✳flâneur — a map of the web's best reading

The other hard retrieval problems

softwaredoug.com · 574 words · saved by 1 readers

I’m happy that as practitioners we now see every thing as living in a vector space. Indeed that’s exciting! We can store these vectors, query them, and find things most similar. The “vector database” (or really –retrieval system–) will become the center for user interactivity. I’m very enthusiastic about this future. What makes me a bit sad though - we think that the only vector dimensions that matter come from dense embeddings. In reality, to get to make step changes in retrieval systems, diverse, orthogonal features matter. Embeddings “blur” what we look at. We squint at our picture of a squirell and we see a vague N-legged furry thing. We recall in our minds all the other things we know: like dogs, cats, maybe lizards. Maybe tables? Stuffed animals? Amazing. Vector search helps us broaden out in a way that once seemed magical. But we also need the un-squinted, explicit information from the photo too! The high precision, engineered, domain-specific features. Features that tell us ind

In my darkest “old man yells at cloud” moments, in todays AI world, I get a little sad. I’m happy that as practitioners we now see every thing as living in a vector space. Indeed that’s exciting! We can store these vectors, query them, and find things most similar. The “vector database” (or really –retrieval system–) will become the center for user interactivity. I’m very enthusiastic about this future. What makes me a bit sad though - we think that the only vector dimensions that matter come from dense embeddings. In reality, to get to make step changes in retrieval systems, diverse, orthogon

Explore this link on the map →

related reading