Awesome Search KG

How to Tune Hybrid Search in Qdrant

4 min read

How to Tune Hybrid Search in Qdrant

Source: https://qdrant.tech/articles/how-to-tune-hybrid-search/ Author: Dylan Couzon

Summary

A tuning workflow for Hybrid Search fusion in Qdrant: confirm fusion beats either prefetch alone, choose between RRF and DBSF on labeled data, set RRF’s k from how many relevant documents each query tends to have, sweep per-prefetch weights last, then validate the winner on held-out queries.

Setup

Measurements use five public datasets (5,183–100,000 documents) — SciFact, ArguAna, WANDS, CodeSearchNet, DBPedia-entity — with all-MiniLM-L6-v2 for dense retrieval, Qdrant’s core BM25 for sparse retrieval, and 200 candidates from each prefetch, scored with nDCG@10.

Confirm Fusion Beats Either Prefetch

Default RRF (k=2, equal weights — Qdrant’s default) against dense-only and sparse-only retrieval:

DatasetDenseSparseRRF (k=2)Over the better one
SciFact0.62390.68860.7175+0.0289
ArguAna0.49050.42240.5216+0.0311
WANDS0.69210.70980.7254+0.0156
CodeSearchNet0.62990.51260.6555+0.0256
DBPedia-entity0.46770.38570.4638−0.0039

Fusion beat both prefetches on four of five datasets, with 95% intervals excluding zero; on DBPedia-entity fusion trailed dense retrieval alone and the interval crossed zero.

RRF vs. DBSF

RRF uses only a candidate’s rank position — a document at rank 1 scores the same whether it led rank 2 by a wide or narrow margin. DBSF puts both retrievers’ scores on one scale per query, using each list’s mean and spread, so a wide score lead survives into the fused ranking. Which one wins depends on the data; the article’s recommendation is to run both against a labeled query set. See Distribution-Based Score Fusion for the formula and the article’s five-dataset DBSF-vs-RRF comparison table.

Choosing RRF’s k

Qdrant scores a document at position pos as 1/((pos+1)/weight + k − 1), which reduces to 1/(pos+k) at equal weights. At Qdrant’s default k=2, rank 1 carries 5.50× the weight of rank 10; at k=61 (the convention from the original RRF paper, one-based k=60, mapped to Qdrant’s zero-based positions), rank 1 carries only 1.15× the weight.

Sweeping k over 1, 2, 5, 20, and 61 across the five datasets: datasets with about one relevant document per query (ArguAna, CodeSearchNet, SciFact) did best at k=2 or k=5; datasets with tens or hundreds of relevant documents per query (DBPedia-entity, WANDS) did best at k=20 or k=61. On WANDS, k=2 and k=61 picked a different top result for 42% of queries while nDCG@10 rose 0.0360.

Tied scores are more common at low k — SciFact’s default-RRF top 10 shared a score with an adjacent result in 12.5% of queries on average, versus 2.8% at k=61 and none under DBSF. The fix: request more than 10 results, sort client-side by descending score then ascending ID, keep the first 10.

Tuning Weights

A weight pair multiplies each prefetch’s contribution; the pair is absolute, so (1, 2) and (2, 4) are different settings because the formula divides position by weight. Weights are only valid for the k they were tested at — on WANDS at k=5, (1, 2) scores 0.7390 and (2, 4) scores 0.7508, but at k=61 (WANDS’s best k) equal weights win, 0.7614 vs. 0.7567. A prefetch’s own standalone score doesn’t say which way to lean: on DBPedia-entity, dense scores higher solo (0.4677 vs. 0.3857) yet the winning pair weights sparse 3× and gains 0.0060; CodeSearchNet leans the other way, weighting dense 2× for a 0.0096 gain. A weight of 0.0 keeps a prefetch’s documents in the result at score 0.0 rather than excluding them.

Validate on Held-Out Queries

A configuration can win on the queries used to select it and still fail on held-out queries. On SciFact’s 300 queries, no tested configuration — including DBSF’s +0.0148 gain — had a 95% interval excluding zero; across 200 random splits, a selected configuration retained 67%–95% of its gain on held-out data. Keeping the default was the correct answer on one of the five datasets.

Tuning Order

  1. Confirm fusion beats either prefetch alone.
  2. Pick RRF or DBSF on labeled data.
  3. Set k from relevant documents per query.
  4. Sweep a few weight pairs at that k.
  5. Validate the winner on held-out queries before shipping.
  • Hybrid Queries - Qdrant — the DBSF formula and API this article’s DBSF results build on
  • RRF is Not Enough — Doug Turnbull; the case that fusion strategy can’t fix upstream retrieval quality, complementary to this article’s premise that fusion parameters matter

Graph View

Backlinks