A TypeScript RAG Stack That Turns Project Decisions Into a Queryable Asset

Most RAG tutorials stop at a demo that stuffs PDFs into a vector store and calls it done. CorpRAG models the harder, higher-value problem: maintaining a living index of project rationale that an AI agent can query through a standard protocol, with every answer pointing back to a specific source file.

Project context — why JWT was chosen over sessions, why a table was split, what constraint a business rule addressed — decays faster than code. CorpRAG captures these decisions as structured Markdown documents and runs them through a full RAG pipeline: chunking, embedding with nomic-embed-text via Ollama, and similarity search in LanceDB. The system separates knowledge by business domain and tags each chunk with metadata for filtering and source traceability.

Instead of bolting a chatbot directly onto the retrieval layer, the project exposes query tools through the Model Context Protocol. An MCP-capable AI client can call domain-specific tools like ext_query or fr_query on demand, keeping context windows small and answers grounded in specific source documents. The architecture decouples retrieval from generation, so the knowledge base can be re-indexed without touching client prompts.

The column that accompanies the code treats RAG as a data pipeline to debug, not a magic API. It walks through each link — document loading, splitting, embedding, vector storage, and MCP tool wrapping — and stresses that answer quality is the product of every stage multiplied together.