RAG Semantic Search

This site demonstrates the business value of Retrieval-Augmented Generation (RAG): instead of relying on an AI model's built-in knowledge, answers are generated from your own knowledge base — content you control is retrieved and injected into the AI's prompt, so responses reflect your curated, up-to-date information rather than whatever the model happened to learn during training.

Under the hood: web pages are scraped (Firecrawl), cleaned, chunked, and embedded (bge-m3) into a Qdrant vector store. Queries run a two-stage retrieval: semantic search casts a wide net for recall — fetching every chunk that might be relevant — then a cross-encoder (bge-reranker-v2-m3) reranks those candidates for precision, keeping only the best. A local LLM (llama3.1:8b) composes the answer with per-source attribution. Pick a collection to try different use cases — legal, medical, financial, or any other knowledge base you load — and toggle Pro Mode to inspect the scored chunks behind each answer and browse the documents in the collection.

⚠️ This is a Demo
This is a self-hosted, unauthenticated demo of a full RAG pipeline: scrape → clean → chunk → embed → search → rerank → answer. That makes the knowledge base easy to demonstrate, but also easy to poison — exactly the kind of real-world risk companies face when write access is not protected. In production this sits behind auth, rate limits, domain allowlists, and an identity-aware proxy. Full technical details and the production threat model are in the GitHub README.