ChatPromptTemplate constructs a structured chat prompt by combining a user's question with retrieved document content and metadata before a chat model generates an answer. section 14.5.4 section 14.5.2
Use it as the prompt-building stage in a retrieval-augmented generation pipeline:
- Receive the question and context produced by the retrieval chain. section 14.5.4
- Combine document content with metadata such as source, page number, and embeddings. section 14.5.4
- Format the combined context and question into an organized prompt for ChatOpenAI. section 14.5.4 section 14.5.2
Index question in a RAG chain
For the question “What indexes does MongoDB support?”, the template combines retrieved document content and metadata into a message instructing the model to answer only from the supplied context. ChatOpenAI then generates the response, and StrOutputParser returns the answer as text. section 14.5.4
import os
from langchain_openai import ChatOpenAI
from langchain_community.document_loaders import PyPDFLoader
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import MongoDBAtlasVectorSearch
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.pydantic_v1 import BaseModel
from langchain_core.runnables import (
RunnableLambda,
RunnableParallel,
RunnablePassthrough,
)
from langchain_text_splitters import RecursiveCharacterTextSplitter
from pymongo import MongoClient
# Set DB
if os.environ.get("MONGO_URI", None) is None:
raise Exception("Missing `MONGO_URI` environment variable.")
MONGO_URI = os.environ["MONGO_URI"]
DB_NAME = "langchain"
COLLECTION_NAME = "mongodb"
ATLAS_VECTOR_SEARCH_INDEX_NAME = "default"
client = MongoClient(MONGO_URI)
db = client[DB_NAME]
MONGODB_COLLECTION = db[COLLECTION_NAME]
# Read from MongoDB Atlas Vector Search
vectorstore = MongoDBAtlasVectorSearch.from_connection_string(
MONGO_URI,
DB_NAME + "." + COLLECTION_NAME,
OpenAIEmbeddings(disallowed_special=()),
index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME,
)
retriever = vectorstore.as_retriever(search_kwargs={"k": 1})
# RAG prompt
template = """Answer the question based only on the following context:
{context}
Question: {question}
"""
prompt = ChatPromptTemplate.from_template(template)
# RAG
model = ChatOpenAI()
chain = (
RunnableParallel({"context": retriever,
➥"question": RunnablePassthrough()})
| prompt
| model
| StrOutputParser()
)
# Add typing for input
class Question(BaseModel):
__root__: str
chain = chain.with_types(input_type=Question)
def _ingest(url: str) -> dict:
loader = PyPDFLoader(url)
data = loader.load()
# Split docs
text_splitter = RecursiveCharacterTextSplitter
➥(chunk_size=500, chunk_overlap=0)
docs = text_splitter.split_documents(data)
# Insert the documents in MongoDB Atlas Vector Search
_ = MongoDBAtlasVectorSearch.from_documents(
documents=docs,
embedding=OpenAIEmbeddings(disallowed_special=()),
collection=MONGODB_COLLECTION,
index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME,
)
return {}
ingest = RunnableLambda(_ingest)
Inspect the complete Python chain showing ChatPromptTemplate alongside the retriever, ChatOpenAI, runnable components, and output parser. section 14.5.2
