🔍 Teach AI Your Internal Library (RAG for Custom Codebases)

This project shows how to turn your internal codebase into a searchable, intelligent assistant using Retrieval-Augmented Generation (RAG), without retraining any models or exposing private code.

💡 Why This Matters

Most organizations use custom libraries that AI coding agents can't access. This project bridges that gap by:

• Parsing your code into meaningful chunks
• Generating natural language summaries with a local LLM
• Storing and retrieving these chunks with a vector DB
• Serving them to an LLM via API for intelligent Q&A

---

🧱 Architecture

Codebase → Parser → Chunks → LLM Summaries → Vector DB
                                        ↓
                                 RAG API via FastAPI
                                        ↓
                             → Smart AI Answers

You can also inject this context into any MCP server used by coding agents.

---

📂 Components

1. Parser.py / Parser_chroma.py

• Parses .swift files (or any similar structured language)
• Extracts declarations (classes, funcs, etc.)
• Summarizes each chunk using Phi-3-mini
• Outputs to markdown or ChromaDB

2. rag_server.py

• FastAPI server
• Accepts natural language queries
• Uses semantic search from ChromaDB
• Builds prompts and sends to the LLM
• Returns helpful contextual answers

---

🚀 How to Run

Prerequisites

• Python 3.10+
• Local LLM running at http://localhost:1234 (e.g., Phi-3-mini via LM Studio)
• chroma for vector storage

pip install httpx chromadb fastapi uvicorn

Step 1: Parse & Summarize

python Parser_chroma.py

This will:

• Walk through the ./BP directory
• Extract Swift code chunks
• Generate summaries
• Store them in ChromaDB

Step 2: Start the RAG API Server

uvicorn rag_server:app --reload

Step 3: Ask Questions

Send a POST request:

POST /ask
{
  "query": "What does the LoginCard component do?",
  "top_k": 5
}

---

🧪 Tested On

• Custom Swift-like component library mimicking Bootstrap
• Local setup using Phi-3-mini LLM
• Works offline and secure (no cloud calls)

---

🧠 Future Additions

• Tree-sitter-based language support
• UI interface for interactive exploration
• Support for other languages beyond Swift

---

📜 License

MIT License — Use it, extend it, contribute to it.

---

🙌 Credits

Created by Renjith For devs who want their AI to actually understand their code.