Most AI assistants answer from what the model "remembers" from training. That's fine for trivia. It's not fine when the answer has to match a company's actual, current material, and the model just confidently makes something up.
The problem
A coaching platform needed an assistant that answered strictly from their own content, not general knowledge, and not a hallucinated approximation of it. They also needed to keep editing that content in the tools they already used, not migrate into a new CMS just to feed an AI.
The approach
I embedded the source content into a vector index and retrieved it per query, so every answer is grounded in the client's actual corpus rather than model recall. The assistant runs across multiple model providers behind one interface, so no single provider is a point of failure. Content syncs directly from the client's existing operational tooling, their non-technical team keeps editing where they already work. A scheduler drives recurring prompts with no human in the loop.
Stack: Python, FastAPI, LangChain, OpenAI, Anthropic, Pinecone, Postgres, Docker
Result: answers grounded in real content, delivered in Slack, running unattended on a schedule, with zero migration friction for the non-technical team maintaining it.
The lesson: if your AI's accuracy depends on the model remembering correctly, you don't have an accuracy problem. You have a retrieval problem.
Full breakdown → https://farzamazhar.com/#work