The right context.
At the right moment.
Follow a fictional conversation as a sliding context window detects intent and retrieves evidence cards. Step through the actual Python pipeline’s output.
Conversation
Fictional transcriptEvidence cards
Unverified excerptsSample sources include different accounts and conflicting terms. Compare source context; these excerpts are not a verified answer or a commercial commitment.
What you’re seeing
Four authored scenarios, a 30-second context window, heuristic intent classification and TF-IDF retrieval over bundled fictional knowledge. The cards are generated at build time by the real pipeline.
What runs locally
The source includes a FastAPI/SSE lab, optional WAV transcription and selective local-model escalation with retrieval fallback. Those services are separate from this browser replay.
Prototype boundaries
No live listening, user uploads or shared server session here. Scripted gold-moment results are regression checks, not evidence of accuracy on real meetings. Review the MIT-licensed source, tests and setup instructions on GitHub.