Technology method · wiki:deep
Adaptive / funnel retrieval
Use a short Matryoshka prefix (cheap ANN) to shortlist, then rerank with a longer prefix or full vector — the operational payoff of OpenAI/mxbai truncation.
Adaptive / funnel retrieval turns Matryoshka truncation into a system: query and index at small m for throughput, then score the candidate set at larger m (or full d). Cascades that shrink k while growing m are sometimes called funnel retrieval (MRL adaptive-retrieval literature; vendor blogs on OpenAI Matryoshka + pgvector).
Example shape: embed with OpenAI 3-large at full 3072 or store 1024; first-pass ANN at 256–512; rerank top-k at 1024/full. Same pattern for mxbai native 1024 → first pass 512.
Depends on ANN indexes and stores in Vector DBs. Still a RAG concern: citation quality must be measured at the final fidelity you ship.
See also
In this taxonomy
_catalog/taxonomy/tech-adaptive-funnel-retrieval.md