Skip to main content

Migrating from MapRerankDocumentsChain

MapRerankDocumentsChain implements a strategy for analyzing long texts. The strategy is as follows:

  • Split a text into smaller documents;
  • Map a process to the set of documents, where the process includes generating a score;
  • Rank the results by score and return the maximum.

A common process in this scenario is question-answering using pieces of context from a document. Forcing the model to generate score along with its answer helps to select for answers generated only by relevant context.

An LangGraph implementation allows for the incorporation of tool calling and other features for this problem. Below we will go through both MapRerankDocumentsChain and a corresponding LangGraph implementation on a simple example for illustrative purposes.

Example​

Let's go through an example where we analyze a set of documents. Let's use the following 3 documents:

from langchain_core.documents import Document

documents = [
Document(page_content="Alice has blue eyes", metadata={"title": "book_chapter_2"}),
Document(page_content="Bob has brown eyes", metadata={"title": "book_chapter_1"}),
Document(
page_content="Charlie has green eyes", metadata={"title": "book_chapter_3"}
),
]
API Reference:Document

Legacy​

Details

Below we show an implementation with MapRerankDocumentsChain. We define the prompt template for a question-answering task and instantiate a LLMChain object for this purpose. We define how documents are formatted into the prompt and ensure consistency among the keys in the various prompts.

from langchain.chains import LLMChain, MapRerankDocumentsChain
from langchain.output_parsers.regex import RegexParser
from langchain_core.prompts import PromptTemplate
from langchain_openai import OpenAI

document_variable_name = "context"
llm = OpenAI()
# The prompt here should take as an input variable the
# `document_variable_name`
# The actual prompt will need to be a lot more complex, this is just
# an example.
prompt_template = (
"What color are Bob's eyes? "
"Output both your answer and a score (1-10) of how confident "
"you are in the format: <Answer>\nScore: <Score>.\n\n"
"Provide no other commentary.\n\n"
"Context: {context}"
)
output_parser = RegexParser(
regex=r"(.*?)\nScore: (.*)",
output_keys=["answer", "score"],
)
prompt = PromptTemplate(
template=prompt_template,
input_variables=["context"],
output_parser=output_parser,
)
llm_chain = LLMChain(llm=llm, prompt=prompt)
chain = MapRerankDocumentsChain(
llm_chain=llm_chain,
document_variable_name=document_variable_name,
rank_key="score",
answer_key="answer",
)
response = chain.invoke(documents)
response["output_text"]
/langchain/libs/langchain/langchain/chains/llm.py:369: UserWarning: The apply_and_parse method is deprecated, instead pass an output parser directly to LLMChain.
warnings.warn(
'Brown'

Inspecting the LangSmith trace for the above run, we can see three LLM calls-- one for each document-- and that the scoring mechanism mitigated against hallucinations.

LangGraph​

Details

Below we show a LangGraph implementation of this process. Note that our template is simplified, as we delegate the formatting instructions to the chat model's tool-calling features via the .with_structured_output method.

Here we follow a basic map-reduce workflow to execute the LLM calls in parallel.

We will need to install langgraph:

pip install -qU langgraph
import operator
from typing import Annotated, List, TypedDict

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langgraph.constants import Send
from langgraph.graph import END, START, StateGraph


class AnswerWithScore(TypedDict):
answer: str
score: Annotated[int, ..., "Score from 1-10."]


llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

prompt_template = "What color are Bob's eyes?\n\n" "Context: {context}"
prompt = ChatPromptTemplate.from_template(prompt_template)

# The below chain formats context from a document into a prompt, then
# generates a response structured according to the AnswerWithScore schema.
map_chain = prompt | llm.with_structured_output(AnswerWithScore)

# Below we define the components that will make up the graph


# This will be the overall state of the graph.
# It will contain the input document contents, corresponding
# answers with scores, and a final answer.
class State(TypedDict):
contents: List[str]
answers_with_scores: Annotated[list, operator.add]
answer: str


# This will be the state of the node that we will "map" all
# documents to in order to generate answers with scores
class MapState(TypedDict):
content: str


# Here we define the logic to map out over the documents
# We will use this an edge in the graph
def map_analyses(state: State):
# We will return a list of `Send` objects
# Each `Send` object consists of the name of a node in the graph
# as well as the state to send to that node
return [
Send("generate_analysis", {"content": content}) for content in state["contents"]
]


# Here we generate an answer with score, given a document
async def generate_analysis(state: MapState):
response = await map_chain.ainvoke(state["content"])
return {"answers_with_scores": [response]}


# Here we will select the top answer
def pick_top_ranked(state: State):
ranked_answers = sorted(
state["answers_with_scores"], key=lambda x: -int(x["score"])
)
return {"answer": ranked_answers[0]}


# Construct the graph: here we put everything together to construct our graph
graph = StateGraph(State)
graph.add_node("generate_analysis", generate_analysis)
graph.add_node("pick_top_ranked", pick_top_ranked)
graph.add_conditional_edges(START, map_analyses, ["generate_analysis"])
graph.add_edge("generate_analysis", "pick_top_ranked")
graph.add_edge("pick_top_ranked", END)
app = graph.compile()
from IPython.display import Image

Image(app.get_graph().draw_mermaid_png())

result = await app.ainvoke({"contents": [doc.page_content for doc in documents]})
result["answer"]
{'answer': 'Bob has brown eyes.', 'score': 10}

Inspecting the LangSmith trace for the above run, we can see three LLM calls as before. Using the model's tool-calling features have also enabled us to remove the parsing step.

Next steps​

See these how-to guides for more on question-answering tasks with RAG.

Check out the LangGraph documentation for detail on building with LangGraph, including this guide on the details of map-reduce in LangGraph.


Was this page helpful?