Ghousted/CodeSage

CodeSage Backend

FastAPI service that powers CodeSage — clones a GitHub repo, chunks + embeds it, stores vectors in Pinecone, and answers questions over the resulting index using an LLM via the Hugging Face Inference API.

Required secrets

Set these in Settings → Variables and secrets (as secrets, not variables):

• —PINECONE_API_KEY — your Pinecone API key
• —HUGGINGFACE_API_KEY — HF Inference API token with read access to the chosen LLM

Optional environment variables

Set these as variables (non-secret):

• —PINECONE_INDEX_NAME — defaults to codesage
• —EMBEDDING_MODEL — defaults to sentence-transformers/all-MiniLM-L6-v2
• —LLM_MODEL — defaults to Qwen/Qwen2.5-7B-Instruct
• —ALLOWED_ORIGINS — comma-separated CORS origins (set to your Vercel URL in prod)
• —GITHUB_TOKEN — optional, raises GitHub API rate limits from 60 → 5000/hr
• —LOG_LEVEL — defaults to INFO
• —WARMUP_EMBEDDER — 1 to pre-load the embedding model on boot

Health check

GET /health returns { "status": "ok" } once required secrets are present.