Dracula Chat

Vector similarity search over an FAQ set, with an OpenAI fallback when nothing clears the threshold. Configurable match thresholds and JSONL logging, so afterwards you can tell why it answered the way it did.

I built it to understand RAG properly. It worked well enough that the same approach went into the product at Sopeonow, which is the useful part of the story. That platform got a RAG assistant because of weekend R&D, not because anyone made a procurement decision.