mirror of
https://github.com/meta-llama/llama-stack.git
synced 2025-10-08 21:04:39 +00:00
Removing custom tool and agent utilities and moving them client side
This commit is contained in:
Ashwin Bharambe
parent
fa864f70da
commit
099ac81bc7
17 changed files with 100 additions and 392 deletions
|
|
@ -25,14 +25,10 @@ from llama_stack.apis.inference import * # noqa: F403
|
|||
from llama_stack.apis.memory import * # noqa: F403
|
||||
from llama_stack.apis.safety import * # noqa: F403
|
||||
|
||||
from llama_stack.tools.base import BaseTool
|
||||
from llama_stack.tools.builtin import (
|
||||
interpret_content_as_attachment,
|
||||
SingleMessageBuiltinTool,
|
||||
)
|
||||
|
||||
from .rag.context_retriever import generate_rag_query
|
||||
from .safety import SafetyException, ShieldRunnerMixin
|
||||
from .tools.base import BaseTool
|
||||
from .tools.builtin import interpret_content_as_attachment, SingleMessageBuiltinTool
|
||||
|
||||
|
||||
def make_random_string(length: int = 8):
|
||||
|
|
|
|||
|
|
@ -14,16 +14,16 @@ from llama_stack.apis.inference import Inference
|
|||
from llama_stack.apis.memory import Memory
|
||||
from llama_stack.apis.safety import Safety
|
||||
from llama_stack.apis.agents import * # noqa: F403
|
||||
from llama_stack.tools.builtin import (
|
||||
|
||||
from .agent_instance import ChatAgent
|
||||
from .config import MetaReferenceImplConfig
|
||||
from .tools.builtin import (
|
||||
CodeInterpreterTool,
|
||||
PhotogenTool,
|
||||
SearchTool,
|
||||
WolframAlphaTool,
|
||||
)
|
||||
from llama_stack.tools.safety import with_safety
|
||||
|
||||
from .agent_instance import ChatAgent
|
||||
from .config import MetaReferenceImplConfig
|
||||
from .tools.safety import with_safety
|
||||
|
||||
|
||||
logger = logging.getLogger()
|
||||
|
|
|
|||
|
|
@ -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 unittest
|
||||
|
||||
from llama_models.llama3.api.datatypes import (
|
||||
Attachment,
|
||||
BuiltinTool,
|
||||
CompletionMessage,
|
||||
StopReason,
|
||||
ToolCall,
|
||||
)
|
||||
|
||||
from ..tools.builtin import CodeInterpreterTool
|
||||
|
||||
|
||||
class TestCodeInterpreter(unittest.IsolatedAsyncioTestCase):
|
||||
async def test_matplotlib(self):
|
||||
tool = CodeInterpreterTool()
|
||||
code = """
|
||||
import matplotlib.pyplot as plt
|
||||
import numpy as np
|
||||
|
||||
x = np.array([1, 1])
|
||||
y = np.array([0, 10])
|
||||
|
||||
plt.plot(x, y)
|
||||
plt.title('x = 1')
|
||||
plt.xlabel('x')
|
||||
plt.ylabel('y')
|
||||
plt.grid(True)
|
||||
plt.axvline(x=1, color='r')
|
||||
plt.show()
|
||||
"""
|
||||
message = CompletionMessage(
|
||||
role="assistant",
|
||||
content="",
|
||||
tool_calls=[
|
||||
ToolCall(
|
||||
call_id="call_id",
|
||||
tool_name=BuiltinTool.code_interpreter,
|
||||
arguments={"code": code},
|
||||
)
|
||||
],
|
||||
stop_reason=StopReason.end_of_message,
|
||||
)
|
||||
ret = await tool.run([message])
|
||||
|
||||
self.assertEqual(len(ret), 1)
|
||||
|
||||
output = ret[0].content
|
||||
self.assertIsInstance(output, Attachment)
|
||||
self.assertEqual(output.mime_type, "image/png")
|
||||
|
||||
async def test_path_unlink(self):
|
||||
tool = CodeInterpreterTool()
|
||||
code = """
|
||||
import os
|
||||
from pathlib import Path
|
||||
import tempfile
|
||||
|
||||
dpath = Path(os.environ["MPLCONFIGDIR"])
|
||||
with open(dpath / "test", "w") as f:
|
||||
f.write("hello")
|
||||
|
||||
Path(dpath / "test").unlink()
|
||||
print("_OK_")
|
||||
"""
|
||||
message = CompletionMessage(
|
||||
role="assistant",
|
||||
content="",
|
||||
tool_calls=[
|
||||
ToolCall(
|
||||
call_id="call_id",
|
||||
tool_name=BuiltinTool.code_interpreter,
|
||||
arguments={"code": code},
|
||||
)
|
||||
],
|
||||
stop_reason=StopReason.end_of_message,
|
||||
)
|
||||
ret = await tool.run([message])
|
||||
|
||||
self.assertEqual(len(ret), 1)
|
||||
|
||||
output = ret[0].content
|
||||
self.assertTrue("_OK_" in output)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
|
|
@ -0,0 +1,5 @@
|
|||
# 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.
|
||||
|
|
@ -0,0 +1,20 @@
|
|||
# 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 abc import ABC, abstractmethod
|
||||
from typing import List
|
||||
|
||||
from llama_stack.apis.inference import Message
|
||||
|
||||
|
||||
class BaseTool(ABC):
|
||||
@abstractmethod
|
||||
def get_name(self) -> str:
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
async def run(self, messages: List[Message]) -> List[Message]:
|
||||
raise NotImplementedError
|
||||
|
|
@ -0,0 +1,375 @@
|
|||
# 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 tempfile
|
||||
|
||||
from abc import abstractmethod
|
||||
from typing import List, Optional
|
||||
|
||||
import requests
|
||||
from termcolor import cprint
|
||||
|
||||
from .ipython_tool.code_execution import (
|
||||
CodeExecutionContext,
|
||||
CodeExecutionRequest,
|
||||
CodeExecutor,
|
||||
TOOLS_ATTACHMENT_KEY_REGEX,
|
||||
)
|
||||
|
||||
from llama_stack.apis.inference import * # noqa: F403
|
||||
from llama_stack.apis.agents import * # noqa: F403
|
||||
|
||||
from .base import BaseTool
|
||||
|
||||
|
||||
def interpret_content_as_attachment(content: str) -> Optional[Attachment]:
|
||||
match = re.search(TOOLS_ATTACHMENT_KEY_REGEX, content)
|
||||
if match:
|
||||
snippet = match.group(1)
|
||||
data = json.loads(snippet)
|
||||
return Attachment(
|
||||
content=URL(uri="file://" + data["filepath"]), mime_type=data["mimetype"]
|
||||
)
|
||||
|
||||
return None
|
||||
|
||||
|
||||
class SingleMessageBuiltinTool(BaseTool):
|
||||
async def run(self, messages: List[CompletionMessage]) -> List[ToolResponseMessage]:
|
||||
assert len(messages) == 1, f"Expected single message, got {len(messages)}"
|
||||
|
||||
message = messages[0]
|
||||
assert len(message.tool_calls) == 1, "Expected a single tool call"
|
||||
|
||||
tool_call = messages[0].tool_calls[0]
|
||||
|
||||
query = tool_call.arguments["query"]
|
||||
response: str = await self.run_impl(query)
|
||||
|
||||
message = ToolResponseMessage(
|
||||
call_id=tool_call.call_id,
|
||||
tool_name=tool_call.tool_name,
|
||||
content=response,
|
||||
)
|
||||
return [message]
|
||||
|
||||
@abstractmethod
|
||||
async def run_impl(self, query: str) -> str:
|
||||
raise NotImplementedError()
|
||||
|
||||
|
||||
class PhotogenTool(SingleMessageBuiltinTool):
|
||||
def __init__(self, dump_dir: str) -> None:
|
||||
self.dump_dir = dump_dir
|
||||
|
||||
def get_name(self) -> str:
|
||||
return BuiltinTool.photogen.value
|
||||
|
||||
async def run_impl(self, query: str) -> str:
|
||||
"""
|
||||
Implement this to give the model an ability to generate images.
|
||||
|
||||
Return:
|
||||
info = {
|
||||
"filepath": str(image_filepath),
|
||||
"mimetype": "image/png",
|
||||
}
|
||||
"""
|
||||
raise NotImplementedError()
|
||||
|
||||
|
||||
class SearchTool(SingleMessageBuiltinTool):
|
||||
def __init__(self, engine: SearchEngineType, api_key: str, **kwargs) -> None:
|
||||
self.api_key = api_key
|
||||
if engine == SearchEngineType.bing:
|
||||
self.engine = BingSearch(api_key, **kwargs)
|
||||
elif engine == SearchEngineType.brave:
|
||||
self.engine = BraveSearch(api_key, **kwargs)
|
||||
else:
|
||||
raise ValueError(f"Unknown search engine: {engine}")
|
||||
|
||||
def get_name(self) -> str:
|
||||
return BuiltinTool.brave_search.value
|
||||
|
||||
async def run_impl(self, query: str) -> str:
|
||||
return await self.engine.search(query)
|
||||
|
||||
|
||||
class BingSearch:
|
||||
def __init__(self, api_key: str, top_k: int = 3, **kwargs) -> None:
|
||||
self.api_key = api_key
|
||||
self.top_k = top_k
|
||||
|
||||
async def search(self, query: str) -> str:
|
||||
url = "https://api.bing.microsoft.com/v7.0/search"
|
||||
headers = {
|
||||
"Ocp-Apim-Subscription-Key": self.api_key,
|
||||
}
|
||||
params = {
|
||||
"count": self.top_k,
|
||||
"textDecorations": True,
|
||||
"textFormat": "HTML",
|
||||
"q": query,
|
||||
}
|
||||
|
||||
response = requests.get(url=url, params=params, headers=headers)
|
||||
response.raise_for_status()
|
||||
clean = self._clean_response(response.json())
|
||||
return json.dumps(clean)
|
||||
|
||||
def _clean_response(self, search_response):
|
||||
clean_response = []
|
||||
query = search_response["queryContext"]["originalQuery"]
|
||||
if "webPages" in search_response:
|
||||
pages = search_response["webPages"]["value"]
|
||||
for p in pages:
|
||||
selected_keys = {"name", "url", "snippet"}
|
||||
clean_response.append(
|
||||
{k: v for k, v in p.items() if k in selected_keys}
|
||||
)
|
||||
if "news" in search_response:
|
||||
clean_news = []
|
||||
news = search_response["news"]["value"]
|
||||
for n in news:
|
||||
selected_keys = {"name", "url", "description"}
|
||||
clean_news.append({k: v for k, v in n.items() if k in selected_keys})
|
||||
|
||||
clean_response.append(clean_news)
|
||||
|
||||
return {"query": query, "top_k": clean_response}
|
||||
|
||||
|
||||
class BraveSearch:
|
||||
def __init__(self, api_key: str) -> None:
|
||||
self.api_key = api_key
|
||||
|
||||
async def search(self, query: str) -> str:
|
||||
url = "https://api.search.brave.com/res/v1/web/search"
|
||||
headers = {
|
||||
"X-Subscription-Token": self.api_key,
|
||||
"Accept-Encoding": "gzip",
|
||||
"Accept": "application/json",
|
||||
}
|
||||
payload = {"q": query}
|
||||
response = requests.get(url=url, params=payload, headers=headers)
|
||||
return json.dumps(self._clean_brave_response(response.json()))
|
||||
|
||||
def _clean_brave_response(self, search_response, top_k=3):
|
||||
query = None
|
||||
clean_response = []
|
||||
if "query" in search_response:
|
||||
if "original" in search_response["query"]:
|
||||
query = search_response["query"]["original"]
|
||||
if "mixed" in search_response:
|
||||
mixed_results = search_response["mixed"]
|
||||
for m in mixed_results["main"][:top_k]:
|
||||
r_type = m["type"]
|
||||
results = search_response[r_type]["results"]
|
||||
if r_type == "web":
|
||||
# For web data - add a single output from the search
|
||||
idx = m["index"]
|
||||
selected_keys = [
|
||||
"type",
|
||||
"title",
|
||||
"url",
|
||||
"description",
|
||||
"date",
|
||||
"extra_snippets",
|
||||
]
|
||||
cleaned = {
|
||||
k: v for k, v in results[idx].items() if k in selected_keys
|
||||
}
|
||||
elif r_type == "faq":
|
||||
# For faw data - take a list of all the questions & answers
|
||||
selected_keys = ["type", "question", "answer", "title", "url"]
|
||||
cleaned = []
|
||||
for q in results:
|
||||
cleaned.append(
|
||||
{k: v for k, v in q.items() if k in selected_keys}
|
||||
)
|
||||
elif r_type == "infobox":
|
||||
idx = m["index"]
|
||||
selected_keys = [
|
||||
"type",
|
||||
"title",
|
||||
"url",
|
||||
"description",
|
||||
"long_desc",
|
||||
]
|
||||
cleaned = {
|
||||
k: v for k, v in results[idx].items() if k in selected_keys
|
||||
}
|
||||
elif r_type == "videos":
|
||||
selected_keys = [
|
||||
"type",
|
||||
"url",
|
||||
"title",
|
||||
"description",
|
||||
"date",
|
||||
]
|
||||
cleaned = []
|
||||
for q in results:
|
||||
cleaned.append(
|
||||
{k: v for k, v in q.items() if k in selected_keys}
|
||||
)
|
||||
elif r_type == "locations":
|
||||
# For faw data - take a list of all the questions & answers
|
||||
selected_keys = [
|
||||
"type",
|
||||
"title",
|
||||
"url",
|
||||
"description",
|
||||
"coordinates",
|
||||
"postal_address",
|
||||
"contact",
|
||||
"rating",
|
||||
"distance",
|
||||
"zoom_level",
|
||||
]
|
||||
cleaned = []
|
||||
for q in results:
|
||||
cleaned.append(
|
||||
{k: v for k, v in q.items() if k in selected_keys}
|
||||
)
|
||||
elif r_type == "news":
|
||||
# For faw data - take a list of all the questions & answers
|
||||
selected_keys = [
|
||||
"type",
|
||||
"title",
|
||||
"url",
|
||||
"description",
|
||||
]
|
||||
cleaned = []
|
||||
for q in results:
|
||||
cleaned.append(
|
||||
{k: v for k, v in q.items() if k in selected_keys}
|
||||
)
|
||||
else:
|
||||
cleaned = []
|
||||
|
||||
clean_response.append(cleaned)
|
||||
|
||||
return {"query": query, "top_k": clean_response}
|
||||
|
||||
|
||||
class WolframAlphaTool(SingleMessageBuiltinTool):
|
||||
def __init__(self, api_key: str) -> None:
|
||||
self.api_key = api_key
|
||||
self.url = "https://api.wolframalpha.com/v2/query"
|
||||
|
||||
def get_name(self) -> str:
|
||||
return BuiltinTool.wolfram_alpha.value
|
||||
|
||||
async def run_impl(self, query: str) -> str:
|
||||
params = {
|
||||
"input": query,
|
||||
"appid": self.api_key,
|
||||
"format": "plaintext",
|
||||
"output": "json",
|
||||
}
|
||||
response = requests.get(
|
||||
self.url,
|
||||
params=params,
|
||||
)
|
||||
|
||||
return json.dumps(self._clean_wolfram_alpha_response(response.json()))
|
||||
|
||||
def _clean_wolfram_alpha_response(self, wa_response):
|
||||
remove = {
|
||||
"queryresult": [
|
||||
"datatypes",
|
||||
"error",
|
||||
"timedout",
|
||||
"timedoutpods",
|
||||
"numpods",
|
||||
"timing",
|
||||
"parsetiming",
|
||||
"parsetimedout",
|
||||
"recalculate",
|
||||
"id",
|
||||
"host",
|
||||
"server",
|
||||
"related",
|
||||
"version",
|
||||
{
|
||||
"pods": [
|
||||
"scanner",
|
||||
"id",
|
||||
"error",
|
||||
"expressiontypes",
|
||||
"states",
|
||||
"infos",
|
||||
"position",
|
||||
"numsubpods",
|
||||
]
|
||||
},
|
||||
"assumptions",
|
||||
],
|
||||
}
|
||||
for main_key in remove:
|
||||
for key_to_remove in remove[main_key]:
|
||||
try:
|
||||
if key_to_remove == "assumptions":
|
||||
if "assumptions" in wa_response[main_key]:
|
||||
del wa_response[main_key][key_to_remove]
|
||||
if isinstance(key_to_remove, dict):
|
||||
for sub_key in key_to_remove:
|
||||
if sub_key == "pods":
|
||||
for i in range(len(wa_response[main_key][sub_key])):
|
||||
if (
|
||||
wa_response[main_key][sub_key][i]["title"]
|
||||
== "Result"
|
||||
):
|
||||
del wa_response[main_key][sub_key][i + 1 :]
|
||||
break
|
||||
sub_items = wa_response[main_key][sub_key]
|
||||
for i in range(len(sub_items)):
|
||||
for sub_key_to_remove in key_to_remove[sub_key]:
|
||||
if sub_key_to_remove in sub_items[i]:
|
||||
del sub_items[i][sub_key_to_remove]
|
||||
elif key_to_remove in wa_response[main_key]:
|
||||
del wa_response[main_key][key_to_remove]
|
||||
except KeyError:
|
||||
pass
|
||||
return wa_response
|
||||
|
||||
|
||||
class CodeInterpreterTool(BaseTool):
|
||||
def __init__(self) -> None:
|
||||
ctx = CodeExecutionContext(
|
||||
matplotlib_dump_dir=tempfile.mkdtemp(),
|
||||
)
|
||||
self.code_executor = CodeExecutor(ctx)
|
||||
|
||||
def get_name(self) -> str:
|
||||
return BuiltinTool.code_interpreter.value
|
||||
|
||||
async def run(self, messages: List[CompletionMessage]) -> List[ToolResponseMessage]:
|
||||
message = messages[0]
|
||||
assert len(message.tool_calls) == 1, "Expected a single tool call"
|
||||
|
||||
tool_call = messages[0].tool_calls[0]
|
||||
script = tool_call.arguments["code"]
|
||||
|
||||
req = CodeExecutionRequest(scripts=[script])
|
||||
res = self.code_executor.execute(req)
|
||||
|
||||
pieces = [res["process_status"]]
|
||||
for out_type in ["stdout", "stderr"]:
|
||||
res_out = res[out_type]
|
||||
if res_out != "":
|
||||
pieces.extend([f"[{out_type}]", res_out, f"[/{out_type}]"])
|
||||
if out_type == "stderr":
|
||||
cprint(f"ipython tool error: ↓\n{res_out}", color="red")
|
||||
|
||||
message = ToolResponseMessage(
|
||||
call_id=tool_call.call_id,
|
||||
tool_name=tool_call.tool_name,
|
||||
content="\n".join(pieces),
|
||||
)
|
||||
return [message]
|
||||
|
|
@ -0,0 +1,5 @@
|
|||
# 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.
|
||||
|
|
@ -0,0 +1,133 @@
|
|||
# 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 errno
|
||||
|
||||
# Disabling potentially dangerous functions
|
||||
import os as _os
|
||||
from functools import partial
|
||||
|
||||
os_funcs_to_disable = [
|
||||
"kill",
|
||||
"system",
|
||||
"putenv",
|
||||
"remove",
|
||||
"removedirs",
|
||||
"rmdir",
|
||||
"fchdir",
|
||||
"setuid",
|
||||
"fork",
|
||||
"forkpty",
|
||||
"killpg",
|
||||
"rename",
|
||||
"renames",
|
||||
"truncate",
|
||||
"replace",
|
||||
# "unlink", # Commenting as this was blocking matpltlib from rendering plots correctly
|
||||
"fchmod",
|
||||
"fchown",
|
||||
"chmod",
|
||||
"chown",
|
||||
"chroot",
|
||||
"fchdir",
|
||||
"lchflags",
|
||||
"lchmod",
|
||||
"lchown",
|
||||
"chdir",
|
||||
]
|
||||
|
||||
|
||||
def call_not_allowed(*args, **kwargs):
|
||||
raise OSError(errno.EPERM, "Call are not permitted in this environment")
|
||||
|
||||
|
||||
for func_name in os_funcs_to_disable:
|
||||
if hasattr(_os, func_name):
|
||||
setattr(_os, func_name, partial(call_not_allowed, _func_name=f"os.{func_name}"))
|
||||
|
||||
import shutil as _shutil
|
||||
|
||||
for func_name in ["rmtree", "move", "chown"]:
|
||||
if hasattr(_shutil, func_name):
|
||||
setattr(
|
||||
_shutil,
|
||||
func_name,
|
||||
partial(call_not_allowed, _func_name=f"shutil.{func_name}"),
|
||||
)
|
||||
|
||||
import subprocess as _subprocess
|
||||
|
||||
|
||||
def popen_not_allowed(*args, **kwargs):
|
||||
raise _subprocess.CalledProcessError(
|
||||
-1,
|
||||
args[0] if args else "unknown",
|
||||
stderr="subprocess.Popen is not allowed in this environment",
|
||||
)
|
||||
|
||||
|
||||
_subprocess.Popen = popen_not_allowed
|
||||
|
||||
|
||||
import atexit as _atexit
|
||||
import builtins as _builtins
|
||||
import io as _io
|
||||
import json as _json
|
||||
import sys as _sys
|
||||
|
||||
# NB! The following "unused" imports crucial, make sure not not to remove
|
||||
# them with linters - they're used in code_execution.py
|
||||
from contextlib import ( # noqa
|
||||
contextmanager as _contextmanager,
|
||||
redirect_stderr as _redirect_stderr,
|
||||
redirect_stdout as _redirect_stdout,
|
||||
)
|
||||
from multiprocessing.connection import Connection as _Connection
|
||||
|
||||
# Mangle imports to avoid polluting model execution namespace.
|
||||
|
||||
_IO_SINK = _io.StringIO()
|
||||
_NETWORK_TIMEOUT = 5
|
||||
_NETWORK_CONNECTIONS = None
|
||||
|
||||
|
||||
def _open_connections():
|
||||
global _NETWORK_CONNECTIONS
|
||||
if _NETWORK_CONNECTIONS is not None:
|
||||
# Ensure connections only opened once.
|
||||
return _NETWORK_CONNECTIONS
|
||||
req_w_fd, resp_r_fd = _sys.argv[1], _sys.argv[2]
|
||||
req_con = _Connection(int(req_w_fd), readable=False)
|
||||
resp_con = _Connection(int(resp_r_fd), writable=False)
|
||||
_NETWORK_CONNECTIONS = (req_con, resp_con)
|
||||
return _NETWORK_CONNECTIONS
|
||||
|
||||
|
||||
_builtins._open_connections = _open_connections
|
||||
|
||||
|
||||
@_atexit.register
|
||||
def _close_connections():
|
||||
global _NETWORK_CONNECTIONS
|
||||
if _NETWORK_CONNECTIONS is None:
|
||||
return
|
||||
for con in _NETWORK_CONNECTIONS:
|
||||
con.close()
|
||||
del _NETWORK_CONNECTIONS
|
||||
|
||||
|
||||
def _network_call(request):
|
||||
# NOTE: We communicate with the parent process in json, encoded
|
||||
# in raw bytes. We do this because native send/recv methods use
|
||||
# pickle which involves execution of arbitrary code.
|
||||
_open_connections()
|
||||
req_con, resp_con = _NETWORK_CONNECTIONS
|
||||
|
||||
req_con.send_bytes(_json.dumps(request).encode("utf-8"))
|
||||
if resp_con.poll(timeout=_NETWORK_TIMEOUT) is None:
|
||||
raise Exception(f"Network request timed out: {_json.dumps(request)}")
|
||||
else:
|
||||
return _json.loads(resp_con.recv_bytes().decode("utf-8"))
|
||||
|
|
@ -0,0 +1,256 @@
|
|||
# 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 base64
|
||||
import json
|
||||
import multiprocessing
|
||||
import os
|
||||
import re
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import textwrap
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
from datetime import datetime
|
||||
from io import BytesIO
|
||||
from pathlib import Path
|
||||
from typing import List
|
||||
|
||||
from PIL import Image
|
||||
|
||||
from .utils import get_code_env_prefix
|
||||
|
||||
TOOLS_ATTACHMENT_KEY = "__tools_attachment__"
|
||||
TOOLS_ATTACHMENT_KEY_REGEX = re.compile(r"__tools_attachment__=(\{.*?\})")
|
||||
|
||||
DIRNAME = Path(__file__).parent
|
||||
|
||||
CODE_EXEC_TIMEOUT = 20
|
||||
CODE_ENV_PREFIX = get_code_env_prefix()
|
||||
|
||||
STDOUTERR_SINK_WRAPPER_TEMPLATE = """\
|
||||
with _redirect_stdout(_IO_SINK), _redirect_stderr(_IO_SINK):
|
||||
{code}\
|
||||
"""
|
||||
|
||||
TRYEXCEPT_WRAPPER_TEMPLATE = """\
|
||||
try:
|
||||
{code}
|
||||
except:
|
||||
pass\
|
||||
"""
|
||||
|
||||
|
||||
def generate_bwrap_command(bind_dirs: List[str]) -> str:
|
||||
"""
|
||||
Generate the bwrap command string for binding all
|
||||
directories in the current directory read-only.
|
||||
"""
|
||||
bwrap_args = ""
|
||||
bwrap_args += "--ro-bind / / "
|
||||
# Add the --dev flag to mount device files
|
||||
bwrap_args += "--dev /dev "
|
||||
for d in bind_dirs:
|
||||
bwrap_args += f"--bind {d} {d} "
|
||||
|
||||
# Add the --unshare-all flag to isolate the sandbox from the rest of the system
|
||||
bwrap_args += "--unshare-all "
|
||||
# Add the --die-with-parent flag to ensure the child process dies when bwrap's parent dies
|
||||
bwrap_args += "--die-with-parent "
|
||||
return bwrap_args
|
||||
|
||||
|
||||
@dataclass
|
||||
class CodeExecutionContext:
|
||||
matplotlib_dump_dir: str
|
||||
use_proxy: bool = False
|
||||
|
||||
|
||||
@dataclass
|
||||
class CodeExecutionRequest:
|
||||
scripts: List[str]
|
||||
only_last_cell_stdouterr: bool = True
|
||||
only_last_cell_fail: bool = True
|
||||
seed: int = 0
|
||||
strip_fpaths_in_stderr: bool = True
|
||||
|
||||
|
||||
class CodeExecutor:
|
||||
def __init__(self, context: CodeExecutionContext):
|
||||
self.context = context
|
||||
|
||||
def execute(self, req: CodeExecutionRequest) -> dict:
|
||||
scripts = req.scripts
|
||||
for i in range(len(scripts) - 1):
|
||||
if req.only_last_cell_stdouterr:
|
||||
scripts[i] = STDOUTERR_SINK_WRAPPER_TEMPLATE.format(
|
||||
code=textwrap.indent(scripts[i], " " * 4)
|
||||
)
|
||||
if req.only_last_cell_fail:
|
||||
scripts[i] = TRYEXCEPT_WRAPPER_TEMPLATE.format(
|
||||
code=textwrap.indent(scripts[i], " " * 4)
|
||||
)
|
||||
|
||||
# Seeds prefix:
|
||||
seed = req.seed
|
||||
seeds_prefix = f"""\
|
||||
def _set_seeds():
|
||||
import random
|
||||
random.seed({seed})
|
||||
import numpy as np
|
||||
np.random.seed({seed})
|
||||
_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)
|
||||
|
||||
try:
|
||||
python_path = os.environ.get("PYTHONPATH", "")
|
||||
env = dict(
|
||||
os.environ,
|
||||
PYTHONHASHSEED=str(seed),
|
||||
MPLCONFIGDIR=dpath,
|
||||
MPLBACKEND="module://matplotlib_custom_backend",
|
||||
PYTHONPATH=f"{DIRNAME}:{python_path}",
|
||||
)
|
||||
stdout, stderr, returncode = do_subprocess(
|
||||
cmd=cmd,
|
||||
env=env,
|
||||
ctx=self.context,
|
||||
)
|
||||
|
||||
stderr = stderr.strip()
|
||||
if req.strip_fpaths_in_stderr:
|
||||
pattern = r'File "([^"]+)", line (\d+)'
|
||||
stderr = re.sub(pattern, r"line \2", stderr)
|
||||
|
||||
return {
|
||||
"process_status": "completed",
|
||||
"returncode": returncode,
|
||||
"stdout": stdout.strip(),
|
||||
"stderr": stderr,
|
||||
}
|
||||
|
||||
except subprocess.TimeoutExpired:
|
||||
return {
|
||||
"process_status": "timeout",
|
||||
"stdout": "Timed out",
|
||||
"stderr": "Timed out",
|
||||
}
|
||||
|
||||
except Exception as e:
|
||||
return {
|
||||
"process_status": "error",
|
||||
"error_type": type(e).__name__,
|
||||
"stderr": str(e),
|
||||
"stdout": str(e),
|
||||
}
|
||||
|
||||
|
||||
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]
|
||||
# Create a list of PIL images from the bytes objects
|
||||
images = [Image.open(BytesIO(img)) for img in images]
|
||||
# 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_dpath = Path(matplotlib_dump_dir, dump_dname)
|
||||
dump_dpath.mkdir(parents=True, exist_ok=True)
|
||||
# save image into a file
|
||||
dump_fname = f"matplotlib_{str(time.time()).replace('.', '_')}_{i}.png"
|
||||
dump_fpath = dump_dpath / dump_fname
|
||||
img.save(dump_fpath, "PNG")
|
||||
image_paths.append(str(dump_fpath))
|
||||
|
||||
# this is kind of convoluted, we send back this response to the subprocess which
|
||||
# prints it out
|
||||
info = {
|
||||
"filepath": str(image_paths[-1]),
|
||||
"mimetype": "image/png",
|
||||
}
|
||||
return f"{TOOLS_ATTACHMENT_KEY}={json.dumps(info)}"
|
||||
|
||||
|
||||
def execute_subprocess_request(request, ctx: CodeExecutionContext):
|
||||
"Route requests from the subprocess (via network Pipes) to the internet/tools."
|
||||
if request["type"] == "matplotlib":
|
||||
return process_matplotlib_response(request, ctx.matplotlib_dump_dir)
|
||||
else:
|
||||
raise Exception(f'Unrecognised network request type: {request["type"]}')
|
||||
|
||||
|
||||
def do_subprocess(*, cmd: list, env: dict, ctx: CodeExecutionContext):
|
||||
# Create Pipes to be used for any external tool/network requests.
|
||||
req_r, req_w = multiprocessing.Pipe(duplex=False)
|
||||
resp_r, resp_w = multiprocessing.Pipe(duplex=False)
|
||||
|
||||
cmd += [str(req_w.fileno()), str(resp_r.fileno())]
|
||||
proc = subprocess.Popen(
|
||||
cmd,
|
||||
pass_fds=(req_w.fileno(), resp_r.fileno()),
|
||||
text=True,
|
||||
stdout=subprocess.PIPE,
|
||||
stderr=subprocess.PIPE,
|
||||
close_fds=True,
|
||||
env=env,
|
||||
)
|
||||
|
||||
# Close unnecessary fds.
|
||||
req_w.close()
|
||||
resp_r.close()
|
||||
|
||||
pipe_close = False
|
||||
done_read = False
|
||||
start = time.monotonic()
|
||||
while proc.poll() is None and not pipe_close:
|
||||
if req_r.poll(0.1):
|
||||
# NB: Python pipe semantics for poll and recv mean that
|
||||
# poll() returns True is a pipe is closed.
|
||||
# CF old school PEP from '09
|
||||
# https://bugs.python.org/issue5573
|
||||
try:
|
||||
request = json.loads(req_r.recv_bytes().decode("utf-8"))
|
||||
response = execute_subprocess_request(request, ctx)
|
||||
|
||||
resp_w.send_bytes(json.dumps(response).encode("utf-8"))
|
||||
except EOFError:
|
||||
# The request pipe is closed - set a marker to exit
|
||||
# after the next attempt at reading stdout/stderr.
|
||||
pipe_close = True
|
||||
|
||||
try:
|
||||
# If lots has been printed, pipe might be full but
|
||||
# proc cannot exit until all the stdout/stderr
|
||||
# been written/read.
|
||||
stdout, stderr = proc.communicate(timeout=0.3)
|
||||
done_read = True
|
||||
except subprocess.TimeoutExpired:
|
||||
# The program has not terminated. Ignore it, there
|
||||
# may be more network/tool requests.
|
||||
continue
|
||||
if time.monotonic() - start > CODE_EXEC_TIMEOUT:
|
||||
proc.terminate()
|
||||
raise subprocess.TimeoutExpired(cmd, CODE_EXEC_TIMEOUT)
|
||||
|
||||
if not done_read:
|
||||
# Solve race condition where process terminates before
|
||||
# we hit the while loop.
|
||||
stdout, stderr = proc.communicate(timeout=0.3)
|
||||
|
||||
resp_w.close()
|
||||
req_r.close()
|
||||
return stdout, stderr, proc.returncode
|
||||
|
|
@ -0,0 +1,87 @@
|
|||
# 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.
|
||||
|
||||
"""
|
||||
A custom Matplotlib backend that overrides the show method to return image bytes.
|
||||
"""
|
||||
|
||||
import base64
|
||||
import io
|
||||
import json as _json
|
||||
|
||||
import matplotlib
|
||||
from matplotlib.backend_bases import FigureManagerBase
|
||||
|
||||
# Import necessary components from Matplotlib
|
||||
from matplotlib.backends.backend_agg import FigureCanvasAgg
|
||||
|
||||
|
||||
class CustomFigureCanvas(FigureCanvasAgg):
|
||||
def show(self):
|
||||
# Save the figure to a BytesIO object
|
||||
buf = io.BytesIO()
|
||||
self.print_png(buf)
|
||||
image_bytes = buf.getvalue()
|
||||
buf.close()
|
||||
return image_bytes
|
||||
|
||||
|
||||
class CustomFigureManager(FigureManagerBase):
|
||||
def __init__(self, canvas, num):
|
||||
super().__init__(canvas, num)
|
||||
|
||||
|
||||
# Mimic module initialization that integrates with the Matplotlib backend system
|
||||
def _create_figure_manager(num, *args, **kwargs):
|
||||
"""
|
||||
Create a custom figure manager instance.
|
||||
"""
|
||||
FigureClass = kwargs.pop("FigureClass", None) # noqa: N806
|
||||
if FigureClass is None:
|
||||
from matplotlib.figure import Figure
|
||||
|
||||
FigureClass = Figure # noqa: N806
|
||||
fig = FigureClass(*args, **kwargs)
|
||||
canvas = CustomFigureCanvas(fig)
|
||||
manager = CustomFigureManager(canvas, num)
|
||||
return manager
|
||||
|
||||
|
||||
def show():
|
||||
"""
|
||||
Handle all figures and potentially return their images as bytes.
|
||||
|
||||
This function iterates over all figures registered with the custom backend,
|
||||
renders them as images in bytes format, and could return a list of bytes objects,
|
||||
one for each figure, or handle them as needed.
|
||||
"""
|
||||
image_data = []
|
||||
for manager in matplotlib._pylab_helpers.Gcf.get_all_fig_managers():
|
||||
# Get the figure from the manager
|
||||
fig = manager.canvas.figure
|
||||
buf = io.BytesIO() # Create a buffer for the figure
|
||||
fig.savefig(buf, format="png") # Save the figure to the buffer in PNG format
|
||||
buf.seek(0) # Go to the beginning of the buffer
|
||||
image_bytes = buf.getvalue() # Retrieve bytes value
|
||||
image_base64 = base64.b64encode(image_bytes).decode("utf-8")
|
||||
image_data.append({"image_base64": image_base64})
|
||||
buf.close()
|
||||
|
||||
req_con, resp_con = _open_connections()
|
||||
|
||||
_json_dump = _json.dumps(
|
||||
{
|
||||
"type": "matplotlib",
|
||||
"image_data": image_data,
|
||||
}
|
||||
)
|
||||
req_con.send_bytes(_json_dump.encode("utf-8"))
|
||||
resp = _json.loads(resp_con.recv_bytes().decode("utf-8"))
|
||||
print(resp)
|
||||
|
||||
|
||||
FigureCanvas = CustomFigureCanvas
|
||||
FigureManager = CustomFigureManager
|
||||
|
|
@ -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.
|
||||
|
||||
import os
|
||||
|
||||
DIR = os.path.dirname(os.path.realpath(__file__))
|
||||
CODE_ENV_PREFIX_FILE = os.path.join(DIR, "code_env_prefix.py")
|
||||
CODE_ENV_PREFIX = None
|
||||
|
||||
|
||||
def get_code_env_prefix() -> str:
|
||||
global CODE_ENV_PREFIX
|
||||
|
||||
if CODE_ENV_PREFIX is None:
|
||||
with open(CODE_ENV_PREFIX_FILE, "r") as f:
|
||||
CODE_ENV_PREFIX = f.read()
|
||||
|
||||
return CODE_ENV_PREFIX
|
||||
|
|
@ -0,0 +1,58 @@
|
|||
# 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 List
|
||||
|
||||
from llama_stack.apis.inference import Message
|
||||
from llama_stack.apis.safety import Safety, ShieldDefinition
|
||||
|
||||
from llama_stack.providers.impls.meta_reference.agents.safety import ShieldRunnerMixin
|
||||
|
||||
from .builtin import BaseTool
|
||||
|
||||
|
||||
class SafeTool(BaseTool, ShieldRunnerMixin):
|
||||
"""A tool that makes other tools safety enabled"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
tool: BaseTool,
|
||||
safety_api: Safety,
|
||||
input_shields: List[ShieldDefinition] = None,
|
||||
output_shields: List[ShieldDefinition] = None,
|
||||
):
|
||||
self._tool = tool
|
||||
ShieldRunnerMixin.__init__(
|
||||
self, safety_api, input_shields=input_shields, output_shields=output_shields
|
||||
)
|
||||
|
||||
def get_name(self) -> str:
|
||||
# return the name of the wrapped tool
|
||||
return self._tool.get_name()
|
||||
|
||||
async def run(self, messages: List[Message]) -> List[Message]:
|
||||
if self.input_shields:
|
||||
await self.run_shields(messages, self.input_shields)
|
||||
# run the underlying tool
|
||||
res = await self._tool.run(messages)
|
||||
if self.output_shields:
|
||||
await self.run_shields(messages, self.output_shields)
|
||||
|
||||
return res
|
||||
|
||||
|
||||
def with_safety(
|
||||
tool: BaseTool,
|
||||
safety_api: Safety,
|
||||
input_shields: List[ShieldDefinition] = None,
|
||||
output_shields: List[ShieldDefinition] = None,
|
||||
) -> SafeTool:
|
||||
return SafeTool(
|
||||
tool,
|
||||
safety_api,
|
||||
input_shields=input_shields,
|
||||
output_shields=output_shields,
|
||||
)
|
||||
|
|
@ -1,184 +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 Optional
|
||||
|
||||
from llama_models.llama3.api.datatypes import * # noqa: F403
|
||||
from llama_models.llama3.api.tool_utils import ToolUtils
|
||||
|
||||
from termcolor import cprint
|
||||
|
||||
from llama_stack.apis.agents import AgentTurnResponseEventType, StepType
|
||||
|
||||
|
||||
class LogEvent:
|
||||
def __init__(
|
||||
self,
|
||||
role: Optional[str] = None,
|
||||
content: str = "",
|
||||
end: str = "\n",
|
||||
color="white",
|
||||
):
|
||||
self.role = role
|
||||
self.content = content
|
||||
self.color = color
|
||||
self.end = "\n" if end is None else end
|
||||
|
||||
def __str__(self):
|
||||
if self.role is not None:
|
||||
return f"{self.role}> {self.content}"
|
||||
else:
|
||||
return f"{self.content}"
|
||||
|
||||
def print(self, flush=True):
|
||||
cprint(f"{str(self)}", color=self.color, end=self.end, flush=flush)
|
||||
|
||||
|
||||
EventType = AgentTurnResponseEventType
|
||||
|
||||
|
||||
class EventLogger:
|
||||
async def log(
|
||||
self,
|
||||
event_generator,
|
||||
stream=True,
|
||||
tool_prompt_format: ToolPromptFormat = ToolPromptFormat.json,
|
||||
):
|
||||
previous_event_type = None
|
||||
previous_step_type = None
|
||||
|
||||
async for chunk in event_generator:
|
||||
if not hasattr(chunk, "event"):
|
||||
# Need to check for custom tool first
|
||||
# since it does not produce event but instead
|
||||
# a Message
|
||||
if isinstance(chunk, ToolResponseMessage):
|
||||
yield chunk, LogEvent(
|
||||
role="CustomTool", content=chunk.content, color="grey"
|
||||
)
|
||||
continue
|
||||
|
||||
event = chunk.event
|
||||
event_type = event.payload.event_type
|
||||
if event_type in {
|
||||
EventType.turn_start.value,
|
||||
EventType.turn_complete.value,
|
||||
}:
|
||||
# Currently not logging any turn realted info
|
||||
yield event, None
|
||||
continue
|
||||
|
||||
step_type = event.payload.step_type
|
||||
# handle safety
|
||||
if (
|
||||
step_type == StepType.shield_call
|
||||
and event_type == EventType.step_complete.value
|
||||
):
|
||||
response = event.payload.step_details.response
|
||||
if not response.is_violation:
|
||||
yield event, LogEvent(
|
||||
role=step_type, content="No Violation", color="magenta"
|
||||
)
|
||||
else:
|
||||
yield event, LogEvent(
|
||||
role=step_type,
|
||||
content=f"{response.violation_type} {response.violation_return_message}",
|
||||
color="red",
|
||||
)
|
||||
|
||||
# handle inference
|
||||
if step_type == StepType.inference:
|
||||
if stream:
|
||||
if event_type == EventType.step_start.value:
|
||||
# TODO: Currently this event is never received
|
||||
yield event, LogEvent(
|
||||
role=step_type, content="", end="", color="yellow"
|
||||
)
|
||||
elif event_type == EventType.step_progress.value:
|
||||
# HACK: if previous was not step/event was not inference's step_progress
|
||||
# this is the first time we are getting model inference response
|
||||
# aka equivalent to step_start for inference. Hence,
|
||||
# start with "Model>".
|
||||
if (
|
||||
previous_event_type != EventType.step_progress.value
|
||||
and previous_step_type != StepType.inference
|
||||
):
|
||||
yield event, LogEvent(
|
||||
role=step_type, content="", end="", color="yellow"
|
||||
)
|
||||
|
||||
if event.payload.tool_call_delta:
|
||||
if isinstance(event.payload.tool_call_delta.content, str):
|
||||
yield event, LogEvent(
|
||||
role=None,
|
||||
content=event.payload.tool_call_delta.content,
|
||||
end="",
|
||||
color="cyan",
|
||||
)
|
||||
else:
|
||||
yield event, LogEvent(
|
||||
role=None,
|
||||
content=event.payload.model_response_text_delta,
|
||||
end="",
|
||||
color="yellow",
|
||||
)
|
||||
else:
|
||||
# step_complete
|
||||
yield event, LogEvent(role=None, content="")
|
||||
|
||||
else:
|
||||
# Not streaming
|
||||
if event_type == EventType.step_complete.value:
|
||||
response = event.payload.step_details.model_response
|
||||
if response.tool_calls:
|
||||
content = ToolUtils.encode_tool_call(
|
||||
response.tool_calls[0], tool_prompt_format
|
||||
)
|
||||
else:
|
||||
content = response.content
|
||||
yield event, LogEvent(
|
||||
role=step_type,
|
||||
content=content,
|
||||
color="yellow",
|
||||
)
|
||||
|
||||
# handle tool_execution
|
||||
if (
|
||||
step_type == StepType.tool_execution
|
||||
and
|
||||
# Only print tool calls and responses at the step_complete event
|
||||
event_type == EventType.step_complete.value
|
||||
):
|
||||
details = event.payload.step_details
|
||||
for t in details.tool_calls:
|
||||
yield event, LogEvent(
|
||||
role=step_type,
|
||||
content=f"Tool:{t.tool_name} Args:{t.arguments}",
|
||||
color="green",
|
||||
)
|
||||
for r in details.tool_responses:
|
||||
yield event, LogEvent(
|
||||
role=step_type,
|
||||
content=f"Tool:{r.tool_name} Response:{r.content}",
|
||||
color="green",
|
||||
)
|
||||
|
||||
if (
|
||||
step_type == StepType.memory_retrieval
|
||||
and event_type == EventType.step_complete.value
|
||||
):
|
||||
details = event.payload.step_details
|
||||
content = interleaved_text_media_as_str(details.inserted_context)
|
||||
content = content[:200] + "..." if len(content) > 200 else content
|
||||
|
||||
yield event, LogEvent(
|
||||
role=step_type,
|
||||
content=f"Retrieved context from banks: {details.memory_bank_ids}.\n====\n{content}\n>",
|
||||
color="cyan",
|
||||
)
|
||||
|
||||
preivous_event_type = event_type
|
||||
previous_step_type = step_type
|
||||
|
|
@ -1,94 +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 AsyncGenerator, List
|
||||
|
||||
from llama_models.llama3.api.datatypes import * # noqa: F403
|
||||
from llama_stack.apis.agents import * # noqa: F403
|
||||
from llama_stack.apis.memory import * # noqa: F403
|
||||
from llama_stack.apis.safety import * # noqa: F403
|
||||
|
||||
from llama_stack.apis.agents import AgentTurnResponseEventType as EventType
|
||||
from llama_stack.tools.custom.datatypes import CustomTool
|
||||
|
||||
|
||||
class AgentWithCustomToolExecutor:
|
||||
def __init__(
|
||||
self,
|
||||
api: Agents,
|
||||
agent_id: str,
|
||||
session_id: str,
|
||||
agent_config: AgentConfig,
|
||||
custom_tools: List[CustomTool],
|
||||
):
|
||||
self.api = api
|
||||
self.agent_id = agent_id
|
||||
self.session_id = session_id
|
||||
self.agent_config = agent_config
|
||||
self.custom_tools = custom_tools
|
||||
|
||||
async def execute_turn(
|
||||
self,
|
||||
messages: List[Message],
|
||||
attachments: Optional[List[Attachment]] = None,
|
||||
max_iters: int = 5,
|
||||
stream: bool = True,
|
||||
) -> AsyncGenerator:
|
||||
tools_dict = {t.get_name(): t for t in self.custom_tools}
|
||||
|
||||
current_messages = messages.copy()
|
||||
n_iter = 0
|
||||
while n_iter < max_iters:
|
||||
n_iter += 1
|
||||
|
||||
request = AgentTurnCreateRequest(
|
||||
agent_id=self.agent_id,
|
||||
session_id=self.session_id,
|
||||
messages=current_messages,
|
||||
attachments=attachments,
|
||||
stream=stream,
|
||||
)
|
||||
|
||||
turn = None
|
||||
async for chunk in self.api.create_agent_turn(request):
|
||||
if chunk.event.payload.event_type != EventType.turn_complete.value:
|
||||
yield chunk
|
||||
else:
|
||||
turn = chunk.event.payload.turn
|
||||
|
||||
message = turn.output_message
|
||||
if len(message.tool_calls) == 0:
|
||||
yield chunk
|
||||
return
|
||||
|
||||
if message.stop_reason == StopReason.out_of_tokens:
|
||||
yield chunk
|
||||
return
|
||||
|
||||
tool_call = message.tool_calls[0]
|
||||
if tool_call.tool_name not in tools_dict:
|
||||
m = ToolResponseMessage(
|
||||
call_id=tool_call.call_id,
|
||||
tool_name=tool_call.tool_name,
|
||||
content=f"Unknown tool `{tool_call.tool_name}` was called. Try again with something else",
|
||||
)
|
||||
next_message = m
|
||||
else:
|
||||
tool = tools_dict[tool_call.tool_name]
|
||||
result_messages = await execute_custom_tool(tool, message)
|
||||
next_message = result_messages[0]
|
||||
|
||||
yield next_message
|
||||
current_messages = [next_message]
|
||||
|
||||
|
||||
async def execute_custom_tool(tool: CustomTool, message: Message) -> List[Message]:
|
||||
result_messages = await tool.run([message])
|
||||
assert (
|
||||
len(result_messages) == 1
|
||||
), f"Expected single message, got {len(result_messages)}"
|
||||
|
||||
return result_messages
|
||||
Loading…
Add table
Add a link
Reference in a new issue