Introducing SID-2, a new search model built on compute-scaling swarms.
Our swarms find better results than Opus and Astra, while being 8x faster and 20x lower cost.
By 2030, AI will write an internet's worth of text every 24 hours. Current search pipelines were built on hand-designed rules, adding more won't scale to meet this incoming volume of data.
SID-2 is available on our API today. Try it on your most demanding searches.
There's a place for old school classic RAG when you don't trust the agent to even know what question to ask search, so you just want to inject some context before work even begins
@chm_hamayun They use the same exact schema. But I do try to get the agents to pass a param to the browser ?mode=agent so I can distinguish their traffic
My search consulting moved away from LLM judges. Now I:
* Vibecode a facsimile of my clients application
* Let agents act as users to use the app to solve search use-cases
When using agent evals, I stay in the lane of sessions, clicks, etc - not NDCG
https://t.co/i6nKX8TYTp
Maven is running a sale this week, including our AI-Powered Search course. It's a great time to get training with leftover training funds
https://t.co/yuA0yXuyuW
Why is 'classic' Learning to Rank still relevant in a world of embeddings and cross-encoders?
Because classic ML still does a good job of accounting for the full range of domain-specific ranking factors that have nothing to do with unstructured text similarity
Instead of LLM-as-a-judge, let agents simulate traffic on search UI's. The search UI adds as much friction as ranking. And its hard to disentangle ranking from the display of search results.
(Of course, human traffic teaches us more, but agents help with cold-start problems)
Turn user search traffic on your site as a corpus to search itself. No model training. No deploying GPUs. Just simple aggregations
That's what I did with Alamo Drafthouse's search. They have a fat head query traffic, benefiting from just knowing the right answer for every query
https://t.co/LcBSUhNkdM
Introducing e-means, an algorithm beating k-means in 10x fewer epochs.
With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
The recording is now available for enrolled students, with late enrollment still possible.
Today, we covered a broad overview of many different AI query modalities (single vector, multi-vector/late interaction, personalized search with clustering-based guardrails, various hybrid search algorithms, behaviors vector spaces and queries, etc.
Tomorrow is Blitzrank for Evals. Thursday is Learning to Rank / Building Ranking Classifiers.
Many weeks still to go after that (until Dec 4).