The day Biznetrix committed to AI inside the conversation. Fourteen migrations in one session — conversations, messages, then a five-table KB stack (sources, documents, chunks, embeddings, search logs), then the AI policy layer (response rules, prompt templates, escalation policies, response logs), contacts, and outbound webhooks. Twenty-seven PHP files wired outside migrations on the same day, meaning everything needed to use these tables was built the same session they were created.
The five-table KB structure is the backbone of the whole chat intelligence system. Sources hold the origin, documents break it into pages, chunks are the retrieval units, embeddings hold the vector representation, search logs track what worked and what did not. Every layer serves a purpose. Designing it this way from the start rather than bolting it on later was the right call.
AI escalation policies — deciding when to hand off to a human — were designed on day one of the AI layer, not as an afterthought. That kind of intentional thinking is what separates an AI feature from an AI product.