Filtered Vector Search Is a Regime Map, Not a Leaderboard: FAISS, pgvector, and Qdrant under Selectivity, Churn, and Hybrid Fusion

Production vector search is almost never the unfiltered, read-only problem that leaderboards measure. Retrieval systems filter by tenant and permission, update continuously, and mix lexical with dense signals. This study measures three deployed shapes of vector search, a library (FAISS 1.15.1), a SQL extension (pgvector 0.8.0 on PostgreSQL 16), and a dedicated engine (Qdrant, client 1.20.0, embedded local mode), on one corpus (SQuAD v1.1 paragraphs chunked to 58,999 vectors of 256 dimensions), one embedding model, one query set, and one exact ground truth, with filtering, churn, and hybrid fusion as the variables. At matched recall on the unfiltered cell (breadth 40 for every backend), unfiltered recall@10 is FAISS 0.967, pgvector 0.978, Qdrant 1.000. Under a random-label filter at about 1 percent selectivity, FAISS falls to 0.413 with a 0.267 shortfall rate, while pgvector with iterative scanning holds 1.000 and Qdrant holds 1.000. The famous pgvector overfiltering trap does not appear where the arithmetic predicts it (the tight cell, where the planner leaves the approximate index and answers exactly); it appears one band up, at about 10 percent selectivity, where defaults return recall 0.486, a 0.990 shortfall rate, and a mean of 4.86 results against a prediction of 4.75. After a 10 percent delete and insert churn cycle, FAISS without bookkeeping serves 11.3 percent stale results, while pgvector and Qdrant serve none. Hybrid fusion over a shared BM25 ranking helps two backends past the pre-registered gain line and misses it on the third. Six hypotheses were registered in advance; four are supported, one is partially supported, and one, the trap hypothesis, is refuted in an instructive direction. The deliverable is a regime map with named boundaries, not a ranking.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24282172
Primary Topic
Information Retrieval and Search Behavior
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Filtered Vector Search Is a Regime Map, Not a Leaderboard: FAISS, pgvector, and Qdrant under Selectivity, Churn, and Hybrid Fusion

Jesun Ahmad Ushno
4 links
Zenodo (CERN European Organization for Nuclear Research)
Information Retrieval and Search Behavior
preprint

Filtered Vector Search Is a Regime Map, Not a Leaderboard: FAISS, pgvector, and Qdrant under Selectivity, Churn, and Hybrid Fusion

Jesun Ahmad Ushno
preprint en

Abstract

Production vector search is almost never the unfiltered, read-only problem that leaderboards measure. Retrieval systems filter by tenant and permission, update continuously, and mix lexical with dense signals. This study measures three deployed shapes of vector search, a library (FAISS 1.15.1), a SQL extension (pgvector 0.8.0 on PostgreSQL 16), and a dedicated engine (Qdrant, client 1.20.0, embedded local mode), on one corpus (SQuAD v1.1 paragraphs chunked to 58,999 vectors of 256 dimensions), one embedding model, one query set, and one exact ground truth, with filtering, churn, and hybrid fusion as the variables. At matched recall on the unfiltered cell (breadth 40 for every backend), unfiltered recall@10 is FAISS 0.967, pgvector 0.978, Qdrant 1.000. Under a random-label filter at about 1 percent selectivity, FAISS falls to 0.413 with a 0.267 shortfall rate, while pgvector with iterative scanning holds 1.000 and Qdrant holds 1.000. The famous pgvector overfiltering trap does not appear where the arithmetic predicts it (the tight cell, where the planner leaves the approximate index and answers exactly); it appears one band up, at about 10 percent selectivity, where defaults return recall 0.486, a 0.990 shortfall rate, and a mean of 4.86 results against a prediction of 4.75. After a 10 percent delete and insert churn cycle, FAISS without bookkeeping serves 11.3 percent stale results, while pgvector and Qdrant serve none. Hybrid fusion over a shared BM25 ranking helps two backends past the pre-registered gain line and misses it on the third. Six hypotheses were registered in advance; four are supported, one is partially supported, and one, the trap hypothesis, is refuted in an instructive direction. The deliverable is a regime map with named boundaries, not a ranking.

Zenodo (CERN European Organization for Nuclear Research)
Information Retrieval and Search Behavior
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon