Why keyword search hits its limits
What BM25 does well, what it cannot do, and why vector similarity search complements the lexical index rather than replacing it.
Read the article Rt
Research team
Vela Applied Research
The collective working on embedding models, recall quality and evaluation
About this team
The research team works on everything that precedes and surrounds the index: embedding model selection, recall measurement, evaluation protocols, quantization, and how similarity metrics behave across domains and languages.
Its publications are collective notes. They focus on methods and reproducible results rather than individual journeys, and are not attributed to any named person.
Articles published by Research team (2)
Choosing your embedding dimension
Memory cost per vector, how dimension affects recall, quantization and truncation: how to decide before indexing a large corpus.
Read the article