Built-In Search Ranking in 2026: How Hybrid Fusion Works
Explains how Seahorse Cloud fuses Dense and Sparse candidate ranks with RRF, what RRF K, Top K, and EF Search control, and how to compare search modes.
Key takeaways
- Hybrid fuses candidate ranks: Built-in RRF combines Dense and Sparse rankings.
- RRF K tunes rank influence: The documented default is 60, with a range of 10–100. It does not set how many results are returned.
- Choose settings by function: Search mode selects retrieval behavior, Top K limits the returned results, and EF Search controls HNSW approximate-nearest-neighbor search effort.
Seahorse Cloud Semantic Search includes Reciprocal Rank Fusion (RRF) in Hybrid mode. It combines the ranked candidate lists from Dense semantic search and Sparse lexical search into one ranking. This combines lexical and semantic retrieval signals without a separate fusion service.
This guide explains the RRF formula and K setting, compares Hybrid, Dense, and Sparse modes, and distinguishes rank fusion from result count and ANN search effort. See the official Semantic Search manual for the current options and defaults.
How built-in RRF combines rankings
For each candidate document, RRF adds its contributions from the ranked lists: score = Σ 1 / (k + rank). A candidate ranked highly in either list can gain a strong combined score, while candidates appearing in both lists receive contributions from both. The candidate with the higher sum ranks first.
In Hybrid mode, RRF K (k in the formula) defaults to 60 and accepts values from 10 to 100. A smaller K increases the relative advantage of top-ranked results; a larger K makes the differences between ranks flatter. K affects the fusion calculation, not the number of returned results.
Illustrative example only—not a benchmark. With K = 60, candidate A is rank 1 in one list and rank 10 in the other; candidate B is rank 3 in both:
| Candidate | RRF score | Result |
|---|---|---|
| A | 1/61 + 1/70 ≈ 0.03068 | Below B |
| B | 1/63 + 1/63 ≈ 0.03175 | B ranks above A |
Choose a search mode
| Mode | What it combines or finds | Useful when |
|---|---|---|
| Hybrid | Combines Dense and Sparse candidate ranks with RRF; default mode | Queries mix natural-language meaning and terms |
| Dense | Semantic similarity | Paraphrases, synonyms, and concept-based questions matter |
| Sparse | Lexical, keyword-based matching | Proper terms, codes, or other distinctive terms matter |
Sparse search is not a guarantee of byte-exact substring lookup; matching behavior can depend on tokenization and configuration. For example, for “ERR-42 retry-policy recovery,” compare Sparse when the identifier and specific terms matter, Dense when users may describe the issue differently, and Hybrid when both signals are useful.
Distinguish the search controls
| Setting | What it controls | Documented default and range |
|---|---|---|
| Top K | Maximum number of results returned | 5; 1–200 |
| RRF K | How Dense and Sparse ranks are fused in Hybrid mode | 60; 10–100 |
| EF Search | HNSW approximate-nearest-neighbor search effort; applies to Hybrid and Dense | 200; 10–1000 |
Increasing Top K can return more results and increase processing work. EF Search controls the approximate-search quality–latency trade-off, while RRF K adjusts how candidate ranks are combined.
Embedding endpoints, vector dimension, M, and EF Construction are creation-time storage and index settings. The Storage creation manual says those settings cannot be edited after creation. Search mode, Top K, EF Search, and RRF K are query-time options in the Semantic Search interface, where each setting applies as documented.
A practical way to compare modes
Choose a small set of representative questions, run them against the same corpus in the relevant modes, and inspect the top results for relevance. Adjust Top K to the number of results your application needs; tune RRF K only when comparing Hybrid rankings. This gives you evidence from your own queries without assuming that one mode or setting wins universally.
How is built-in RRF different from an external reranker?
Built-in RRF combines the positions of Dense and Sparse candidates using 1 / (k + rank). It does not learn a relevance model or perform cross-encoder reranking. An external reranker may still help for a particular application; compare results on representative queries to determine whether it adds value.
How should I choose a mode and Top K for real queries?
Use Hybrid as a starting point when queries combine meaning and specific terms, Dense for paraphrases and conceptual questions, and Sparse for distinctive lexical terms or identifiers. Test a small set of representative queries against the same corpus, inspect the top results, and set Top K to the result count your use case needs.
Compare search modes in your own queries
Try Hybrid, Dense, and Sparse on a representative query set, then inspect the top results. The console is available to configured users.
