Ask a question in Bangla and get an answer that points to the minute it was said. This notebook builds retrieval-augmented generation (RAG) over match or talk-show transcripts.
What you do
- Split transcripts into passages with timestamps and embed them with multilingual-e5 (MIT).
- Retrieve the best passages for a question and have Qwen2.5-7B answer using only those passages.
- See the source minute next to every answer.
Good to know
- If the passages do not contain the answer, the notebook is set to say so rather than guess.
- Use transcripts you have the right to use. About 10 minutes on a 24 GB AI Computer.