Developer Builds Agentic RAG AI Agent Using FAISS, BM25 and Qwen to Limit Hallucinations

A developer shared a personal AI engineering project in which they built an Agentic Retrieval-Augmented Generation (RAG) system designed to answer questions strictly from a provided knowledge base. The system combines FAISS for semantic vector search and BM25 for keyword-based retrieval, merging both results through a hybrid search pipeline before passing context to the Qwen2.5-72B-Instruct language model. The core motivation was to prevent the model from relying on its pre-trained general knowledge, instead grounding all responses in retrieved documents. During development, the creator abandoned an initial CodeAgent approach after the model generated unsupported tool calls, opting instead for a simpler and more reliable direct RAG pipeline. The project demonstrates a practical method for constraining large language model outputs to a defined knowledge domain, with the model instructed to reject questions it cannot answer from available context.