With the previous approach, we are using a list of titles. Any item got mistaken or item order shift will break the filtering. This approach is pretty fragile.
Switched to marking every video with a unique id, and passing these ids to the llm. It makes sure in every step, the video id and title are connected. This increased the accuracy of the filter and made it much more reliable.
With the previous approach, we are using a list of titles. Any item got mistaken or item order shift will break the filtering. This approach is pretty fragile. Switched to marking every video with a unique id, and passing these ids to the llm. It makes sure in every step, the video id and title are connected. This increased the accuracy of the filter and made it much more reliable.