A RAG API that outsources the heavy lifting to everyone else

What it does

rag-search is a FastAPI server that takes a query, fetches web results via Serper, embeds them with OpenAI (text-embedding-ada-002), and runs them through a Zilliz vector collection before filtering and reranking the top hits. The README treats the whole thing as a wiring exercise: bring API keys for search, embeddings, vector storage, and bearer auth, and the server orchestrates the rest.

The interesting bit

The project exposes every stage of the pipeline—search depth, reranking, score thresholds, top-k limits—as JSON knobs on the /rag-search endpoint. That makes it less a framework and more a preset retrieval pipeline for teams already paying for Serper, OpenAI, and Zilliz.

Key highlights

• Single /rag-search POST endpoint with per-request toggles for is_reranking, is_filter, and is_detail.
• Hard-wired for Serper web search, OpenAI embeddings, and Zilliz/Milvus vector storage.
• Bearer-token auth via AUTH_API_KEY.
• Response format and any LLM generation step are not shown in the README.

Caveats

• Documentation stops at the quickstart; architecture diagrams, response schemas, and error handling are left to the source code.
• Swapping out Serper, OpenAI, or Zilliz would require editing code, not just config files.
• Needs four separate external API keys and a live Zilliz instance before it returns its first result.

Verdict

Useful if you want a drop-in, opinionated retrieval API and your stack already looks exactly like this. Skip it if you need pluggable providers, local models, or enough documentation to debug without reading the source.