Cheap RAG in Go with Gemini File Search: No Vector DB, Two Calls, One Hosted Store

Developer Maneshwar shares his experience building a cost-effective Retrieval-Augmented Generation (RAG) system using Go and Google's Gemini File Search. By leveraging Gemini's native file indexing, the author avoids the complexity and costs associated with traditional vector databases and embedding pipelines. The architecture utilizes a two-call approach: the first call analyzes the input to generate generalized retrieval queries, while the second call uses the File Search tool to retrieve relevant context and draft a response. Key technical challenges, such as document chunking, managing API key-specific stores, and ensuring deterministic output through self-auditing checks, are addressed in detail. The system also employs a shared manifest to prevent duplicate uploads to the Gemini store. This implementation demonstrates how to build a robust, citation-backed AI tool for internal documentation without the overhead of managing specialized infrastructure.