I want to be able to search for a specific snippet in a website/document or a specific timestamp in a video/podcast. The service should receive a query and return the possible locations of the paragraph/phrase that best answers the query string. Of course, this won't be able to take the whole context, but this is ok here.
This will probably use techniques like embedding, and it will contrast to LLMs by returning the specific snippet where the thing searched is mentioned instead of the summary or key points of the whole document. This will eliminate almost all concerns about LLMs hallucination and false citation.
Example: "Show HN: Semantic search for video (semanticvideosearch.com)" (the URL doesn't work anymore, but you get the gist).
One of the best ways to deliver it is through the browser extension (e.g. highlight the snippet).
Related: https://kagifeedback.org/d/1171-smart-snippets-as-an-alternative-for-snippet-length (e.g. best snippet = best result of the semantic search using the user's query string).
