Hybrid Search Blueprint Series: Semantic Boosting

Erik Hatcher (MongoDB) presents Semantic Boosting — a two-shot hybrid retrieval technique that runs a vector search first, then injects the results as boost clauses into a final lexical search — preserving full-text features like faceting, highlighting, and pagination.

Third article in the series:

1. Survey of the hybrid search landscape
2. Reciprocal Rank Fusion (RRF) and Relative Score Fusion (RSF)
3. This article — Semantic Boosting

---

What Is Semantic Boosting?

A two-phase approach:

1. Vector search phase — run $vectorSearch to retrieve the top-N semantically relevant documents and their similarity scores
2. Inject + boost phase — extract document IDs and scores; add each as an explicit equals boost clause (weighted by score × multiplier) into a final $search lexical query

The final result set is entirely produced by the lexical engine. This gives developers access to all standard text search features — facets, highlights, pagination — on hybrid results.

Why Not RRF or Score Fusion?

• RRF / RSF merge ranked lists from two independent queries; pagination and faceting require extra wiring
• Semantic Boosting makes the lexical engine the single authoritative output; vector influence is injected as score signals, not a separate ranked list

The technique is a special case of signal incorporation — the same pattern used to boost from click/purchase signals, except the signal source is the vector query.

Example (MongoDB Atlas Search)

# 1. Vector search → get IDs + scores
vector_results = collection.aggregate([
    {"$vectorSearch": {
        "queryVector": embed_query(q),
        "index": "vector_index",
        "path": "plot_embedding_voyage_3_large",
        "numCandidates": 150,
        "limit": 10
    }},
    {"$project": {"_id": 1}},
    {"$addFields": {"score": {"$meta": "vectorSearchScore"}}}
])
 
# 2. Build boost clauses from vector results
vector_boost_clauses = [
    {"equals": {"path": "_id", "value": r["_id"],
                "score": {"boost": {"value": r["score"] * 10}}}}
    for r in vector_results
]
 
# 3. Lexical search + injected boosts
should_clauses = [
    {"text": {"query": q, "path": ["title"], "score": {"boost": {"value": 2}}}},
    {"text": {"query": q, "path": ["plot"]}}
] + vector_boost_clauses

The boost multiplier (× 10) is a tuning knob — requires calibration across a query sample.

Embeddings

Uses Voyage AI’s voyage-4-large model (2048 dimensions, dot product similarity) for query embedding at query time. Documents are pre-embedded at index time.

Key Advantages

Feature	RRF/RSF	Semantic Boosting
Faceting	Requires extra logic	Native (lexical engine handles it)
Highlighting	Requires extra logic	Native
Pagination	Requires merging	Native
Score tunability	Limited	Full (boost multiplier per signal)

Related Concepts

• Semantic Boosting — the core technique introduced
• Hybrid Search — broader family this belongs to
• Reciprocal Rank Fusion — alternative fusion strategy
• Relative Score Fusion — alternative score-based fusion
• Results Boosting — Semantic Boosting is a special case of signal boosting
• Dense Vector Retrieval — the first-phase signal source
• Linear Score Combination — another score-based combination approach

Related Articles

• RRF is Not Enough — critiques RRF’s limitations in hybrid contexts
• Hybrid search > sum of its parts? Berlin Buzzwords 2022

People

• Erik Hatcher

Companies

• MongoDB
• Voyage AI