hint: 8 days ago txtai released their arxiv embeddings
Show HN: I made a website to semantically search ArXiv papers
71–80 of 107 posts
Re: Show HN: I made a website to semantically search ArXiv papers
#72I tried a simple search by author and it didn’t work. All the fancy stuff is great, but I’d expect the basics still work, in the end it’s a search engine for papers.
Re: Show HN: I made a website to semantically search ArXiv papers
#73I tried a simple search by author and it didn’t work. All the fancy stuff is great, but I’d expect the basics still work, in the end it’s a search engine for papers.
Maybe use the right tool for the job? Author names generally don’t have a lot of semantics associated with them and definitely not in the abstract.
Re: Show HN: I made a website to semantically search ArXiv papers
#74Earlier quoted context omitted.
That looks great for speed, but what about recall?
That's has a major downgrade. For binary embeddings, the top 10 results are same as fp32, albeit shuffled. However after the 10th result, I think quality degrades quite a bit. I was planning to add a reranking strategy for binary embeddings. What do you think?
Try this trick that I learned from Cohere:
- Fetch top 10*k (i.e. 100) results using the hamming distance
- Rerank by taking dot product between query embedding (full precision) and binary doc embeddings
- Show top-10 results after re-ranking
Re: Show HN: I made a website to semantically search ArXiv papers
#75Nice but I have to point out that a systematic review cannot be done with semantic search and should never be done in a preprint collection.
but it can provide recommendations
Re: Show HN: I made a website to semantically search ArXiv papers
#76This looks great, thanks for building this.
Something on similar lines which many may link, Research Rabbit - https://www.researchrabbit.ai/
Re: Show HN: I made a website to semantically search ArXiv papers
#77Earlier quoted context omitted.
That's has a major downgrade. For binary embeddings, the top 10 results are same as fp32, albeit shuffled. However after the 10th result, I think quality degrades quite a bit. I was planning to add a reranking strategy for binary embeddings. What do you think?
Try this trick that I learned from Cohere: - Fetch top 10*k (i.e. 100) results using the hamming distance - Rerank by taking dot product between query embedding (full precision) and binary doc embeddings - Show top-10 results after re-ranking
This is pretty cool. The dot product would give the unnormalized cosine similarity from a smaller pool. Thank you so much!
Re: Show HN: I made a website to semantically search ArXiv papers
#78Earlier quoted context omitted.
One of the factors is how users phrase their queries. On some level people are used to full text search but semantic shines when they ask literal questions with terminology that may not match the answer.
Exactly. Full text paradigm has it's own pros and I believe we need those tools in the new vector search to take full advantage. I am planning to add keywords feature where if a user enters something in "quotes", the would need to be in the shown results. Just like you can do with a google search.
You might be interested in hybrid search which issues both a full text and semantic search and then merges the results via reciprocal rank fusion.
Re: Show HN: I made a website to semantically search ArXiv papers
#79Earlier quoted context omitted.
Exactly. Full text paradigm has it's own pros and I believe we need those tools in the new vector search to take full advantage. I am planning to add keywords feature where if a user enters something in "quotes", the would need to be in the shown results. Just like you can do with a google search.
You might be interested in hybrid search which issues both a full text and semantic search and then merges the results via reciprocal rank fusion.
Thank you! I shall play with it this weekend :D
Re: Show HN: I made a website to semantically search ArXiv papers
#80This is awesome! If you’re interested, you could add a search tool client for your backend in paper-qa ( https://github.com/Future-House/paper-qa ). Then paper-qa users would be able to use your semantic search as part of its workflow.
I advise against it since binarized hamming distance isn't exactly that good unless your vector length is say a million.