AI-powered FAQ retrieval system for live customer support using hybrid search (BM25 + dense retrieval), reranking, grounding checks, and validation loops to deliver accurate, reliable answer suggestions in real time.Implemented multi-turn query handling, confidence-based decision logic, and stateful orchestration with LangGraph for retries and workflow management.
An exploration of how you fix an AI answer while it’s still being written.
In most AI chats, if one line of the reply is wrong, you wait for the whole answer, then retype the request and start over. Here you tap the sentence that’s off, choose Correct, and say what should change in your own words. The fix waits in a queue in the input bar while the plan keeps writing, then folds into the reply in place. You see the old line struck through and the new one highlighted, with an Undo.
Every claim in the plan cites its source, and your correction appears beside them as a source too. That way you can always see why the plan says what it says.