Semantic Boosting

A two-shot hybrid retrieval technique that executes a vector search first, then injects the resulting document IDs and similarity scores as explicit boost clauses into a final full-text lexical query.

Introduced (named) by Erik Hatcher at MongoDB in the context of MongoDB Atlas Search — the third technique in his hybrid search series, after the landscape survey and the classic fusion mechanisms in Reciprocal Rank Fusion and Relative Score Fusion.

---

How It Works

1. Phase 1 — Vector search: Run $vectorSearch to retrieve the top-N semantically similar candidates with their similarity scores
2. Phase 2 — Score extraction: Extract document IDs and scores; apply a weighting multiplier to determine boost magnitude
3. Phase 3 — Boosted lexical search: Inject each (ID, score) pair as a boost clause into a compound.should array inside the final $search query

The final output is generated entirely by the lexical search engine.

Key Property

Because the final result set comes from the lexical engine, all standard text search features work natively on hybrid results:

• Faceting — no extra merging logic needed
• Highlighting — matched terms annotated by the lexical engine
• Pagination — single result cursor

This distinguishes Semantic Boosting from Reciprocal Rank Fusion and Relative Score Fusion, which merge two separate ranked lists and require additional plumbing for pagination and facets.

It also sidesteps Score Normalization entirely: there is only ever one ranked output, produced by the lexical engine, so there are never two incomparable score scales to reconcile — the vector score enters as a boost magnitude rather than as a competing ranking.

Relationship to Signal Incorporation

Semantic Boosting is a special case of the broader signal incorporation pattern — where external signals (clicks, purchases, query-doc associations) are pre-computed and fed as boost factors into a lexical query. Here, the “signal” is the real-time vector similarity score.

Tuning

The boost multiplier (e.g., score × 10) is the primary tuning knob. Calibration requires:

• Sampling a representative query set
• Comparing metric scores (NDCG, MRR) across multiplier values
• Balancing vector influence against lexical relevance

Limitations

• Requires running two queries sequentially (latency cost)
• The candidate pool from Phase 1 must be large enough — if a relevant document is missed by vector search, it can still be found lexically but won’t receive the semantic boost
• Multiplier tuning is empirical; no principled derivation

Related Concepts

• Hybrid Search — the broader category
• Reciprocal Rank Fusion — alternative fusion; merges two ranked lists
• Relative Score Fusion — alternative fusion; normalizes and combines scores
• Results Boosting — parent technique; Semantic Boosting is a special case
• Dense Vector Retrieval — Phase 1 signal source
• Linear Score Combination — another score-combination approach

Articles

• Hybrid Search Blueprint Series Semantic Boosting — origin article
• Reciprocal Rank Fusion and Relative Score Fusion — the fusion alternatives this technique is positioned against
• Survey of the Hybrid Search Landscape — opens the series

People

• Erik Hatcher — named and described the technique