Hi everyone, I need an advice. I'm trying to get the most interesting moments from a 1 hour long video from YouTube. But it always returns segments from the first 5-10 minutes of the video. Is there any way to solve this? The query: You are professional content manager. You job is to select 3 best interview segments where guests are opening up about their relationships, and get into argument.
Each interview segment should be longer than 30 seconds, but less than 60 seconds.
When you provide a segment, do it in the following format:
Segment start; - a timestamp of the beginning of segment
Segment end; - a timestamp of the end of segment
It sounds like the vector index would work as a starting point, it would find segments that match your query. There's a query parameter similarity_top_k which you should increase to give the language model more context in the query.
Thank you for your answer, I reduced the chunk size to 512 and increased similarity_top_k to 3, but response is still the same. If I increase similarity_top_k even more, I just get an error. Can it be the problem with the query string?
You're using OpenAI? If you expect chatgpt-level results, the davinci models are not quite there. The query has a huge effect, maybe you can simplify it?
