Apppublic

Dharanidhar1709/Repository-Analysis-Assistant

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes

App README

Codebase Intelligence

AI-powered semantic search, Q&A, and summarization over any codebase.

Live Demo: https://ai-codebase-assistant-production.up.railway.app

What it does

Upload a .zip or paste a GitHub URL and instantly:

—Ask natural language questions about your code
—Search functions and classes semantically
—Summarize the entire repository with AI

Tech Stack

Layer	Tech
Backend	FastAPI, Python
Embeddings	SentenceTransformers (`all-MiniLM-L6-v2`)
Vector DB	ChromaDB
LLM	Groq API (llama-3.3-70b)
Frontend	React + Vite
Deploy	Railway + Docker

Architecture

ZIP / GitHub URL → Clone/Extract → File Chunking → Embeddings → ChromaDB
                                                                      ↓
User Query → Embed Query → Semantic Search → Top-K Chunks → Groq LLM → Answer

Features

—GitHub URL indexing — paste any public repo URL, no zip needed
—Multi-language support — Python (AST), Jupyter Notebooks, Java, C, C++, JS, TS, Go, Rust, Ruby
—AST-based Python chunking — extracts functions, classes, methods with metadata
—Notebook support — indexes .ipynb code cells as individual chunks
—Multi-session isolation — each upload gets a unique ChromaDB collection, multiple users never collide
—Large repo handling — smart sampling across files, truncated file trees to stay within LLM context limits
—Rich embeddings — chunks embedded with file context, docstrings, dependencies for better semantic search
—Auto docstring generation — generates missing docstrings and re-indexes
—Dual UI themes — Light (SaaS) and Dark (Hacker) modes

How Large Repos Are Handled

For repos with 100s–1000s of files:

—File tree is capped at 100 entries with a count of remaining files
—Summary samples 1 representative chunk per unique file (up to 15 files) instead of top-N chunks from same file
—Ask/Search are unaffected — full semantic search runs across all indexed chunks
—Embedding batch size tuned to 16 to balance RAM and speed on free tier
—Large repos (1000+ chunks) may take 3-5 minutes to index due to CPU constraints

Run Locally

bash

git clone https://github.com/Dharanidhar-17-09-2006/ai-codebase-assistant
cd ai-codebase-assistant
pip install -r requirements.txt
# add GROQ_API_KEY to .env
cd frontend && npm install && npm run build && cd ..
uvicorn app.main:app --reload
# open http://localhost:8000

API Endpoints

Method	Endpoint	Description
POST	`/repo/upload`	Upload zip, returns `collection_name`
POST	`/repo/index-url`	Index from GitHub URL
POST	`/repo/ask`	RAG Q&A over indexed codebase
POST	`/repo/search`	Semantic search over chunks
POST	`/repo/summarize`	AI summary of repository
POST	`/repo/generate-docstrings`	Auto-generate missing docstrings
DELETE	`/repo/session/{upload_id}`	Clean up ChromaDB collections

Resume

Built an AI codebase assistant using a RAG pipeline (SentenceTransformers + ChromaDB + Groq LLM) with FastAPI backend, supporting Python AST chunking, Jupyter notebook indexing, multi-language support (Java, C, Go, Rust etc.), GitHub URL indexing, multi-user session isolation, smart large-repo summarization, and a React frontend with dual themes — deployed on Railway with Docker.