdown(self):
pass
async def register_tool(self, tool: Tool) -> None:
pass
async def unregister_tool(self, tool_id: str) -> None:
return
async def insert(
self,
documents: List[RAGDocument],
@ -103,7 +112,7 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
actual_chunks = [chunk.data_element_path_or_content for chunk in chunks]
await self.vector_io_api.insert_chunks(
chunks=actual_chunks,
chunks=actual_chunks, # type: ignore
vector_db_id=vector_db_id,
)
@ -140,11 +149,11 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
return RAGQueryResult(content=None)
# sort by score
chunks, scores = zip(*sorted(zip(chunks, scores, strict=False), key=lambda x: x[1], reverse=True), strict=False)
chunks, scores = zip(*sorted(zip(chunks, scores, strict=False), key=lambda x: x[1], reverse=True), strict=False) # type: ignore
chunks = chunks[: query_config.max_chunks]
tokens = 0
picked = [
picked: list[InterleavedContentItem] = [
TextContentItem(
text=f"knowledge_search tool found {len(chunks)} chunks:\nBEGIN of knowledge_search tool results.\n"
)
@ -173,27 +182,29 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
async def list_runtime_tools(
self, tool_group_id: Optional[str] = None, mcp_endpoint: Optional[URL] = None
) -> List[ToolDef]:
) -> ListToolDefsResponse:
# Parameters are not listed since these methods are not yet invoked automatically
# by the LLM. The method is only implemented so things like /tools can list without
# encountering fatals.
return [
ToolDef(
name="insert_into_memory",
description="Insert documents into memory",
),
ToolDef(
name="knowledge_search",
description="Search for information in a database.",
parameters=[
ToolParameter(
name="query",
description="The query to search for. Can be a natural language sentence or keywords.",
parameter_type="string",
),
],
),
]
return ListToolDefsResponse(
data=[
ToolDef(
name="insert_into_memory",
description="Insert documents into memory",
),
ToolDef(
name="knowledge_search",
description="Search for information in a database.",
parameters=[
ToolParameter(
name="query",
description="The query to search for. Can be a natural language sentence or keywords.",
parameter_type="string",
),
],
),
]
)
async def invoke_tool(self, tool_name: str, kwargs: Dict[str, Any]) -> ToolInvocationResult:
vector_db_ids = kwargs.get("vector_db_ids", [])

4
llama_stack/providers/inline/vector_io/chroma/config.py
View file

@ -13,5 +13,5 @@ class ChromaVectorIOConfig(BaseModel):
db_path: str
@classmethod
def sample_config(cls) -> Dict[str, Any]:
return {"db_path": "{env.CHROMADB_PATH}"}
def sample_run_config(cls, db_path: str = "${env.CHROMADB_PATH}", **kwargs: Any) -> Dict[str, Any]:
return {"db_path": db_path}

24
llama_stack/providers/inline/vector_io/faiss/faiss.py
View file

@ -15,11 +15,13 @@ import faiss
import numpy as np
from numpy.typing import NDArray
from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.common.content_types import InterleavedContent
from llama_stack.apis.inference.inference import Inference
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.datatypes import VectorDBsProtocolPrivate
from llama_stack.providers.utils.kvstore import kvstore_impl
from llama_stack.providers.utils.kvstore.api import KVStore
from llama_stack.providers.utils.memory.vector_store import (
EmbeddingIndex,
VectorDBWithIndex,
@ -35,16 +37,14 @@ FAISS_INDEX_PREFIX = f"faiss_index:{VERSION}::"
class FaissIndex(EmbeddingIndex):
chunk_by_index: Dict[int, str]
def __init__(self, dimension: int, kvstore=None, bank_id: str = None):
def __init__(self, dimension: int, kvstore: KVStore | None = None, bank_id: str | None = None):
self.index = faiss.IndexFlatL2(dimension)
self.chunk_by_index = {}
self.chunk_by_index: dict[int, Chunk] = {}
self.kvstore = kvstore
self.bank_id = bank_id
@classmethod
async def create(cls, dimension: int, kvstore=None, bank_id: str = None):
async def create(cls, dimension: int, kvstore: KVStore | None = None, bank_id: str | None = None):
instance = cls(dimension, kvstore, bank_id)
await instance.initialize()
return instance
@ -114,11 +114,11 @@ class FaissIndex(EmbeddingIndex):
class FaissVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
def __init__(self, config: FaissVectorIOConfig, inference_api: Api.inference) -> None:
def __init__(self, config: FaissVectorIOConfig, inference_api: Inference) -> None:
self.config = config
self.inference_api = inference_api
self.cache = {}
self.kvstore = None
self.cache: dict[str, VectorDBWithIndex] = {}
self.kvstore: KVStore | None = None
async def initialize(self) -> None:
self.kvstore = await kvstore_impl(self.config.kvstore)
@ -144,6 +144,8 @@ class FaissVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
self,
vector_db: VectorDB,
) -> None:
assert self.kvstore is not None
key = f"{VECTOR_DBS_PREFIX}{vector_db.identifier}"
await self.kvstore.set(
key=key,
@ -161,6 +163,8 @@ class FaissVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
return [i.vector_db for i in self.cache.values()]
async def unregister_vector_db(self, vector_db_id: str) -> None:
assert self.kvstore is not None
if vector_db_id not in self.cache:
logger.warning(f"Vector DB {vector_db_id} not found")
return

19
llama_stack/providers/inline/vector_io/qdrant/__init__.py Normal file
View file

@ -0,0 +1,19 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Dict
from llama_stack.providers.datatypes import Api, ProviderSpec
from .config import QdrantVectorIOConfig
async def get_adapter_impl(config: QdrantVectorIOConfig, deps: Dict[Api, ProviderSpec]):
from llama_stack.providers.remote.vector_io.qdrant.qdrant import QdrantVectorIOAdapter
impl = QdrantVectorIOAdapter(config, deps[Api.inference])
await impl.initialize()
return impl

23
llama_stack/providers/inline/vector_io/qdrant/config.py Normal file
View file

@ -0,0 +1,23 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel
from llama_stack.schema_utils import json_schema_type
@json_schema_type
class QdrantVectorIOConfig(BaseModel):
path: str
@classmethod
def sample_run_config(cls, __distro_dir__: str) -> Dict[str, Any]:
return {
"path": "${env.QDRANT_PATH:~/.llama/" + __distro_dir__ + "}/" + "qdrant.db",
}

284
llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py
View file

@ -4,6 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import asyncio
import hashlib
import logging
import sqlite3
@ -15,9 +16,10 @@ import numpy as np
import sqlite_vec
from numpy.typing import NDArray
from llama_stack.apis.inference.inference import Inference
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.datatypes import VectorDBsProtocolPrivate
from llama_stack.providers.utils.memory.vector_store import EmbeddingIndex, VectorDBWithIndex
logger = logging.getLogger(__name__)
@ -28,6 +30,15 @@ def serialize_vector(vector: List[float]) -> bytes:
return struct.pack(f"{len(vector)}f", *vector)
def _create_sqlite_connection(db_path):
"""Create a SQLite connection with sqlite_vec extension loaded."""
connection = sqlite3.connect(db_path)
connection.enable_load_extension(True)
sqlite_vec.load(connection)
connection.enable_load_extension(False)
return connection
class SQLiteVecIndex(EmbeddingIndex):
"""
An index implementation that stores embeddings in a SQLite virtual table using sqlite-vec.
@ -36,40 +47,56 @@ class SQLiteVecIndex(EmbeddingIndex):
- A virtual table (vec_chunks_{bank_id}) that holds the serialized vector.
"""
def __init__(self, dimension: int, connection: sqlite3.Connection, bank_id: str):
def __init__(self, dimension: int, db_path: str, bank_id: str):
self.dimension = dimension
self.connection = connection
self.db_path = db_path
self.bank_id = bank_id
self.metadata_table = f"chunks_{bank_id}".replace("-", "_")
self.vector_table = f"vec_chunks_{bank_id}".replace("-", "_")
@classmethod
async def create(cls, dimension: int, connection: sqlite3.Connection, bank_id: str):
instance = cls(dimension, connection, bank_id)
async def create(cls, dimension: int, db_path: str, bank_id: str):
instance = cls(dimension, db_path, bank_id)
await instance.initialize()
return instance
async def initialize(self) -> None:
cur = self.connection.cursor()
# Create the table to store chunk metadata.
cur.execute(f"""
CREATE TABLE IF NOT EXISTS {self.metadata_table} (
id TEXT PRIMARY KEY,
chunk TEXT
);
""")
# Create the virtual table for embeddings.
cur.execute(f"""
CREATE VIRTUAL TABLE IF NOT EXISTS {self.vector_table}
USING vec0(embedding FLOAT[{self.dimension}], id TEXT);
""")
self.connection.commit()
def _init_tables():
connection = _create_sqlite_connection(self.db_path)
cur = connection.cursor()
try:
# Create the table to store chunk metadata.
cur.execute(f"""
CREATE TABLE IF NOT EXISTS {self.metadata_table} (
id TEXT PRIMARY KEY,
chunk TEXT
);
""")
# Create the virtual table for embeddings.
cur.execute(f"""
CREATE VIRTUAL TABLE IF NOT EXISTS {self.vector_table}
USING vec0(embedding FLOAT[{self.dimension}], id TEXT);
""")
connection.commit()
finally:
cur.close()
connection.close()
async def delete(self):
cur = self.connection.cursor()
cur.execute(f"DROP TABLE IF EXISTS {self.metadata_table};")
cur.execute(f"DROP TABLE IF EXISTS {self.vector_table};")
self.connection.commit()
await asyncio.to_thread(_init_tables)
async def delete(self) -> None:
def _drop_tables():
connection = _create_sqlite_connection(self.db_path)
cur = connection.cursor()
try:
cur.execute(f"DROP TABLE IF EXISTS {self.metadata_table};")
cur.execute(f"DROP TABLE IF EXISTS {self.vector_table};")
connection.commit()
finally:
cur.close()
connection.close()
await asyncio.to_thread(_drop_tables)
async def add_chunks(self, chunks: List[Chunk], embeddings: NDArray, batch_size: int = 500):
"""
@ -78,42 +105,57 @@ class SQLiteVecIndex(EmbeddingIndex):
embedding (serialized to raw bytes) into the virtual table using the assigned rowid.
If any insert fails, the transaction is rolled back to maintain consistency.
"""
cur = self.connection.cursor()
try:
# Start transaction
cur.execute("BEGIN TRANSACTION")
for i in range(0, len(chunks), batch_size):
batch_chunks = chunks[i : i + batch_size]
batch_embeddings = embeddings[i : i + batch_size]
# Prepare metadata inserts
metadata_data = [
(generate_chunk_id(chunk.metadata["document_id"], chunk.content), chunk.model_dump_json())
for chunk in batch_chunks
]
# Insert metadata (ON CONFLICT to avoid duplicates)
cur.executemany(
f"""
INSERT INTO {self.metadata_table} (id, chunk)
VALUES (?, ?)
ON CONFLICT(id) DO UPDATE SET chunk = excluded.chunk;
""",
metadata_data,
)
# Prepare embeddings inserts
embedding_data = [
(generate_chunk_id(chunk.metadata["document_id"], chunk.content), serialize_vector(emb.tolist()))
for chunk, emb in zip(batch_chunks, batch_embeddings, strict=True)
]
# Insert embeddings in batch
cur.executemany(f"INSERT INTO {self.vector_table} (id, embedding) VALUES (?, ?);", embedding_data)
self.connection.commit()
assert all(isinstance(chunk.content, str) for chunk in chunks), "SQLiteVecIndex only supports text chunks"
except sqlite3.Error as e:
self.connection.rollback() # Rollback on failure
logger.error(f"Error inserting into {self.vector_table}: {e}")
def _execute_all_batch_inserts():
connection = _create_sqlite_connection(self.db_path)
cur = connection.cursor()
finally:
cur.close() # Ensure cursor is closed
try:
# Start transaction a single transcation for all batches
cur.execute("BEGIN TRANSACTION")
for i in range(0, len(chunks), batch_size):
batch_chunks = chunks[i : i + batch_size]
batch_embeddings = embeddings[i : i + batch_size]
# Prepare metadata inserts
metadata_data = [
(generate_chunk_id(chunk.metadata["document_id"], chunk.content), chunk.model_dump_json())
for chunk in batch_chunks
if isinstance(chunk.content, str)
]
# Insert metadata (ON CONFLICT to avoid duplicates)
cur.executemany(
f"""
INSERT INTO {self.metadata_table} (id, chunk)
VALUES (?, ?)
ON CONFLICT(id) DO UPDATE SET chunk = excluded.chunk;
""",
metadata_data,
)
# Prepare embeddings inserts
embedding_data = [
(
generate_chunk_id(chunk.metadata["document_id"], chunk.content),
serialize_vector(emb.tolist()),
)
for chunk, emb in zip(batch_chunks, batch_embeddings, strict=True)
if isinstance(chunk.content, str)
]
# Insert embeddings in batch
cur.executemany(f"INSERT INTO {self.vector_table} (id, embedding) VALUES (?, ?);", embedding_data)
connection.commit()
except sqlite3.Error as e:
connection.rollback() # Rollback on failure
logger.error(f"Error inserting into {self.vector_table}: {e}")
raise
finally:
cur.close()
connection.close()
# Process all batches in a single thread
await asyncio.to_thread(_execute_all_batch_inserts)
async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
"""
@ -122,18 +164,28 @@ class SQLiteVecIndex(EmbeddingIndex):
"""
emb_list = embedding.tolist() if isinstance(embedding, np.ndarray) else list(embedding)
emb_blob = serialize_vector(emb_list)
cur = self.connection.cursor()
query_sql = f"""
SELECT m.id, m.chunk, v.distance
FROM {self.vector_table} AS v
JOIN {self.metadata_table} AS m ON m.id = v.id
WHERE v.embedding MATCH ? AND k = ?
ORDER BY v.distance;
"""
cur.execute(query_sql, (emb_blob, k))
rows = cur.fetchall()
chunks = []
scores = []
def _execute_query():
connection = _create_sqlite_connection(self.db_path)
cur = connection.cursor()
try:
query_sql = f"""
SELECT m.id, m.chunk, v.distance
FROM {self.vector_table} AS v
JOIN {self.metadata_table} AS m ON m.id = v.id
WHERE v.embedding MATCH ? AND k = ?
ORDER BY v.distance;
"""
cur.execute(query_sql, (emb_blob, k))
return cur.fetchall()
finally:
cur.close()
connection.close()
rows = await asyncio.to_thread(_execute_query)
chunks, scores = [], []
for _id, chunk_json, distance in rows:
try:
chunk = Chunk.model_validate_json(chunk_json)
@ -154,67 +206,85 @@ class SQLiteVecVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
and creates a cache of VectorDBWithIndex instances (each wrapping a SQLiteVecIndex).
"""
def __init__(self, config, inference_api: Api.inference) -> None:
def __init__(self, config, inference_api: Inference) -> None:
self.config = config
self.inference_api = inference_api
self.cache: Dict[str, VectorDBWithIndex] = {}
self.connection: Optional[sqlite3.Connection] = None
async def initialize(self) -> None:
# Open a connection to the SQLite database (the file is specified in the config).
self.connection = sqlite3.connect(self.config.db_path)
self.connection.enable_load_extension(True)
sqlite_vec.load(self.connection)
self.connection.enable_load_extension(False)
cur = self.connection.cursor()
# Create a table to persist vector DB registrations.
cur.execute("""
CREATE TABLE IF NOT EXISTS vector_dbs (
id TEXT PRIMARY KEY,
metadata TEXT
);
""")
self.connection.commit()
# Load any existing vector DB registrations.
cur.execute("SELECT metadata FROM vector_dbs")
rows = cur.fetchall()
def _setup_connection():
# Open a connection to the SQLite database (the file is specified in the config).
connection = _create_sqlite_connection(self.config.db_path)
cur = connection.cursor()
try:
# Create a table to persist vector DB registrations.
cur.execute("""
CREATE TABLE IF NOT EXISTS vector_dbs (
id TEXT PRIMARY KEY,
metadata TEXT
);
""")
connection.commit()
# Load any existing vector DB registrations.
cur.execute("SELECT metadata FROM vector_dbs")
rows = cur.fetchall()
return rows
finally:
cur.close()
connection.close()
rows = await asyncio.to_thread(_setup_connection)
for row in rows:
vector_db_data = row[0]
vector_db = VectorDB.model_validate_json(vector_db_data)
index = await SQLiteVecIndex.create(vector_db.embedding_dimension, self.connection, vector_db.identifier)
index = await SQLiteVecIndex.create(
vector_db.embedding_dimension, self.config.db_path, vector_db.identifier
)
self.cache[vector_db.identifier] = VectorDBWithIndex(vector_db, index, self.inference_api)
async def shutdown(self) -> None:
if self.connection:
self.connection.close()
self.connection = None
# nothing to do since we don't maintain a persistent connection
pass
async def register_vector_db(self, vector_db: VectorDB) -> None:
if self.connection is None:
raise RuntimeError("SQLite connection not initialized")
cur = self.connection.cursor()
cur.execute(
"INSERT OR REPLACE INTO vector_dbs (id, metadata) VALUES (?, ?)",
(vector_db.identifier, vector_db.model_dump_json()),
)
self.connection.commit()
index = await SQLiteVecIndex.create(vector_db.embedding_dimension, self.connection, vector_db.identifier)
def _register_db():
connection = _create_sqlite_connection(self.config.db_path)
cur = connection.cursor()
try:
cur.execute(
"INSERT OR REPLACE INTO vector_dbs (id, metadata) VALUES (?, ?)",
(vector_db.identifier, vector_db.model_dump_json()),
)
connection.commit()
finally:
cur.close()
connection.close()
await asyncio.to_thread(_register_db)
index = await SQLiteVecIndex.create(vector_db.embedding_dimension, self.config.db_path, vector_db.identifier)
self.cache[vector_db.identifier] = VectorDBWithIndex(vector_db, index, self.inference_api)
async def list_vector_dbs(self) -> List[VectorDB]:
return [v.vector_db for v in self.cache.values()]
async def unregister_vector_db(self, vector_db_id: str) -> None:
if self.connection is None:
raise RuntimeError("SQLite connection not initialized")
if vector_db_id not in self.cache:
logger.warning(f"Vector DB {vector_db_id} not found")
return
await self.cache[vector_db_id].index.delete()
del self.cache[vector_db_id]
cur = self.connection.cursor()
cur.execute("DELETE FROM vector_dbs WHERE id = ?", (vector_db_id,))
self.connection.commit()
def _delete_vector_db_from_registry():
connection = _create_sqlite_connection(self.config.db_path)
cur = connection.cursor()
try:
cur.execute("DELETE FROM vector_dbs WHERE id = ?", (vector_db_id,))
connection.commit()
finally:
cur.close()
connection.close()
await asyncio.to_thread(_delete_vector_db_from_registry)
async def insert_chunks(self, vector_db_id: str, chunks: List[Chunk], ttl_seconds: Optional[int] = None) -> None:
if vector_db_id not in self.cache: