Class-Recording RAG
A production RAG assistant that grounds each student in their own class recordings. Handling student audio, so the deep write-up is shared on request.
What it is
A production retrieval assistant at Impact Solutions Lab that gives each student a chatbot grounded in their own class recordings. Recordings come in daily, are transcribed, indexed per student, and answered from, so a student can ask what a class covered and get a grounded reply drawn only from their own sessions.
Because this system handles students' audio, this page shows the system design in the abstract and nothing else. No recordings, no transcripts, no student data.
System design
- Dual-language transcription. Class audio is transcribed with WhisperX, choosing the stronger of two languages per segment, with silence filtered to avoid hallucinated text.
- Per-student isolation. Each student's content is indexed under their own scope, and retrieval is gated so an answer can only ever draw from that student's own recordings. This boundary is enforced in the pipeline, not left to a prompt.
- Automated ingestion. New recordings are picked up on a schedule and processed idempotently, so the same session is never indexed twice.
The rest is on request
The full technical write-up, the ingestion and identity model, the retrieval design, the tradeoffs, and the lessons, is kept private because of the data involved. If you'd like to walk through it, reach out and I'm happy to share it selectively.