Repo QA Assistant (Browser-Only)

🚀 A repo-aware AI assistant that can answer questions about your codebase with citations, built to run entirely in the browser — no servers, no API keys, zero cost.

---

✨ Features

• Retrieval-Augmented Generation (RAG): combines semantic + exact code search with a lightweight text generator.
• Embeddings & Search:
  • sentence-transformers/all-MiniLM-L6-v2 for embeddings (semantic similarity).
  • Riggrep MCP for fast exact search.
  • Chroma MCP as a vector database for semantic search.
• Generation:
  • Xenova/LaMini-Flan-T5-248M (~248M params, quantized) for lightweight instruction following in the browser.
• Client-Side Deployment:
  • All models run in the browser using Transformers.js with WebGPU/WebAssembly.
  • No backend, no API costs.
• Memory:
  • Short-term memory of recent turns for continuity.
  • Persistent chat history saved in localStorage.
  • Ability to pin notes (embedded and retrieved like code chunks).
  • Optional rolling summaries of the dialogue.
• UI:
  • Built with Next.js + React + Tailwind CSS.
  • Chat interface with sources listed per answer.
  • Adjustable Top-K slider.
  • Reset chat, add notes, see retrieval scores.
• Deployment:
  • Deployed serverlessly on Vercel.
  • First load downloads models (cached in IndexedDB for reuse).

---

🏗️ Architecture

Repo → Python ingestion (build_index.py) → index.json (chunks + embeddings) | v Next.js App (Vercel) ←→ Transformers.js (browser) | v User Q → Embedder → Cosine Similarity (top-K chunks) → Generator | v Answer + Citations

---

⚡ How It Works

1. Preprocessing

   • A Python script (build_index.py) walks through repo files.
   • Splits code/docs into overlapping chunks.
   • Generates embeddings with all-MiniLM-L6-v2.
   • Saves everything into public/index.json.

2. Retrieval

   • On each query, the question is embedded in-browser.
   • Chunks are scored by cosine similarity.
   • Top-K most relevant chunks (and pinned notes) are selected.

3. Generation

   • A prompt is built with:
     • The user’s question,
     • Recent chat history,
     • Retrieved context chunks,
     • Guidance for handling subjective questions.
   • LaMini-Flan-T5-248M generates a concise answer.
   • Sources are displayed with file + line spans.

4. Deployment

   • Everything runs client-side.
   • First model download is heavy (~100–250 MB), but cached after.

---

🧪 Tech Stack

• Frontend: Next.js, React, Tailwind CSS
• AI/ML:
  • Transformers.js
  • Hugging Face Models (MiniLM, LaMini-Flan-T5-248M)
  • Chroma DB (for semantic search, via MCP)
  • Riggrep (for exact code search, via MCP)
• Ingestion: Python, sentence-transformers
• Deployment: Vercel (serverless, static hosting)

---

📚 Skills Learned

• Retrieval-Augmented Generation (RAG)
• Embeddings & semantic search
• Vector databases (Chroma)
• Exact search (riggrep)
• Browser-only ML deployment (Transformers.js, WebGPU/WASM)
• Next.js/React/Tailwind frontend development
• Python scripting for data preprocessing
• Serverless deployment (Vercel)
• Trade-off analysis: cost vs performance vs usability
• Communicating model constraints in UI/UX

---

🚧 Limitations

• Small model (248M params) → short context window (~512–1024 tokens), limited reasoning.
• First load is slow → large model files (~100–250 MB) must download to the browser.
• Answers are factual extracts → no deep reasoning across many files.
• Best for demo / proof-of-concept, not production-level repo QA.

🚀 Try It

Deployed on Vercel:
👉 Live Demo Link