I share search tips, blog articles, and free events I'm hosting about the search+retreval industry, vector databases, information retrieval and more.
Building a better Jev reranker
|
In a Jev reranker you ask Jev to give P(Yes) of the answer to a question like Is this document relevant to the query? And you rank by P(yes) It works! But... that's kinda boring. I think we can do better: we could also have an LLM generate the questions per query. IE query "desk for kids" has us evaluate products with criteria like Is it less than 36 inches?
Is it a playful design?
Does it explicitly say its designed for children?
Then rerank this query using the sum of P(yes) of those questions. That's what I did in my latest blog post. It seems to improve relevance, with the tradeoff of the LLM call and more Jev questions per query. Still, as a way to generate a better Jev judge rubric for each query, very useful. And as decision models become cheaper, targeted towards domains, and locally hosted - this could gradually become more feasible production strategy. Enjoy! https://softwaredoug.com/blog/2026/10/08/bag-of-decisions
-Doug Slack Community * Events · Consulting · Training (use code search-tips) You're subscribed to Doug Turnbull's daily search tips where I share tips, blog articles, events, and more. You can always manage your profile: |
Doug Turnbull
I share search tips, blog articles, and free events I'm hosting about the search+retreval industry, vector databases, information retrieval and more.
Over the next two weeks, Trey Grainger and I are hosting a free online conference, The Future of AI Search w/ talks from some of your favorite search people. The lineup Unreasonable Effectiveness of BM25 for Agentic Search by Hornet.dev founder Jo Kristian-Bergum Introducing Luxir: Can we build the fastest search engine? by Solr creator Yonik Seeley Building topk-embed-v1: frontier multi-vector embeddings by Marek Galovic, CEO Top K Can a Decision Model Think like a Search Engineer by David...
Hey all, Just a note tomorrow I'll be giving a talk on High-Scale Agent Memory in Turbopuffer Its a first step into a larger world! Agent memory helps new sessions feel like they're continuing a conversation not restarting one. Come, have fun, see the mistakes I've made, and get an intro to Turbopuffer. Event link - https://maven.com/p/e48f16/high-scale-agentic-memory-with-turbopuffer Cheers, -Doug Slack Community * Events · Consulting · Training (use code search-tips) You're subscribed to...
This week the Retrieval Augmented Gathering online conference continues! Last week we explored index-only SPLADE, got a demo of tensors in the search engine, and how to think of agentic memory as telemetry. This week we've got talks below: Embeddings don't solve RAG (today) Solving retrieval for agents isn't as simple as chunk + embed. For one, similarity is not relevance. An decides what's relevant depending on not just raw text, but everything that 'surrounds' it. Places it in context....