langchain/templates/basic-critique-revise/basic_critique_revise/chain.py
2024-01-03 13:28:05 -08:00

144 lines
3.7 KiB
Python

import json
from datetime import datetime
from enum import Enum
from operator import itemgetter
from typing import Any, Dict, Sequence
from langchain.chains.openai_functions import convert_to_openai_function
from langchain_community.chat_models import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field, ValidationError, conint
from langchain_core.runnables import (
Runnable,
RunnableBranch,
RunnableLambda,
RunnablePassthrough,
)
class TaskType(str, Enum):
call = "Call"
message = "Message"
todo = "Todo"
in_person_meeting = "In-Person Meeting"
email = "Email"
mail = "Mail"
text = "Text"
open_house = "Open House"
class Task(BaseModel):
title: str = Field(..., description="The title of the tasks, reminders and alerts")
due_date: datetime = Field(
..., description="Due date. Must be a valid ISO date string with timezone"
)
task_type: TaskType = Field(None, description="The type of task")
class Tasks(BaseModel):
"""JSON definition for creating tasks, reminders and alerts"""
tasks: Sequence[Task]
template = """Respond to the following user query to the best of your ability:
{query}"""
generate_prompt = ChatPromptTemplate.from_template(template)
function_args = {"functions": [convert_to_openai_function(Tasks)]}
task_function_call_model = ChatOpenAI(model="gpt-3.5-turbo").bind(**function_args)
output_parser = RunnableLambda(
lambda x: json.loads(
x.additional_kwargs.get("function_call", {}).get("arguments", '""')
)
)
revise_template = """
Based on the provided context, fix the incorrect result of the original prompt
and the provided errors. Only respond with an answer that satisfies the
constraints laid out in the original prompt and fixes the Pydantic errors.
Hint: Datetime fields must be valid ISO date strings.
<context>
<original_prompt>
{original_prompt}
</original_prompt>
<incorrect_result>
{completion}
</incorrect_result>
<errors>
{error}
</errors>
</context>"""
revise_prompt = ChatPromptTemplate.from_template(revise_template)
revise_chain = revise_prompt | task_function_call_model | output_parser
def output_validator(output):
try:
Tasks.validate(output["completion"])
except ValidationError as e:
return str(e)
return None
class IntermediateType(BaseModel):
error: str
completion: Dict
original_prompt: str
max_revisions: int
validation_step = RunnablePassthrough().assign(error=RunnableLambda(output_validator))
def revise_loop(input: IntermediateType) -> IntermediateType:
revise_step = RunnablePassthrough().assign(completion=revise_chain)
else_step: Runnable[IntermediateType, IntermediateType] = RunnableBranch(
(lambda x: x["error"] is None, RunnablePassthrough()),
revise_step | validation_step,
).with_types(input_type=IntermediateType)
for _ in range(max(0, input["max_revisions"] - 1)):
else_step = RunnableBranch(
(lambda x: x["error"] is None, RunnablePassthrough()),
revise_step | validation_step | else_step,
)
return else_step
revise_lambda = RunnableLambda(revise_loop)
class InputType(BaseModel):
query: str
max_revisions: conint(ge=1, le=10) = 5
chain: Runnable[Any, Any] = (
{
"original_prompt": generate_prompt,
"max_revisions": itemgetter("max_revisions"),
}
| RunnablePassthrough().assign(
completion=(
RunnableLambda(itemgetter("original_prompt"))
| task_function_call_model
| output_parser
)
)
| validation_step
| revise_lambda
| RunnableLambda(itemgetter("completion"))
).with_types(input_type=InputType)