Conversational AI Assistant

Strategy & Knowledge • Production system

Live: ask.bastian-brand.com

What it is

A conversational AI agent that answers questions from a private knowledge base — grounded, with cited sources, and refusing to invent when the corpus has nothing. It’s not a flat “chat over documents”: the agent decides which retrieval tool to call (search summaries, search full transcripts, list sources, fetch recent or per-channel items), gathers evidence, and only then answers, naming the source of every claim. This is a live text demo of a system that also runs as a voice agent.

The knowledge base here is built from ~40 curated AI & coding YouTube channels, so you can ask it what’s new in AI and get a sourced answer.

How I built it

The retrieval brain is an OpenAI-compatible service that runs an agentic tool-loop (function-calling) over a self-hosted pgvector RAG store: hybrid retrieval, a source-citation layer that tags every hit with its channel, title and date, and a grounding prompt that forbids ungrounded claims. In production it’s the custom-LLM behind a VAPI voice assistant; this page exposes the same brain as text. For a public demo I stripped the action tools (so it can only read, never send anything) and rate-limited every request.

Highlights

• Agentic retrieval (the model picks the tool), not a single canned RAG call
• Grounded + cited: every claim carries its source; no source, no claim
• One brain, two channels — the same engine powers voice and text

Stack

Python · FastAPI · OpenRouter (function-calling) · pgvector RAG · VAPI (voice)