Skip to content

CLAUDE.md

Latest commit

 

History

History
71 lines (48 loc) · 2.63 KB

CLAUDE.md

File metadata and controls

71 lines (48 loc) · 2.63 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

CodeBase Intelligence Hub is a learning project for building a RAG-based codebase Q&A system using LangChain, LangGraph, and multi-agent patterns. The project is structured as a phased tutorial progressing from basic LangChain usage through multi-agent orchestration and API serving.

The LLM backend uses an OpenAI-compatible API — currently configured to use GLM (Zhipu AI) models but can work with any OpenAI-compatible endpoint.

Environment Setup

use langchain

# Copy env template and fill in API keys
cp .env.example .env

# Install dependencies with uv (preferred)
uv sync

# Or with pip
pip install -e .

Required .env variables:

  • OPENAI_API_KEY — API key for the LLM provider
  • OPENAI_API_BASE — Base URL (e.g., https://open.bigmodel.cn/api/paas/v4 for GLM)

Commands

# Run the main entry point
uv run python main.py
# or
python main.py

# Run a specific phase script (once implemented)
uv run python -m retrieval.loaders

There is currently no test runner or linter configured in pyproject.toml.

Architecture

The project directory grows incrementally across phases (see docs/phased-tutorial/00-总览与阶段索引.md):

Directory Phase Purpose
main.py 0–1 Entry point; minimal LangChain chain
config/settings.py 1+ Pydantic-based settings loaded from .env
retrieval/ 2+ Document loaders, text splitters, vector store
chains/ 1+ LCEL chain definitions; rag.py added in Phase 4
agents/ 5+ Tool-using agents with ReAct pattern
graph/ 6+ LangGraph StateGraph definitions (state, builder)
api/ 8 FastAPI + LangServe HTTP endpoint

Core data flow (final system): User query → LangGraph orchestrator → routes to specialist agents (retrieval agent uses ChromaDB vector store) → results aggregated → streamed response via FastAPI/LangServe.

Key LangChain patterns used:

  • LCEL (prompt | llm | parser) for chain composition
  • ChatOpenAI with custom base_url for OpenAI-compatible providers
  • langchain_chroma + langchain_text_splitters for RAG pipeline
  • StateGraph from LangGraph for multi-agent orchestration
  • LangServe for serving chains as REST endpoints

Configuration

config/settings.py uses pydantic-settings with BaseSettings — all config comes from environment variables or .env file. The Settings class is the single source of truth for runtime config; prefer importing from it over os.getenv() directly.