5. Hybrid Search: The Best of Both Worlds

Overview

Hybrid search combines keyword search (precision on exact matches) with semantic search (understanding of meaning). It's the approach used by modern production systems (Elasticsearch + embeddings, Weaviate hybrid, Pinecone with filters).

---

Hybrid search strategies

Strategy 1: Filter + Semantic

1. Filter with keyword (e.g. documents that mention "Python")
2. Rank with semantic (order by vector similarity)

Example:

Query: "Python best practices"

Step 1 (keyword): Filter the docs that contain "Python"
→ 1000 documents found

Step 2 (semantic): Rank those 1000 by cosine
→ The top-10 most relevant to "best practices"

The advantage: It guarantees that "Python" appears (keyword) while still finding the most relevant results (semantic).

---

Strategy 2: Rank fusion

1. Run keyword search → Ranking A (top-100)
2. Run semantic search → Ranking B (top-100)
3. Fuse the two rankings (e.g. RRF - Reciprocal Rank Fusion)

RRF (Reciprocal Rank Fusion):

score(doc) = α/rank_keyword + β/rank_semantic

Example:
Doc X: rank_keyword=5, rank_semantic=3
→ score = 1/5 + 1/3 ≈ 0.53

Doc Y: rank_keyword=1, rank_semantic=50
→ score = 1/1 + 1/50 ≈ 1.02

Doc Y has the higher final score → it appears first

The advantage: Documents that appear in both rankings (keyword AND semantic) get a boost.

---

Strategy 3: Cascade

1. Try keyword search
   → If there are enough results (e.g. > 10) → return them
   → If there are few results (< 10) → continue

2. Try semantic search
   → Supplement with the semantic results

The advantage: Efficient (keyword is faster). It uses semantic only when necessary.

---

Weaviate hybrid search (a real example)

Query: "machine learning"

Configuration:
- alpha = 0.5 (a 50% keyword, 50% semantic balance)
- alpha = 0.0 (100% keyword)
- alpha = 1.0 (100% semantic)

The internal algorithm:

final_score = alpha × semantic_score + (1-alpha) × keyword_score

Typical use: alpha=0.5 (balanced) or alpha=0.7 (favoring semantic).

---

Elasticsearch + embeddings (a real example)

Setup:

1. Index the normal text (keyword with BM25)
2. Add an "embedding" field (a 1536D vector)
3. On a query:
   - Run keyword search (BM25)
   - Run vector search (kNN)
   - Combine the scores (a weighted sum)

---

When to use hybrid search

Use hybrid when:

• ✅ Queries are mixed (some exact, others conceptual)
• ✅ You need precision (keyword) AND coverage (semantic)
• ✅ You have the resources for both methods
• ✅ Users expect something "Google-like" (it understands intent but respects exact matches)

Examples:

• A technical documentation search engine
• E-commerce (an exact product name vs a conceptual description)
• Support systems (exact ticket IDs vs descriptions)

---

Hybrid trade-offs

Advantages:

• ✅ The best of both worlds
• ✅ More robust (if one fails, the other compensates)

Disadvantages:

• ❌ Greater complexity (maintaining two indexes)
• ❌ Higher cost (embeddings + an inverted index)
• ❌ Tuning the weights (alpha) requires experimentation

---

Real implementations

Weaviate:

• Native support for hybrid search
• An alpha parameter, tunable per query

Elasticsearch:

• Keyword (BM25) + kNN search (embeddings)
• A script to combine the scores

Pinecone:

• Vector search + metadata filtering (pseudo-hybrid)
• Keyword filters before the vector search

Custom:

• Run keyword (Elasticsearch) and semantic (Pinecone) in parallel
• Fuse them in the application (RRF)

---

Summary

Key points:

• Hybrid: Keyword + semantic
• Strategies: Filter+semantic, rank fusion, cascade
• Advantages: Precision + coverage
• Disadvantages: More complexity, higher cost
• In production: Weaviate, Elasticsearch, custom fusion

---

Next capsule: 06-when-to-use-each.md — Decisions by use case.