Built a retrieval-augmented generation chatbot for a blockchain/digital ecosystem client whose product documentation was scattered across files, leaving users and internal teams with no fast, reliable way to get answers. The bot pulls strictly from the client's own docs — a Google Drive folder acts as the single source of truth, where any file created or updated is automatically downloaded, chunked, embedded with Cohere, and upserted into a Supabase vector store, so the knowledge base stays current without manual re-indexing. The chat agent, running on a Groq-hosted LLM for fast responses, is instructed to always query the vector store first, optimizing each user question into a targeted search before answering — and it's explicitly constrained to respond only from retrieved data, admitting when nothing relevant is found rather than guessing. Postgres-backed chat memory keeps context across a session, so it handles natural follow-ups instead of treating every message as a fresh conversation. The result: a self-updating support assistant that answers accurately and never fabricates product details.