Principle:Langchain ai Langchain Response Parsing

Knowledge Sources	LangChain
Domains	NLP, Data_Transformation
Last Updated	2026-02-11 00:00 GMT

Overview

A post-processing step that transforms raw provider API responses into LangChain's standardized message and metadata format.

Description

After the provider API returns a raw response (e.g., an OpenAI completion object), response parsing extracts the relevant information and structures it into LangChain's unified format:

• Content extraction: Text content, refusal messages, or audio data
• Tool call parsing: Structured tool call objects from the model's function calling output
• Usage metadata: Token counts (input, output, total) with breakdown details (cached tokens, reasoning tokens)
• Response metadata: Model name, system fingerprint, finish reason

This normalization ensures that downstream code (output parsers, chains, agents) can work with any provider's output through a consistent interface.

Usage

This principle is applied automatically after every API call. It is essential for multi-provider applications where the same downstream logic must handle responses from different LLM providers.

Theoretical Basis

Response parsing follows the Normalizer Pattern:

# Abstract algorithm (not real code)
def parse_response(raw_response):
    for choice in raw_response.choices:
        message = AIMessage(
            content=choice.message.content,
            tool_calls=parse_tool_calls(choice.message.tool_calls),
            usage_metadata=parse_usage(raw_response.usage),
            response_metadata=extract_metadata(raw_response),
        )
        yield ChatGeneration(message=message)
    return ChatResult(generations=[...])

Related Pages

Implemented By

• Implementation:Langchain_ai_Langchain_BaseChatOpenAI_Create_Chat_Result