ai-code-assistant

Project Summary

This project is a lightweight code assistant API for internal experimentation and learning. It reads local reference files, builds a searchable vector index, and answers coding questions through a simple FastAPI service. The goal is practical support for day-to-day development tasks, not a production-grade platform.

How It Works

1. ingest.py reads documents/sample.txt.
2. The text is split into chunks.
3. Embeddings are created with local Ollama embedding models.
4. Chunks are stored in a FAISS vector index under vectorstore/.
5. app.py loads that index and exposes a question-answer API.

API Endpoints

• GET /
  • Basic service info and endpoint list.
• GET /health
  • Health check endpoint.
• POST /ask
  • Accepts a JSON payload and returns an answer from retrieval + Ollama chat model.

Request body for /ask:

{
	"question": "What does this project do?"
}

Response example:

{
	"answer": "..."
}

Local Setup

1. Install dependencies:

./env3/bin/python -m pip install -r requirements.txt

2. Make sure Ollama is running and required models are available:

ollama serve
ollama pull nomic-embed-text
ollama pull llama3.2:latest

3. Build vector index:

./env3/bin/python ingest.py

4. Start API server:

./env3/bin/python -m uvicorn app:app --reload

5. Open docs:

http://127.0.0.1:8000/docs

Environment Variables

• OLLAMA_BASE_URL default: http://localhost:11434
• OLLAMA_EMBED_MODEL default: nomic-embed-text
• OLLAMA_CHAT_MODEL default: llama3.2:latest

Project Structure

• app.py: FastAPI application and QA endpoint.
• ingest.py: document loading, chunking, embedding, and FAISS index creation.
• documents/sample.txt: source content used for indexing.
• vectorstore/: generated FAISS index files (index.faiss, index.pkl).

Notes

• Run ingestion before starting the API; otherwise startup fails because the FAISS index is missing.
• This is a learning/demo project, so error handling and production hardening are intentionally minimal.