phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-07-01 03:58:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into watsonx_hc

Browse source
This commit is contained in:
Sumit Jaiswal 2025-06-16 14:44:46 +05:30 • committed by GitHub
commit f5388e252d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
48 changed files with 2179 additions and 66 deletions
Download patch file Download diff file Expand all files Collapse all files

16
llama_stack/apis/agents/openai_responses.py
View file

@ -81,6 +81,15 @@ class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
type: Literal["web_search_call"] = "web_search_call"
@json_schema_type
class OpenAIResponseOutputMessageFileSearchToolCall(BaseModel):
id: str
queries: list[str]
status: str
type: Literal["file_search_call"] = "file_search_call"
results: list[dict[str, Any]] | None = None
@json_schema_type
class OpenAIResponseOutputMessageFunctionToolCall(BaseModel):
call_id: str
@ -119,6 +128,7 @@ class OpenAIResponseOutputMessageMCPListTools(BaseModel):
OpenAIResponseOutput = Annotated[
OpenAIResponseMessage
| OpenAIResponseOutputMessageWebSearchToolCall
| OpenAIResponseOutputMessageFileSearchToolCall
| OpenAIResponseOutputMessageFunctionToolCall
| OpenAIResponseOutputMessageMCPCall
| OpenAIResponseOutputMessageMCPListTools,
@ -362,6 +372,7 @@ class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
OpenAIResponseInput = Annotated[
# Responses API allows output messages to be passed in as input
OpenAIResponseOutputMessageWebSearchToolCall
| OpenAIResponseOutputMessageFileSearchToolCall
| OpenAIResponseOutputMessageFunctionToolCall
| OpenAIResponseInputFunctionToolCallOutput
|
@ -397,9 +408,10 @@ class FileSearchRankingOptions(BaseModel):
@json_schema_type
class OpenAIResponseInputToolFileSearch(BaseModel):
type: Literal["file_search"] = "file_search"
vector_store_id: list[str]
vector_store_ids: list[str]
filters: dict[str, Any] | None = None
max_num_results: int | None = Field(default=10, ge=1, le=50)
ranking_options: FileSearchRankingOptions | None = None
# TODO: add filters
class ApprovalFilter(BaseModel):

3
llama_stack/apis/inference/inference.py
View file

@ -1038,6 +1038,8 @@ class InferenceProvider(Protocol):
# vLLM-specific parameters
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
# for fill-in-the-middle type completion
suffix: str | None = None,
) -> OpenAICompletion:
"""Generate an OpenAI-compatible completion for the given prompt using the specified model.
@ -1058,6 +1060,7 @@ class InferenceProvider(Protocol):
:param temperature: (Optional) The temperature to use.
:param top_p: (Optional) The top p to use.
:param user: (Optional) The user to use.
:param suffix: (Optional) The suffix that should be appended to the completion.
:returns: An OpenAICompletion.
"""
...

46
llama_stack/apis/tools/rag_tool.py
View file

@ -15,6 +15,48 @@ from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@json_schema_type
class RRFRanker(BaseModel):
"""
Reciprocal Rank Fusion (RRF) ranker configuration.
:param type: The type of ranker, always "rrf"
:param impact_factor: The impact factor for RRF scoring. Higher values give more weight to higher-ranked results.
Must be greater than 0. Default of 60 is from the original RRF paper (Cormack et al., 2009).
"""
type: Literal["rrf"] = "rrf"
impact_factor: float = Field(default=60.0, gt=0.0) # default of 60 for optimal performance
@json_schema_type
class WeightedRanker(BaseModel):
"""
Weighted ranker configuration that combines vector and keyword scores.
:param type: The type of ranker, always "weighted"
:param alpha: Weight factor between 0 and 1.
0 means only use keyword scores,
1 means only use vector scores,
values in between blend both scores.
"""
type: Literal["weighted"] = "weighted"
alpha: float = Field(
default=0.5,
ge=0.0,
le=1.0,
description="Weight factor between 0 and 1. 0 means only keyword scores, 1 means only vector scores.",
)
Ranker = Annotated[
RRFRanker | WeightedRanker,
Field(discriminator="type"),
]
register_schema(Ranker, name="Ranker")
@json_schema_type
class RAGDocument(BaseModel):
"""
@ -76,7 +118,8 @@ class RAGQueryConfig(BaseModel):
:param chunk_template: Template for formatting each retrieved chunk in the context.
Available placeholders: {index} (1-based chunk ordinal), {chunk.content} (chunk content string), {metadata} (chunk metadata dict).
Default: "Result {index}\\nContent: {chunk.content}\\nMetadata: {metadata}\\n"
:param mode: Search mode for retrieval—either "vector" or "keyword". Default "vector".
:param mode: Search mode for retrieval—either "vector", "keyword", or "hybrid". Default "vector".
:param ranker: Configuration for the ranker to use in hybrid search. Defaults to RRF ranker.
"""
# This config defines how a query is generated using the messages
@ -86,6 +129,7 @@ class RAGQueryConfig(BaseModel):
max_chunks: int = 5
chunk_template: str = "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n"
mode: str | None = None
ranker: Ranker | None = Field(default=None) # Only used for hybrid mode
@field_validator("chunk_template")
def validate_chunk_template(cls, v: str) -> str:

65
llama_stack/apis/vector_io/vector_io.py
View file

@ -8,7 +8,7 @@
#
# 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, Literal, Protocol, runtime_checkable
from typing import Annotated, Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, Field
@ -16,6 +16,7 @@ from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
from llama_stack.strong_typing.schema import register_schema
class Chunk(BaseModel):
@ -133,6 +134,50 @@ class VectorStoreDeleteResponse(BaseModel):
deleted: bool = True
@json_schema_type
class VectorStoreChunkingStrategyAuto(BaseModel):
type: Literal["auto"] = "auto"
@json_schema_type
class VectorStoreChunkingStrategyStaticConfig(BaseModel):
chunk_overlap_tokens: int = 400
max_chunk_size_tokens: int = Field(800, ge=100, le=4096)
@json_schema_type
class VectorStoreChunkingStrategyStatic(BaseModel):
type: Literal["static"] = "static"
static: VectorStoreChunkingStrategyStaticConfig
VectorStoreChunkingStrategy = Annotated[
VectorStoreChunkingStrategyAuto | VectorStoreChunkingStrategyStatic, Field(discriminator="type")
]
register_schema(VectorStoreChunkingStrategy, name="VectorStoreChunkingStrategy")
@json_schema_type
class VectorStoreFileLastError(BaseModel):
code: Literal["server_error"] | Literal["rate_limit_exceeded"]
message: str
@json_schema_type
class VectorStoreFileObject(BaseModel):
"""OpenAI Vector Store File object."""
id: str
object: str = "vector_store.file"
attributes: dict[str, Any] = Field(default_factory=dict)
chunking_strategy: VectorStoreChunkingStrategy
created_at: int
last_error: VectorStoreFileLastError | None = None
status: Literal["completed"] | Literal["in_progress"] | Literal["cancelled"] | Literal["failed"]
usage_bytes: int = 0
vector_store_id: str
class VectorDBStore(Protocol):
def get_vector_db(self, vector_db_id: str) -> VectorDB | None: ...
@ -290,3 +335,21 @@ class VectorIO(Protocol):
:returns: A VectorStoreSearchResponse containing the search results.
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files", method="POST")
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
"""Attach a file to a vector store.
:param vector_store_id: The ID of the vector store to attach the file to.
:param file_id: The ID of the file to attach to the vector store.
:param attributes: The key-value attributes stored with the file, which can be used for filtering.
:param chunking_strategy: The chunking strategy to use for the file.
:returns: A VectorStoreFileObject representing the attached file.
"""
...

2
llama_stack/distribution/routers/inference.py
View file

@ -426,6 +426,7 @@ class InferenceRouter(Inference):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
logger.debug(
f"InferenceRouter.openai_completion: {model=}, {stream=}, {prompt=}",
@ -456,6 +457,7 @@ class InferenceRouter(Inference):
user=user,
guided_choice=guided_choice,
prompt_logprobs=prompt_logprobs,
suffix=suffix,
)
provider = self.routing_table.get_provider_impl(model_obj.identifier)

18
llama_stack/distribution/routers/vector_io.py
View file

@ -19,6 +19,7 @@ from llama_stack.apis.vector_io import (
VectorStoreObject,
VectorStoreSearchResponsePage,
)
from llama_stack.apis.vector_io.vector_io import VectorStoreChunkingStrategy, VectorStoreFileObject
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import RoutingTable
@ -254,3 +255,20 @@ class VectorIORouter(VectorIO):
ranking_options=ranking_options,
rewrite_query=rewrite_query,
)
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
logger.debug(f"VectorIORouter.openai_attach_file_to_vector_store: {vector_store_id}, {file_id}")
# Route based on vector store ID
provider = self.routing_table.get_provider_impl(vector_store_id)
return await provider.openai_attach_file_to_vector_store(
vector_store_id=vector_store_id,
file_id=file_id,
attributes=attributes,
chunking_strategy=chunking_strategy,
)

58
llama_stack/providers/inline/agents/meta_reference/openai_responses.py
View file

@ -24,6 +24,7 @@ from llama_stack.apis.agents.openai_responses import (
OpenAIResponseInputMessageContentImage,
OpenAIResponseInputMessageContentText,
OpenAIResponseInputTool,
OpenAIResponseInputToolFileSearch,
OpenAIResponseInputToolMCP,
OpenAIResponseMessage,
OpenAIResponseObject,
@ -34,6 +35,7 @@ from llama_stack.apis.agents.openai_responses import (
OpenAIResponseOutput,
OpenAIResponseOutputMessageContent,
OpenAIResponseOutputMessageContentOutputText,
OpenAIResponseOutputMessageFileSearchToolCall,
OpenAIResponseOutputMessageFunctionToolCall,
OpenAIResponseOutputMessageMCPListTools,
OpenAIResponseOutputMessageWebSearchToolCall,
@ -62,7 +64,7 @@ from llama_stack.apis.inference.inference import (
OpenAIToolMessageParam,
OpenAIUserMessageParam,
)
from llama_stack.apis.tools.tools import ToolGroups, ToolRuntime
from llama_stack.apis.tools import RAGQueryConfig, ToolGroups, ToolRuntime
from llama_stack.log import get_logger
from llama_stack.models.llama.datatypes import ToolDefinition, ToolParamDefinition
from llama_stack.providers.utils.inference.openai_compat import convert_tooldef_to_openai_tool
@ -198,7 +200,8 @@ class OpenAIResponsePreviousResponseWithInputItems(BaseModel):
class ChatCompletionContext(BaseModel):
model: str
messages: list[OpenAIMessageParam]
tools: list[ChatCompletionToolParam] | None = None
response_tools: list[OpenAIResponseInputTool] | None = None
chat_tools: list[ChatCompletionToolParam] | None = None
mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP]
temperature: float | None
response_format: OpenAIResponseFormatParam
@ -388,7 +391,8 @@ class OpenAIResponsesImpl:
ctx = ChatCompletionContext(
model=model,
messages=messages,
tools=chat_tools,
response_tools=tools,
chat_tools=chat_tools,
mcp_tool_to_server=mcp_tool_to_server,
temperature=temperature,
response_format=response_format,
@ -417,7 +421,7 @@ class OpenAIResponsesImpl:
completion_result = await self.inference_api.openai_chat_completion(
model=ctx.model,
messages=messages,
tools=ctx.tools,
tools=ctx.chat_tools,
stream=True,
temperature=ctx.temperature,
response_format=ctx.response_format,
@ -606,6 +610,12 @@ class OpenAIResponsesImpl:
if not tool:
raise ValueError(f"Tool {tool_name} not found")
chat_tools.append(make_openai_tool(tool_name, tool))
elif input_tool.type == "file_search":
tool_name = "knowledge_search"
tool = await self.tool_groups_api.get_tool(tool_name)
if not tool:
raise ValueError(f"Tool {tool_name} not found")
chat_tools.append(make_openai_tool(tool_name, tool))
elif input_tool.type == "mcp":
always_allowed = None
never_allowed = None
@ -667,6 +677,7 @@ class OpenAIResponsesImpl:
tool_call_id = tool_call.id
function = tool_call.function
tool_kwargs = json.loads(function.arguments) if function.arguments else {}
if not function or not tool_call_id or not function.name:
return None, None
@ -680,12 +691,26 @@ class OpenAIResponsesImpl:
endpoint=mcp_tool.server_url,
headers=mcp_tool.headers or {},
tool_name=function.name,
kwargs=json.loads(function.arguments) if function.arguments else {},
kwargs=tool_kwargs,
)
else:
if function.name == "knowledge_search":
response_file_search_tool = next(
t for t in ctx.response_tools if isinstance(t, OpenAIResponseInputToolFileSearch)
)
if response_file_search_tool:
if response_file_search_tool.filters:
logger.warning("Filters are not yet supported for file_search tool")
if response_file_search_tool.ranking_options:
logger.warning("Ranking options are not yet supported for file_search tool")
tool_kwargs["vector_db_ids"] = response_file_search_tool.vector_store_ids
tool_kwargs["query_config"] = RAGQueryConfig(
mode="vector",
max_chunks=response_file_search_tool.max_num_results,
)
result = await self.tool_runtime_api.invoke_tool(
tool_name=function.name,
kwargs=json.loads(function.arguments) if function.arguments else {},
kwargs=tool_kwargs,
)
except Exception as e:
error_exc = e
@ -713,6 +738,27 @@ class OpenAIResponsesImpl:
)
if error_exc or (result.error_code and result.error_code > 0) or result.error_message:
message.status = "failed"
elif function.name == "knowledge_search":
message = OpenAIResponseOutputMessageFileSearchToolCall(
id=tool_call_id,
queries=[tool_kwargs.get("query", "")],
status="completed",
)
if "document_ids" in result.metadata:
message.results = []
for i, doc_id in enumerate(result.metadata["document_ids"]):
text = result.metadata["chunks"][i] if "chunks" in result.metadata else None
score = result.metadata["scores"][i] if "scores" in result.metadata else None
message.results.append(
{
"file_id": doc_id,
"filename": doc_id,
"text": text,
"score": score,
}
)
if error_exc or (result.error_code and result.error_code > 0) or result.error_message:
message.status = "failed"
else:
raise ValueError(f"Unknown tool {function.name} called")

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

@ -121,8 +121,10 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
vector_db_id=vector_db_id,
query=query,
params={
"max_chunks": query_config.max_chunks,
"mode": query_config.mode,
"max_chunks": query_config.max_chunks,
"score_threshold": 0.0,
"ranker": query_config.ranker,
},
)
for vector_db_id in vector_db_ids
@ -170,6 +172,8 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
content=picked,
metadata={
"document_ids": [c.metadata["document_id"] for c in chunks[: len(picked)]],
"chunks": [c.content for c in chunks[: len(picked)]],
"scores": scores[: len(picked)],
},
)

2
llama_stack/providers/inline/vector_io/faiss/__init__.py
View file

@ -16,6 +16,6 @@ async def get_provider_impl(config: FaissVectorIOConfig, deps: dict[Api, Any]):
assert isinstance(config, FaissVectorIOConfig), f"Unexpected config type: {type(config)}"
impl = FaissVectorIOAdapter(config, deps[Api.inference])
impl = FaissVectorIOAdapter(config, deps[Api.inference], deps.get(Api.files, None))
await impl.initialize()
return impl

15
llama_stack/providers/inline/vector_io/faiss/faiss.py
View file

@ -15,6 +15,7 @@ import faiss
import numpy as np
from numpy.typing import NDArray
from llama_stack.apis.files import Files
from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.inference.inference import Inference
from llama_stack.apis.vector_dbs import VectorDB
@ -130,11 +131,23 @@ class FaissIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in FAISS")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in FAISS")
class FaissVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):
def __init__(self, config: FaissVectorIOConfig, inference_api: Inference) -> None:
def __init__(self, config: FaissVectorIOConfig, inference_api: Inference, files_api: Files | None) -> None:
self.config = config
self.inference_api = inference_api
self.files_api = files_api
self.cache: dict[str, VectorDBWithIndex] = {}
self.kvstore: KVStore | None = None
self.openai_vector_stores: dict[str, dict[str, Any]] = {}

2
llama_stack/providers/inline/vector_io/sqlite_vec/__init__.py
View file

@ -15,6 +15,6 @@ async def get_provider_impl(config: SQLiteVectorIOConfig, deps: dict[Api, Any]):
from .sqlite_vec import SQLiteVecVectorIOAdapter
assert isinstance(config, SQLiteVectorIOConfig), f"Unexpected config type: {type(config)}"
impl = SQLiteVecVectorIOAdapter(config, deps[Api.inference])
impl = SQLiteVecVectorIOAdapter(config, deps[Api.inference], deps.get(Api.files, None))
await impl.initialize()
return impl

154
llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py
View file

@ -17,6 +17,7 @@ import numpy as np
import sqlite_vec
from numpy.typing import NDArray
from llama_stack.apis.files.files import Files
from llama_stack.apis.inference.inference import Inference
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import (
@ -26,14 +27,20 @@ from llama_stack.apis.vector_io import (
)
from llama_stack.providers.datatypes import VectorDBsProtocolPrivate
from llama_stack.providers.utils.memory.openai_vector_store_mixin import OpenAIVectorStoreMixin
from llama_stack.providers.utils.memory.vector_store import EmbeddingIndex, VectorDBWithIndex
from llama_stack.providers.utils.memory.vector_store import (
RERANKER_TYPE_RRF,
RERANKER_TYPE_WEIGHTED,
EmbeddingIndex,
VectorDBWithIndex,
)
logger = logging.getLogger(__name__)
# Specifying search mode is dependent on the VectorIO provider.
VECTOR_SEARCH = "vector"
KEYWORD_SEARCH = "keyword"
SEARCH_MODES = {VECTOR_SEARCH, KEYWORD_SEARCH}
HYBRID_SEARCH = "hybrid"
SEARCH_MODES = {VECTOR_SEARCH, KEYWORD_SEARCH, HYBRID_SEARCH}
def serialize_vector(vector: list[float]) -> bytes:
@ -50,6 +57,59 @@ def _create_sqlite_connection(db_path):
return connection
def _normalize_scores(scores: dict[str, float]) -> dict[str, float]:
"""Normalize scores to [0,1] range using min-max normalization."""
if not scores:
return {}
min_score = min(scores.values())
max_score = max(scores.values())
score_range = max_score - min_score
if score_range > 0:
return {doc_id: (score - min_score) / score_range for doc_id, score in scores.items()}
return {doc_id: 1.0 for doc_id in scores}
def _weighted_rerank(
vector_scores: dict[str, float],
keyword_scores: dict[str, float],
alpha: float = 0.5,
) -> dict[str, float]:
"""ReRanker that uses weighted average of scores."""
all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
normalized_vector_scores = _normalize_scores(vector_scores)
normalized_keyword_scores = _normalize_scores(keyword_scores)
return {
doc_id: (alpha * normalized_keyword_scores.get(doc_id, 0.0))
+ ((1 - alpha) * normalized_vector_scores.get(doc_id, 0.0))
for doc_id in all_ids
}
def _rrf_rerank(
vector_scores: dict[str, float],
keyword_scores: dict[str, float],
impact_factor: float = 60.0,
) -> dict[str, float]:
"""ReRanker that uses Reciprocal Rank Fusion."""
# Convert scores to ranks
vector_ranks = {
doc_id: i + 1 for i, (doc_id, _) in enumerate(sorted(vector_scores.items(), key=lambda x: x[1], reverse=True))
}
keyword_ranks = {
doc_id: i + 1 for i, (doc_id, _) in enumerate(sorted(keyword_scores.items(), key=lambda x: x[1], reverse=True))
}
all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
rrf_scores = {}
for doc_id in all_ids:
vector_rank = vector_ranks.get(doc_id, float("inf"))
keyword_rank = keyword_ranks.get(doc_id, float("inf"))
# RRF formula: score = 1/(k + r) where k is impact_factor and r is the rank
rrf_scores[doc_id] = (1.0 / (impact_factor + vector_rank)) + (1.0 / (impact_factor + keyword_rank))
return rrf_scores
class SQLiteVecIndex(EmbeddingIndex):
"""
An index implementation that stores embeddings in a SQLite virtual table using sqlite-vec.
@ -254,8 +314,6 @@ class SQLiteVecIndex(EmbeddingIndex):
"""
Performs keyword-based search using SQLite FTS5 for relevance-ranked full-text search.
"""
if query_string is None:
raise ValueError("query_string is required for keyword search.")
def _execute_query():
connection = _create_sqlite_connection(self.db_path)
@ -293,6 +351,81 @@ class SQLiteVecIndex(EmbeddingIndex):
scores.append(score)
return QueryChunksResponse(chunks=chunks, scores=scores)
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str = RERANKER_TYPE_RRF,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
"""
Hybrid search using a configurable re-ranking strategy.
Args:
embedding: The query embedding vector
query_string: The text query for keyword search
k: Number of results to return
score_threshold: Minimum similarity score threshold
reranker_type: Type of reranker to use ("rrf" or "weighted")
reranker_params: Parameters for the reranker
Returns:
QueryChunksResponse with combined results
"""
if reranker_params is None:
reranker_params = {}
# Get results from both search methods
vector_response = await self.query_vector(embedding, k, score_threshold)
keyword_response = await self.query_keyword(query_string, k, score_threshold)
# Convert responses to score dictionaries using generate_chunk_id
vector_scores = {
generate_chunk_id(chunk.metadata["document_id"], str(chunk.content)): score
for chunk, score in zip(vector_response.chunks, vector_response.scores, strict=False)
}
keyword_scores = {
generate_chunk_id(chunk.metadata["document_id"], str(chunk.content)): score
for chunk, score in zip(keyword_response.chunks, keyword_response.scores, strict=False)
}
# Combine scores using the specified reranker
if reranker_type == RERANKER_TYPE_WEIGHTED:
alpha = reranker_params.get("alpha", 0.5)
combined_scores = _weighted_rerank(vector_scores, keyword_scores, alpha)
else:
# Default to RRF for None, RRF, or any unknown types
impact_factor = reranker_params.get("impact_factor", 60.0)
combined_scores = _rrf_rerank(vector_scores, keyword_scores, impact_factor)
# Sort by combined score and get top k results
sorted_items = sorted(combined_scores.items(), key=lambda x: x[1], reverse=True)
top_k_items = sorted_items[:k]
# Filter by score threshold
filtered_items = [(doc_id, score) for doc_id, score in top_k_items if score >= score_threshold]
# Create a map of chunk_id to chunk for both responses
chunk_map = {}
for c in vector_response.chunks:
chunk_id = generate_chunk_id(c.metadata["document_id"], str(c.content))
chunk_map[chunk_id] = c
for c in keyword_response.chunks:
chunk_id = generate_chunk_id(c.metadata["document_id"], str(c.content))
chunk_map[chunk_id] = c
# Use the map to look up chunks by their IDs
chunks = []
scores = []
for doc_id, score in filtered_items:
if doc_id in chunk_map:
chunks.append(chunk_map[doc_id])
scores.append(score)
return QueryChunksResponse(chunks=chunks, scores=scores)
class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):
"""
@ -301,9 +434,10 @@ class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoc
and creates a cache of VectorDBWithIndex instances (each wrapping a SQLiteVecIndex).
"""
def __init__(self, config, inference_api: Inference) -> None:
def __init__(self, config, inference_api: Inference, files_api: Files | None) -> None:
self.config = config
self.inference_api = inference_api
self.files_api = files_api
self.cache: dict[str, VectorDBWithIndex] = {}
self.openai_vector_stores: dict[str, dict[str, Any]] = {}
@ -343,7 +477,9 @@ class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoc
vector_db_data = row[0]
vector_db = VectorDB.model_validate_json(vector_db_data)
index = await SQLiteVecIndex.create(
vector_db.embedding_dimension, self.config.db_path, vector_db.identifier
vector_db.embedding_dimension,
self.config.db_path,
vector_db.identifier,
)
self.cache[vector_db.identifier] = VectorDBWithIndex(vector_db, index, self.inference_api)
@ -369,7 +505,11 @@ class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoc
connection.close()
await asyncio.to_thread(_register_db)
index = await SQLiteVecIndex.create(vector_db.embedding_dimension, self.config.db_path, 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 list_vector_dbs(self) -> list[VectorDB]:

4
llama_stack/providers/registry/vector_io.py
View file

@ -24,6 +24,7 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.vector_io.faiss.FaissVectorIOConfig",
deprecation_warning="Please use the `inline::faiss` provider instead.",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
InlineProviderSpec(
api=Api.vector_io,
@ -32,6 +33,7 @@ def available_providers() -> list[ProviderSpec]:
module="llama_stack.providers.inline.vector_io.faiss",
config_class="llama_stack.providers.inline.vector_io.faiss.FaissVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
# NOTE: sqlite-vec cannot be bundled into the container image because it does not have a
# source distribution and the wheels are not available for all platforms.
@ -42,6 +44,7 @@ def available_providers() -> list[ProviderSpec]:
module="llama_stack.providers.inline.vector_io.sqlite_vec",
config_class="llama_stack.providers.inline.vector_io.sqlite_vec.SQLiteVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
InlineProviderSpec(
api=Api.vector_io,
@ -51,6 +54,7 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.vector_io.sqlite_vec.SQLiteVectorIOConfig",
deprecation_warning="Please use the `inline::sqlite-vec` provider (notice the hyphen instead of underscore) instead.",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
remote_provider_spec(
Api.vector_io,

1
llama_stack/providers/remote/inference/fireworks/fireworks.py
View file

@ -318,6 +318,7 @@ class FireworksInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProv
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
model_obj = await self.model_store.get_model(model)

1
llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -316,6 +316,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
provider_model_id = await self._get_provider_model_id(model)

37
llama_stack/providers/remote/inference/ollama/ollama.py
View file

@ -33,7 +33,6 @@ from llama_stack.apis.inference import (
JsonSchemaResponseFormat,
LogProbConfig,
Message,
OpenAIEmbeddingsResponse,
ResponseFormat,
SamplingParams,
TextTruncation,
@ -46,6 +45,8 @@ from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionChunk,
OpenAICompletion,
OpenAIEmbeddingsResponse,
OpenAIEmbeddingUsage,
OpenAIMessageParam,
OpenAIResponseFormatParam,
)
@ -62,8 +63,10 @@ from llama_stack.providers.utils.inference.model_registry import (
from llama_stack.providers.utils.inference.openai_compat import (
OpenAICompatCompletionChoice,
OpenAICompatCompletionResponse,
b64_encode_openai_embeddings_response,
get_sampling_options,
prepare_openai_completion_params,
prepare_openai_embeddings_params,
process_chat_completion_response,
process_chat_completion_stream_response,
process_completion_response,
@ -386,7 +389,35 @@ class OllamaInferenceAdapter(
dimensions: int | None = None,
user: str | None = None,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()
model_obj = await self._get_model(model)
if model_obj.model_type != ModelType.embedding:
raise ValueError(f"Model {model} is not an embedding model")
if model_obj.provider_resource_id is None:
raise ValueError(f"Model {model} has no provider_resource_id set")
# Note, at the moment Ollama does not support encoding_format, dimensions, and user parameters
params = prepare_openai_embeddings_params(
model=model_obj.provider_resource_id,
input=input,
encoding_format=encoding_format,
dimensions=dimensions,
user=user,
)
response = await self.openai_client.embeddings.create(**params)
data = b64_encode_openai_embeddings_response(response.data, encoding_format)
usage = OpenAIEmbeddingUsage(
prompt_tokens=response.usage.prompt_tokens,
total_tokens=response.usage.total_tokens,
)
# TODO: Investigate why model_obj.identifier is used instead of response.model
return OpenAIEmbeddingsResponse(
data=data,
model=model_obj.identifier,
usage=usage,
)
async def openai_completion(
self,
@ -409,6 +440,7 @@ class OllamaInferenceAdapter(
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
if not isinstance(prompt, str):
raise ValueError("Ollama does not support non-string prompts for completion")
@ -432,6 +464,7 @@ class OllamaInferenceAdapter(
temperature=temperature,
top_p=top_p,
user=user,
suffix=suffix,
)
return await self.openai_client.completions.create(**params) # type: ignore

2
llama_stack/providers/remote/inference/openai/openai.py
View file

@ -90,6 +90,7 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
if guided_choice is not None:
logging.warning("guided_choice is not supported by the OpenAI API. Ignoring.")
@ -117,6 +118,7 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
temperature=temperature,
top_p=top_p,
user=user,
suffix=suffix,
)
return await self._openai_client.completions.create(**params)

1
llama_stack/providers/remote/inference/passthrough/passthrough.py
View file

@ -242,6 +242,7 @@ class PassthroughInferenceAdapter(Inference):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
client = self._get_client()
model_obj = await self.model_store.get_model(model)

1
llama_stack/providers/remote/inference/together/together.py
View file

@ -299,6 +299,7 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
model_obj = await self.model_store.get_model(model)
params = await prepare_openai_completion_params(

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

@ -559,6 +559,7 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
self._lazy_initialize_client()
model_obj = await self._get_model(model)

1
llama_stack/providers/remote/inference/watsonx/watsonx.py
View file

@ -313,6 +313,7 @@ class WatsonXInferenceAdapter(Inference, ModelRegistryHelper):
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
model_obj = await self.model_store.get_model(model)
params = await prepare_openai_completion_params(

21
llama_stack/providers/remote/vector_io/chroma/chroma.py
View file

@ -23,6 +23,7 @@ from llama_stack.apis.vector_io import (
VectorStoreObject,
VectorStoreSearchResponsePage,
)
from llama_stack.apis.vector_io.vector_io import VectorStoreChunkingStrategy, VectorStoreFileObject
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.inline.vector_io.chroma import ChromaVectorIOConfig as InlineChromaVectorIOConfig
from llama_stack.providers.utils.memory.vector_store import (
@ -104,6 +105,17 @@ class ChromaIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in Chroma")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in Chroma")
class ChromaVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
def __init__(
@ -241,3 +253,12 @@ class ChromaVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
rewrite_query: bool | None = False,
) -> VectorStoreSearchResponsePage:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Chroma")
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Chroma")

21
llama_stack/providers/remote/vector_io/milvus/milvus.py
View file

@ -25,6 +25,7 @@ from llama_stack.apis.vector_io import (
VectorStoreObject,
VectorStoreSearchResponsePage,
)
from llama_stack.apis.vector_io.vector_io import VectorStoreChunkingStrategy, VectorStoreFileObject
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.inline.vector_io.milvus import MilvusVectorIOConfig as InlineMilvusVectorIOConfig
from llama_stack.providers.utils.memory.vector_store import (
@ -102,6 +103,17 @@ class MilvusIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in Milvus")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in Milvus")
class MilvusVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
def __init__(
@ -240,6 +252,15 @@ class MilvusVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
) -> VectorStoreSearchResponsePage:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Qdrant")
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Milvus")
def generate_chunk_id(document_id: str, chunk_text: str) -> str:
"""Generate a unique chunk ID using a hash of document ID and chunk text."""

11
llama_stack/providers/remote/vector_io/pgvector/pgvector.py
View file

@ -128,6 +128,17 @@ class PGVectorIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in PGVector")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in PGVector")
async def delete(self):
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
cur.execute(f"DROP TABLE IF EXISTS {self.table_name}")

21
llama_stack/providers/remote/vector_io/qdrant/qdrant.py
View file

@ -23,6 +23,7 @@ from llama_stack.apis.vector_io import (
VectorStoreObject,
VectorStoreSearchResponsePage,
)
from llama_stack.apis.vector_io.vector_io import VectorStoreChunkingStrategy, VectorStoreFileObject
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.inline.vector_io.qdrant import QdrantVectorIOConfig as InlineQdrantVectorIOConfig
from llama_stack.providers.utils.memory.vector_store import (
@ -111,6 +112,17 @@ class QdrantIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in Qdrant")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in Qdrant")
async def delete(self):
await self.client.delete_collection(collection_name=self.collection_name)
@ -241,3 +253,12 @@ class QdrantVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
rewrite_query: bool | None = False,
) -> VectorStoreSearchResponsePage:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Qdrant")
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Qdrant")

11
llama_stack/providers/remote/vector_io/weaviate/weaviate.py
View file

@ -92,6 +92,17 @@ class WeaviateIndex(EmbeddingIndex):
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in Weaviate")
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError("Hybrid search is not supported in Weaviate")
class WeaviateVectorIOAdapter(
VectorIO,

16
llama_stack/providers/utils/inference/litellm_openai_mixin.py
View file

@ -4,8 +4,6 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import base64
import struct
from collections.abc import AsyncGenerator, AsyncIterator
from typing import Any
@ -37,7 +35,6 @@ from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionChunk,
OpenAICompletion,
OpenAIEmbeddingData,
OpenAIEmbeddingsResponse,
OpenAIEmbeddingUsage,
OpenAIMessageParam,
@ -48,6 +45,7 @@ from llama_stack.distribution.request_headers import NeedsRequestProviderData
from llama_stack.log import get_logger
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
from llama_stack.providers.utils.inference.openai_compat import (
b64_encode_openai_embeddings_response,
convert_message_to_openai_dict_new,
convert_openai_chat_completion_choice,
convert_openai_chat_completion_stream,
@ -293,16 +291,7 @@ class LiteLLMOpenAIMixin(
)
# Convert response to OpenAI format
data = []
for i, embedding_data in enumerate(response["data"]):
# we encode to base64 if the encoding format is base64 in the request
if encoding_format == "base64":
byte_data = b"".join(struct.pack("f", f) for f in embedding_data["embedding"])
embedding = base64.b64encode(byte_data).decode("utf-8")
else:
embedding = embedding_data["embedding"]
data.append(OpenAIEmbeddingData(embedding=embedding, index=i))
data = b64_encode_openai_embeddings_response(response.data, encoding_format)
usage = OpenAIEmbeddingUsage(
prompt_tokens=response["usage"]["prompt_tokens"],
@ -336,6 +325,7 @@ class LiteLLMOpenAIMixin(
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
model_obj = await self.model_store.get_model(model)
params = await prepare_openai_completion_params(

56
llama_stack/providers/utils/inference/openai_compat.py
View file

@ -3,8 +3,10 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import base64
import json
import logging
import struct
import time
import uuid
import warnings
@ -108,6 +110,7 @@ from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAICompletion,
OpenAICompletionChoice,
OpenAIEmbeddingData,
OpenAIMessageParam,
OpenAIResponseFormatParam,
ToolConfig,
@ -1287,6 +1290,7 @@ class OpenAICompletionToLlamaStackMixin:
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
suffix: str | None = None,
) -> OpenAICompletion:
if stream:
raise ValueError(f"{self.__class__.__name__} doesn't support streaming openai completions")
@ -1483,3 +1487,55 @@ class OpenAIChatCompletionToLlamaStackMixin:
model=model,
object="chat.completion",
)
def prepare_openai_embeddings_params(
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
):
if model is None:
raise ValueError("Model must be provided for embeddings")
input_list = [input] if isinstance(input, str) else input
params: dict[str, Any] = {
"model": model,
"input": input_list,
}
if encoding_format is not None:
params["encoding_format"] = encoding_format
if dimensions is not None:
params["dimensions"] = dimensions
if user is not None:
params["user"] = user
return params
def b64_encode_openai_embeddings_response(
response_data: dict, encoding_format: str | None = "float"
) -> list[OpenAIEmbeddingData]:
"""
Process the OpenAI embeddings response to encode the embeddings in base64 format if specified.
"""
data = []
for i, embedding_data in enumerate(response_data):
if encoding_format == "base64":
byte_array = bytearray()
for embedding_value in embedding_data.embedding:
byte_array.extend(struct.pack("f", float(embedding_value)))
response_embedding = base64.b64encode(byte_array).decode("utf-8")
else:
response_embedding = embedding_data.embedding
data.append(
OpenAIEmbeddingData(
embedding=response_embedding,
index=i,
)
)
return data

97
llama_stack/providers/utils/memory/openai_vector_store_mixin.py
View file

@ -5,11 +5,13 @@
# the root directory of this source tree.
import logging
import mimetypes
import time
import uuid
from abc import ABC, abstractmethod
from typing import Any
from llama_stack.apis.files import Files
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import (
QueryChunksResponse,
@ -20,6 +22,15 @@ from llama_stack.apis.vector_io import (
VectorStoreSearchResponse,
VectorStoreSearchResponsePage,
)
from llama_stack.apis.vector_io.vector_io import (
Chunk,
VectorStoreChunkingStrategy,
VectorStoreChunkingStrategyAuto,
VectorStoreChunkingStrategyStatic,
VectorStoreFileLastError,
VectorStoreFileObject,
)
from llama_stack.providers.utils.memory.vector_store import content_from_data_and_mime_type, make_overlapped_chunks
logger = logging.getLogger(__name__)
@ -36,6 +47,7 @@ class OpenAIVectorStoreMixin(ABC):
# These should be provided by the implementing class
openai_vector_stores: dict[str, dict[str, Any]]
files_api: Files | None
@abstractmethod
async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
@ -67,6 +79,16 @@ class OpenAIVectorStoreMixin(ABC):
"""Unregister a vector database (provider-specific implementation)."""
pass
@abstractmethod
async def insert_chunks(
self,
vector_db_id: str,
chunks: list[Chunk],
ttl_seconds: int | None = None,
) -> None:
"""Insert chunks into a vector database (provider-specific implementation)."""
pass
@abstractmethod
async def query_chunks(
self, vector_db_id: str, query: Any, params: dict[str, Any] | None = None
@ -383,3 +405,78 @@ class OpenAIVectorStoreMixin(ABC):
if metadata[key] != value:
return False
return True
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
attributes = attributes or {}
chunking_strategy = chunking_strategy or VectorStoreChunkingStrategyAuto()
vector_store_file_object = VectorStoreFileObject(
id=file_id,
attributes=attributes,
chunking_strategy=chunking_strategy,
created_at=int(time.time()),
status="in_progress",
vector_store_id=vector_store_id,
)
if not hasattr(self, "files_api") or not self.files_api:
vector_store_file_object.status = "failed"
vector_store_file_object.last_error = VectorStoreFileLastError(
code="server_error",
message="Files API is not available",
)
return vector_store_file_object
if isinstance(chunking_strategy, VectorStoreChunkingStrategyStatic):
max_chunk_size_tokens = chunking_strategy.static.max_chunk_size_tokens
chunk_overlap_tokens = chunking_strategy.static.chunk_overlap_tokens
else:
# Default values from OpenAI API spec
max_chunk_size_tokens = 800
chunk_overlap_tokens = 400
try:
file_response = await self.files_api.openai_retrieve_file(file_id)
mime_type, _ = mimetypes.guess_type(file_response.filename)
content_response = await self.files_api.openai_retrieve_file_content(file_id)
content = content_from_data_and_mime_type(content_response.body, mime_type)
chunks = make_overlapped_chunks(
file_id,
content,
max_chunk_size_tokens,
chunk_overlap_tokens,
attributes,
)
if not chunks:
vector_store_file_object.status = "failed"
vector_store_file_object.last_error = VectorStoreFileLastError(
code="server_error",
message="No chunks were generated from the file",
)
return vector_store_file_object
await self.insert_chunks(
vector_db_id=vector_store_id,
chunks=chunks,
)
except Exception as e:
logger.error(f"Error attaching file to vector store: {e}")
vector_store_file_object.status = "failed"
vector_store_file_object.last_error = VectorStoreFileLastError(
code="server_error",
message=str(e),
)
return vector_store_file_object
vector_store_file_object.status = "completed"
return vector_store_file_object

55
llama_stack/providers/utils/memory/vector_store.py
View file

@ -32,6 +32,10 @@ from llama_stack.providers.utils.inference.prompt_adapter import (
log = logging.getLogger(__name__)
# Constants for reranker types
RERANKER_TYPE_RRF = "rrf"
RERANKER_TYPE_WEIGHTED = "weighted"
def parse_pdf(data: bytes) -> str:
# For PDF and DOC/DOCX files, we can't reliably convert to string
@ -72,16 +76,18 @@ def content_from_data(data_url: str) -> str:
data = unquote(data)
encoding = parts["encoding"] or "utf-8"
data = data.encode(encoding)
return content_from_data_and_mime_type(data, parts["mimetype"], parts.get("encoding", None))
encoding = parts["encoding"]
if not encoding:
import chardet
detected = chardet.detect(data)
encoding = detected["encoding"]
def content_from_data_and_mime_type(data: bytes | str, mime_type: str | None, encoding: str | None = None) -> str:
if isinstance(data, bytes):
if not encoding:
import chardet
mime_type = parts["mimetype"]
mime_category = mime_type.split("/")[0]
detected = chardet.detect(data)
encoding = detected["encoding"]
mime_category = mime_type.split("/")[0] if mime_type else None
if mime_category == "text":
# For text-based files (including CSV, MD)
return data.decode(encoding)
@ -200,6 +206,18 @@ class EmbeddingIndex(ABC):
async def query_keyword(self, query_string: str, k: int, score_threshold: float) -> QueryChunksResponse:
raise NotImplementedError()
@abstractmethod
async def query_hybrid(
self,
embedding: NDArray,
query_string: str,
k: int,
score_threshold: float,
reranker_type: str,
reranker_params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
raise NotImplementedError()
@abstractmethod
async def delete(self):
raise NotImplementedError()
@ -243,10 +261,29 @@ class VectorDBWithIndex:
k = params.get("max_chunks", 3)
mode = params.get("mode")
score_threshold = params.get("score_threshold", 0.0)
# Get ranker configuration
ranker = params.get("ranker")
if ranker is None:
# Default to RRF with impact_factor=60.0
reranker_type = RERANKER_TYPE_RRF
reranker_params = {"impact_factor": 60.0}
else:
reranker_type = ranker.type
reranker_params = (
{"impact_factor": ranker.impact_factor} if ranker.type == RERANKER_TYPE_RRF else {"alpha": ranker.alpha}
)
query_string = interleaved_content_as_str(query)
if mode == "keyword":
return await self.index.query_keyword(query_string, k, score_threshold)
# Calculate embeddings for both vector and hybrid modes
embeddings_response = await self.inference_api.embeddings(self.vector_db.embedding_model, [query_string])
query_vector = np.array(embeddings_response.embeddings[0], dtype=np.float32)
if mode == "hybrid":
return await self.index.query_hybrid(
query_vector, query_string, k, score_threshold, reranker_type, reranker_params
)
else:
embeddings_response = await self.inference_api.embeddings(self.vector_db.embedding_model, [query_string])
query_vector = np.array(embeddings_response.embeddings[0], dtype=np.float32)
return await self.index.query_vector(query_vector, k, score_threshold)

2
llama_stack/templates/ollama/build.yaml
View file

@ -23,6 +23,8 @@ distribution_spec:
- inline::basic
- inline::llm-as-judge
- inline::braintrust
files:
- inline::localfs
post_training:
- inline::huggingface
tool_runtime:

9
llama_stack/templates/ollama/ollama.py
View file

@ -13,6 +13,7 @@ from llama_stack.distribution.datatypes import (
ShieldInput,
ToolGroupInput,
)
from llama_stack.providers.inline.files.localfs.config import LocalfsFilesImplConfig
from llama_stack.providers.inline.post_training.huggingface import HuggingFacePostTrainingConfig
from llama_stack.providers.inline.vector_io.faiss.config import FaissVectorIOConfig
from llama_stack.providers.remote.inference.ollama import OllamaImplConfig
@ -29,6 +30,7 @@ def get_distribution_template() -> DistributionTemplate:
"eval": ["inline::meta-reference"],
"datasetio": ["remote::huggingface", "inline::localfs"],
"scoring": ["inline::basic", "inline::llm-as-judge", "inline::braintrust"],
"files": ["inline::localfs"],
"post_training": ["inline::huggingface"],
"tool_runtime": [
"remote::brave-search",
@ -49,6 +51,11 @@ def get_distribution_template() -> DistributionTemplate:
provider_type="inline::faiss",
config=FaissVectorIOConfig.sample_run_config(f"~/.llama/distributions/{name}"),
)
files_provider = Provider(
provider_id="meta-reference-files",
provider_type="inline::localfs",
config=LocalfsFilesImplConfig.sample_run_config(f"~/.llama/distributions/{name}"),
)
posttraining_provider = Provider(
provider_id="huggingface",
provider_type="inline::huggingface",
@ -98,6 +105,7 @@ def get_distribution_template() -> DistributionTemplate:
provider_overrides={
"inference": [inference_provider],
"vector_io": [vector_io_provider_faiss],
"files": [files_provider],
"post_training": [posttraining_provider],
},
default_models=[inference_model, embedding_model],
@ -107,6 +115,7 @@ def get_distribution_template() -> DistributionTemplate:
provider_overrides={
"inference": [inference_provider],
"vector_io": [vector_io_provider_faiss],
"files": [files_provider],
"post_training": [posttraining_provider],
"safety": [
Provider(

9
llama_stack/templates/ollama/run-with-safety.yaml
View file

@ -4,6 +4,7 @@ apis:
- agents
- datasetio
- eval
- files
- inference
- post_training
- safety
@ -84,6 +85,14 @@ providers:
provider_type: inline::braintrust
config:
openai_api_key: ${env.OPENAI_API_KEY:}
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:~/.llama/distributions/ollama/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ollama}/files_metadata.db
post_training:
- provider_id: huggingface
provider_type: inline::huggingface

9
llama_stack/templates/ollama/run.yaml
View file

@ -4,6 +4,7 @@ apis:
- agents
- datasetio
- eval
- files
- inference
- post_training
- safety
@ -82,6 +83,14 @@ providers:
provider_type: inline::braintrust
config:
openai_api_key: ${env.OPENAI_API_KEY:}
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:~/.llama/distributions/ollama/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ollama}/files_metadata.db
post_training:
- provider_id: huggingface
provider_type: inline::huggingface

2
llama_stack/templates/starter/build.yaml
View file

@ -17,6 +17,8 @@ distribution_spec:
- inline::sqlite-vec
- remote::chromadb
- remote::pgvector
files:
- inline::localfs
safety:
- inline::llama-guard
agents:

9
llama_stack/templates/starter/run.yaml
View file

@ -4,6 +4,7 @@ apis:
- agents
- datasetio
- eval
- files
- inference
- safety
- scoring
@ -75,6 +76,14 @@ providers:
db: ${env.PGVECTOR_DB:}
user: ${env.PGVECTOR_USER:}
password: ${env.PGVECTOR_PASSWORD:}
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:~/.llama/distributions/starter/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/starter}/files_metadata.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard

8
llama_stack/templates/starter/starter.py
View file

@ -12,6 +12,7 @@ from llama_stack.distribution.datatypes import (
ShieldInput,
ToolGroupInput,
)
from llama_stack.providers.inline.files.localfs.config import LocalfsFilesImplConfig
from llama_stack.providers.inline.inference.sentence_transformers import (
SentenceTransformersInferenceConfig,
)
@ -134,6 +135,7 @@ def get_distribution_template() -> DistributionTemplate:
providers = {
"inference": ([p.provider_type for p in inference_providers] + ["inline::sentence-transformers"]),
"vector_io": ["inline::sqlite-vec", "remote::chromadb", "remote::pgvector"],
"files": ["inline::localfs"],
"safety": ["inline::llama-guard"],
"agents": ["inline::meta-reference"],
"telemetry": ["inline::meta-reference"],
@ -170,6 +172,11 @@ def get_distribution_template() -> DistributionTemplate:
),
),
]
files_provider = Provider(
provider_id="meta-reference-files",
provider_type="inline::localfs",
config=LocalfsFilesImplConfig.sample_run_config(f"~/.llama/distributions/{name}"),
)
embedding_provider = Provider(
provider_id="sentence-transformers",
provider_type="inline::sentence-transformers",
@ -212,6 +219,7 @@ def get_distribution_template() -> DistributionTemplate:
provider_overrides={
"inference": inference_providers + [embedding_provider],
"vector_io": vector_io_providers,
"files": [files_provider],
},
default_models=default_models + [embedding_model],
default_tool_groups=default_tool_groups,