3 items tagged “jo-kristian-bergum”
2024
Looking back, it's clear we overcomplicated things. While embeddings fundamentally changed how we can represent and compare content, they didn't need an entirely new infrastructure category. What we label as "vector databases" are, in reality, search engines with vector capabilities. The market is already correcting this categorization—vector search providers rapidly add traditional search features while established search engines incorporate vector search capabilities. This category convergence isn't surprising: building a good retrieval engine has always been about combining multiple retrieval and ranking strategies. Vector search is just another powerful tool in that toolbox, not a category of its own.
If you have worked in search, you know how freaking hard even getting started with something close to this with traditional methods. Now, you can zero-shot it.
System Instructions: As a query categorization expert, you try to break down the intent of a search query. First, provide your reasoning and then describe the intent using a single category (broad, detailed, comparision)
User: The query from the user is "nike versus adidas for terrain running". The user is a female, age 22.Model: The user is clearly looking to compare two specific brands, Nike and Adidas, for a particular activity, terrain running. While the user's demographics might be helpful in some situations (e.g., recommending specific product lines), the core intent remains a comparison. Category: Comparison
There's a lot of hand-waving around query intent classification; it's always been like that. Now, it's straightforward (add a few examples to improve accuracy). But my point is that you could only dream about building something like this without having access to lots of interaction data.
Cohere int8 & binary Embeddings—Scale Your Vector Database to Large Datasets (via) Jo Kristian Bergum told me “The accuracy retention [of binary embedding vectors] is sensitive to whether the model has been using this binarization as part of the loss function.”
Cohere provide an API for embeddings, and last week added support for returning binary vectors specifically tuned in this way.
250M embeddings (Cohere provide a downloadable dataset of 250M embedded documents from Wikipedia) at float32 (4 bytes) is 954GB.
Cohere claim that reducing to 1 bit per dimension knocks that down to 30 GB (954/32) while keeping “90-98% of the original search quality”.