initial commit

This commit is contained in:
2026-05-11 12:36:20 +05:30
commit 384cbe8019
15377 changed files with 2360544 additions and 0 deletions

View File

@@ -0,0 +1,7 @@
"""Chains for evaluating ReAct style agents."""
from langchain_classic.evaluation.agents.trajectory_eval_chain import (
TrajectoryEvalChain,
)
__all__ = ["TrajectoryEvalChain"]

View File

@@ -0,0 +1,418 @@
"""A chain for evaluating ReAct style agents.
This chain is used to evaluate ReAct style agents by reasoning about
the sequence of actions taken and their outcomes. It uses a language model
chain (LLMChain) to generate the reasoning and scores.
"""
import re
from collections.abc import Sequence
from typing import (
Any,
TypedDict,
cast,
)
from langchain_core.agents import AgentAction
from langchain_core.callbacks import Callbacks
from langchain_core.callbacks.manager import (
AsyncCallbackManagerForChainRun,
CallbackManagerForChainRun,
)
from langchain_core.exceptions import OutputParserException
from langchain_core.language_models import BaseLanguageModel
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.output_parsers import BaseOutputParser
from langchain_core.tools import BaseTool
from pydantic import ConfigDict, Field
from typing_extensions import override
from langchain_classic.chains.llm import LLMChain
from langchain_classic.evaluation.agents.trajectory_eval_prompt import (
EVAL_CHAT_PROMPT,
TOOL_FREE_EVAL_CHAT_PROMPT,
)
from langchain_classic.evaluation.schema import AgentTrajectoryEvaluator, LLMEvalChain
_MAX_SCORE = 5
class TrajectoryEval(TypedDict):
"""A named tuple containing the score and reasoning for a trajectory."""
score: float
"""The score for the trajectory, normalized from 0 to 1."""
reasoning: str
"""The reasoning for the score."""
class TrajectoryOutputParser(BaseOutputParser):
"""Trajectory output parser."""
@property
def _type(self) -> str:
return "agent_trajectory"
def parse(self, text: str) -> TrajectoryEval:
"""Parse the output text and extract the score and reasoning.
Args:
text: The output text to parse.
Returns:
A named tuple containing the normalized score and reasoning.
Raises:
If the score is not found in the output text or if the LLM's score is not a
digit in the range 1-5.
"""
if "Score:" not in text:
msg = f"Could not find score in model eval output: {text}"
raise OutputParserException(msg)
reasoning, score_str = text.split("Score: ", maxsplit=1)
reasoning, score_str = reasoning.strip(), score_str.strip()
# Use regex to extract the score.
# This will get the number in the string, even if it is a float or more than 10.
# E.g. "Score: 1" will return 1, "Score: 3.5" will return 3.5, and
# "Score: 10" will return 10.
# The score should be an integer digit in the range 1-5.
_score = re.search(r"(\d+(\.\d+)?)", score_str)
# If the score is not found or is a float, raise an exception.
if _score is None or "." in _score.group(1):
msg = f"Score is not an integer digit in the range 1-5: {text}"
raise OutputParserException(msg)
score = int(_score.group(1))
# If the score is not in the range 1-5, raise an exception.
if not 1 <= score <= _MAX_SCORE:
msg = f"Score is not a digit in the range 1-5: {text}"
raise OutputParserException(msg)
normalized_score = (score - 1) / (_MAX_SCORE - 1)
return TrajectoryEval(score=normalized_score, reasoning=reasoning)
class TrajectoryEvalChain(AgentTrajectoryEvaluator, LLMEvalChain):
"""A chain for evaluating ReAct style agents.
This chain is used to evaluate ReAct style agents by reasoning about
the sequence of actions taken and their outcomes.
Based on the paper "ReAct: Synergizing Reasoning and Acting in Language Models"
(https://arxiv.org/abs/2210.03629)
Example:
```python
from langchain_classic.agents import AgentType, initialize_agent
from langchain_openai import ChatOpenAI
from langchain_classic.evaluation import TrajectoryEvalChain
from langchain_classic.tools import tool
@tool
def geography_answers(country: str, question: str) -> str:
\"\"\"Very helpful answers to geography questions.\"\"\"
return f"{country}? IDK - We may never know {question}."
model = ChatOpenAI(model="gpt-3.5-turbo", temperature=0)
agent = initialize_agent(
tools=[geography_answers],
llm=model,
agent=AgentType.OPENAI_FUNCTIONS,
return_intermediate_steps=True,
)
question = "How many dwell in the largest minor region in Argentina?"
response = agent(question)
eval_chain = TrajectoryEvalChain.from_llm(
llm=model, agent_tools=[geography_answers], return_reasoning=True
)
result = eval_chain.evaluate_agent_trajectory(
input=question,
agent_trajectory=response["intermediate_steps"],
prediction=response["output"],
reference="Paris",
)
print(result["score"]) # noqa: T201
# 0
```
"""
agent_tools: list[BaseTool] | None = None
"""A list of tools available to the agent."""
eval_chain: LLMChain
"""The language model chain used for evaluation."""
output_parser: TrajectoryOutputParser = Field(
default_factory=TrajectoryOutputParser,
)
"""The output parser used to parse the output."""
return_reasoning: bool = False
"""DEPRECATED. Reasoning always returned."""
model_config = ConfigDict(
extra="ignore",
)
@property
def requires_reference(self) -> bool:
"""Whether this evaluator requires a reference label."""
return False
@property
def _tools_description(self) -> str:
"""Get the description of the agent tools.
Returns:
The description of the agent tools.
"""
if self.agent_tools is None:
return ""
return "\n\n".join(
[
f"""Tool {i}: {tool.name}
Description: {tool.description}"""
for i, tool in enumerate(self.agent_tools, 1)
],
)
@staticmethod
def get_agent_trajectory(
steps: str | Sequence[tuple[AgentAction, str]],
) -> str:
"""Get the agent trajectory as a formatted string.
Args:
steps: The agent trajectory.
Returns:
The formatted agent trajectory.
"""
if isinstance(steps, str):
return steps
return "\n\n".join(
[
f"""Step {i}:
Tool used: {action.tool}
Tool input: {action.tool_input}
Tool output: {output}"""
for i, (action, output) in enumerate(steps, 1)
],
)
@staticmethod
def _format_reference(reference: str | None) -> str:
"""Format the reference text.
Args:
reference: The reference text.
Returns:
The formatted reference text.
"""
if not reference:
return ""
return f"""
The following is the expected answer. Use this to measure correctness:
[GROUND_TRUTH]
{reference}
[END_GROUND_TRUTH]
"""
@classmethod
def from_llm(
cls,
llm: BaseLanguageModel,
agent_tools: Sequence[BaseTool] | None = None,
output_parser: TrajectoryOutputParser | None = None,
**kwargs: Any,
) -> "TrajectoryEvalChain":
"""Create a TrajectoryEvalChain object from a language model chain.
Args:
llm: The language model chain.
agent_tools: A list of tools available to the agent.
output_parser : The output parser used to parse the chain output into a
score.
**kwargs: Additional keyword arguments.
Returns:
The `TrajectoryEvalChain` object.
"""
if not isinstance(llm, BaseChatModel):
msg = "Only chat models supported by the current trajectory eval"
raise NotImplementedError(msg)
prompt = EVAL_CHAT_PROMPT if agent_tools else TOOL_FREE_EVAL_CHAT_PROMPT
eval_chain = LLMChain(llm=llm, prompt=prompt)
return cls(
agent_tools=agent_tools,
eval_chain=eval_chain,
output_parser=output_parser or TrajectoryOutputParser(),
**kwargs,
)
@property
def input_keys(self) -> list[str]:
"""Get the input keys for the chain.
Returns:
The input keys.
"""
return ["question", "agent_trajectory", "answer", "reference"]
@property
def output_keys(self) -> list[str]:
"""Get the output keys for the chain.
Returns:
The output keys.
"""
return ["score", "reasoning"]
def prep_inputs(self, inputs: dict[str, Any] | Any) -> dict[str, str]:
"""Validate and prep inputs."""
inputs["reference"] = self._format_reference(inputs.get("reference"))
return super().prep_inputs(inputs)
def _call(
self,
inputs: dict[str, str],
run_manager: CallbackManagerForChainRun | None = None,
) -> dict[str, Any]:
"""Run the chain and generate the output.
Args:
inputs: The input values for the chain.
run_manager: The callback manager for the chain run.
Returns:
The output values of the chain.
"""
chain_input = {**inputs}
if self.agent_tools:
chain_input["tool_descriptions"] = self._tools_description
_run_manager = run_manager or CallbackManagerForChainRun.get_noop_manager()
raw_output = self.eval_chain.run(
chain_input,
callbacks=_run_manager.get_child(),
)
return cast("dict", self.output_parser.parse(raw_output))
async def _acall(
self,
inputs: dict[str, str],
run_manager: AsyncCallbackManagerForChainRun | None = None,
) -> dict[str, Any]:
"""Run the chain and generate the output.
Args:
inputs: The input values for the chain.
run_manager: The callback manager for the chain run.
Returns:
The output values of the chain.
"""
chain_input = {**inputs}
if self.agent_tools:
chain_input["tool_descriptions"] = self._tools_description
_run_manager = run_manager or AsyncCallbackManagerForChainRun.get_noop_manager()
raw_output = await self.eval_chain.arun(
chain_input,
callbacks=_run_manager.get_child(),
)
return cast("dict", self.output_parser.parse(raw_output))
@override
def _evaluate_agent_trajectory(
self,
*,
prediction: str,
input: str,
agent_trajectory: Sequence[tuple[AgentAction, str]],
reference: str | None = None,
callbacks: Callbacks = None,
tags: list[str] | None = None,
metadata: dict[str, Any] | None = None,
include_run_info: bool = False,
**kwargs: Any,
) -> dict:
"""Evaluate a trajectory.
Args:
prediction: The final predicted response.
input: The input to the agent.
agent_trajectory: The intermediate steps forming the agent trajectory.
reference: The reference answer.
callbacks: Callbacks to use for this chain run.
tags: The tags to apply.
metadata: The metadata to use.
include_run_info: Whether to include run info in the output.
**kwargs: Additional keyword arguments.
Returns:
The evaluation result, which includes the score and optionally
the reasoning for reaching that.
"""
inputs = {
"question": input,
"agent_trajectory": self.get_agent_trajectory(agent_trajectory),
"answer": prediction,
"reference": reference,
}
return self.__call__(
inputs=inputs,
callbacks=callbacks,
tags=tags,
metadata=metadata,
include_run_info=include_run_info,
return_only_outputs=True,
)
@override
async def _aevaluate_agent_trajectory(
self,
*,
prediction: str,
input: str,
agent_trajectory: Sequence[tuple[AgentAction, str]],
reference: str | None = None,
callbacks: Callbacks = None,
tags: list[str] | None = None,
metadata: dict[str, Any] | None = None,
include_run_info: bool = False,
**kwargs: Any,
) -> dict:
"""Asynchronously evaluate a trajectory.
Args:
prediction: The final predicted response.
input: The input to the agent.
agent_trajectory: The intermediate steps forming the agent trajectory.
reference: The reference answer.
callbacks: Callbacks to use for this chain run.
tags: The tags to apply.
metadata: The metadata to use.
include_run_info: Whether to include run info in the output.
**kwargs: Additional keyword arguments.
Returns:
The evaluation result, which includes the score and optionally
the reasoning for reaching that.
"""
inputs = {
"question": input,
"agent_trajectory": self.get_agent_trajectory(agent_trajectory),
"answer": prediction,
"reference": reference,
}
return await self.acall(
inputs=inputs,
callbacks=callbacks,
tags=tags,
metadata=metadata,
include_run_info=include_run_info,
return_only_outputs=True,
)

View File

@@ -0,0 +1,146 @@
"""Prompt for trajectory evaluation chain."""
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
from langchain_core.prompts.chat import (
ChatPromptTemplate,
HumanMessagePromptTemplate,
)
EVAL_TEMPLATE = """An AI language model has been given access to the following set of tools to help answer a user's question.
The tools given to the AI model are:
[TOOL_DESCRIPTIONS]
{tool_descriptions}
[END_TOOL_DESCRIPTIONS]
The question the human asked the AI model was:
[QUESTION]
{question}
[END_QUESTION]{reference}
The AI language model decided to use the following set of tools to answer the question:
[AGENT_TRAJECTORY]
{agent_trajectory}
[END_AGENT_TRAJECTORY]
The AI language model's final answer to the question was:
[RESPONSE]
{answer}
[END_RESPONSE]
Let's to do a detailed evaluation of the AI language model's answer step by step.
We consider the following criteria before giving a score from 1 to 5:
i. Is the final answer helpful?
ii. Does the AI language use a logical sequence of tools to answer the question?
iii. Does the AI language model use the tools in a helpful way?
iv. Does the AI language model use too many steps to answer the question?
v. Are the appropriate tools used to answer the question?""" # noqa: E501
EXAMPLE_INPUT = """An AI language model has been given access to the following set of tools to help answer a user's question.
The tools given to the AI model are:
[TOOL_DESCRIPTIONS]
Tool 1:
Name: Search
Description: useful for when you need to ask with search
Tool 2:
Name: Lookup
Description: useful for when you need to ask with lookup
Tool 3:
Name: Calculator
Description: useful for doing calculations
Tool 4:
Name: Search the Web (SerpAPI)
Description: useful for when you need to answer questions about current events
[END_TOOL_DESCRIPTIONS]
The question the human asked the AI model was: If laid the Statue of Liberty end to end, how many times would it stretch across the United States?
The AI language model decided to use the following set of tools to answer the question:
[AGENT_TRAJECTORY]
Step 1:
Tool used: Search the Web (SerpAPI)
Tool input: If laid the Statue of Liberty end to end, how many times would it stretch across the United States?
Tool output: The Statue of Liberty was given to the United States by France, as a symbol of the two countries' friendship. It was erected atop an American-designed ...
[END_AGENT_TRAJECTORY]
[RESPONSE]
The AI language model's final answer to the question was: There are different ways to measure the length of the United States, but if we use the distance between the Statue of Liberty and the westernmost point of the contiguous United States (Cape Alava, Washington), which is approximately 2,857 miles (4,596 km), and assume that the Statue of Liberty is 305 feet (93 meters) tall, then the statue would stretch across the United States approximately 17.5 times if laid end to end.
[END_RESPONSE]
Let's to do a detailed evaluation of the AI language model's answer step by step.
We consider the following criteria before giving a score from 1 to 5:
i. Is the final answer helpful?
ii. Does the AI language use a logical sequence of tools to answer the question?
iii. Does the AI language model use the tools in a helpful way?
iv. Does the AI language model use too many steps to answer the question?
v. Are the appropriate tools used to answer the question?""" # noqa: E501
EXAMPLE_OUTPUT = """First, let's evaluate the final answer. The final uses good reasoning but is wrong. 2,857 divided by 305 is not 17.5.\
The model should have used the calculator to figure this out. Second does the model use a logical sequence of tools to answer the question?\
The way model uses the search is not helpful. The model should have used the search tool to figure the width of the US or the height of the statue.\
The model didn't use the calculator tool and gave an incorrect answer. The search API should be used for current events or specific questions.\
The tools were not used in a helpful way. The model did not use too many steps to answer the question.\
The model did not use the appropriate tools to answer the question.\
Judgment: Given the good reasoning in the final answer but otherwise poor performance, we give the model a score of 2.
Score: 2""" # noqa: E501
EVAL_CHAT_PROMPT = ChatPromptTemplate.from_messages(
messages=[
SystemMessage(
content="You are a helpful assistant that evaluates language models."
),
HumanMessage(content=EXAMPLE_INPUT),
AIMessage(content=EXAMPLE_OUTPUT),
HumanMessagePromptTemplate.from_template(EVAL_TEMPLATE),
]
)
TOOL_FREE_EVAL_TEMPLATE = """An AI language model has been given access to a set of tools to help answer a user's question.
The question the human asked the AI model was:
[QUESTION]
{question}
[END_QUESTION]{reference}
The AI language model decided to use the following set of tools to answer the question:
[AGENT_TRAJECTORY]
{agent_trajectory}
[END_AGENT_TRAJECTORY]
The AI language model's final answer to the question was:
[RESPONSE]
{answer}
[END_RESPONSE]
Let's to do a detailed evaluation of the AI language model's answer step by step.
We consider the following criteria before giving a score from 1 to 5:
i. Is the final answer helpful?
ii. Does the AI language use a logical sequence of tools to answer the question?
iii. Does the AI language model use the tools in a helpful way?
iv. Does the AI language model use too many steps to answer the question?
v. Are the appropriate tools used to answer the question?""" # noqa: E501
TOOL_FREE_EVAL_CHAT_PROMPT = ChatPromptTemplate.from_messages(
messages=[
SystemMessage(
content="You are a helpful assistant that evaluates language models."
),
HumanMessage(content=EXAMPLE_INPUT),
AIMessage(content=EXAMPLE_OUTPUT),
HumanMessagePromptTemplate.from_template(TOOL_FREE_EVAL_TEMPLATE),
]
)