A user question can be a quick factual lookup, a request for structured analysis, or a mix of both stitched into one sentence. I wanted the backend to recognize that split automatically, answer each part with the node built for it, and still return one coherent response, rather than forcing every request through a single generic answering path.

I also wanted the UI to stay responsive during a multi-step answer, so the client shows progress as the graph works instead of a blank screen until the final text arrives.

The chat endpoint returns application/x-ndjson: one JSON object per line, typed as progress, text, error, or final_response. The React frontend consumes this stream incrementally, so partial answer text and status updates appear as the graph produces them rather than all at once.

Conversation history is saved to disk per session and reloaded by conversation ID, which is what makes a session resumable. Once a conversation passes 20 messages, older turns are compressed into a summary and the 10 most recent messages are kept in full.