Free Lesson
Neural search at BM25 latency
60 min
Sep 22, 2026 1:00 PM
By continuing, you agree to Maven's Terms and Privacy Policy.
What you'll learn
Better documents help agent+human queries
What does "asymmetric" actually mean, and why doesn't dropping the model on the query side wreck your relevance?
What do you trade for a cheaper query?
Something has to give. Is it relevance, index size, or indexing time?
Where does it fit and how to take it to production?
Pros and cons vs just BM25? Along with benchmarks across different models and search engines.
Why this topic matters
SPLADE uses a neural model to expand your documents' keywords, so keyword search can match synonyms (and other semantically similar keywords). But it runs a model on every query, which adds about 50ms latency and needs a GPU. Inference-free SPLADE skips that step. It does all the model work upfront on your documents, so queries cost about as much as BM25. We'll talk about how that works, what you give up, and how to take it to production.
