You give it a transcript in [MM:SS - MM:SS] text format and a question. It
returns the chunk that best answers it, with its timestamp.
There are two search methods. The first is TF-IDF from scikit-learn, using
single words and word pairs, with sublinear term-frequency scaling. The second
embeds every chunk with the all-MiniLM-L6-v2 sentence-transformer model and
returns the chunk with the highest cosine similarity, if it scores above 0.3.
The sample transcript has 242 chunks.
You can use it through a Streamlit app (upload a transcript, then ask) or an interactive CLI. Every question and answer is logged to a file.
I never built a labelled set of test questions, so I don't have an accuracy number for either method.