4. Comparison: Keyword vs Semantic Search

Overview

Here you compare keyword search and semantic search side by side: advantages, limitations, use cases. You'll understand that they're not mutually exclusive: many systems use both (hybrid search).

---

Full comparison

Aspect	Keyword Search	Semantic Search
What it searches for	Exact words	Meaning (similar vectors)
Synonyms	❌ No (without a dictionary)	✅ Yes (similar embeddings)
Related concepts	❌ No	✅ Yes
Exact words	✅ Yes (guaranteed)	⚠️ Not guaranteed
Speed	✅ Very fast (inverted indexes)	✅ Fast (with HNSW/IVF)
Setup cost	Low (just index the text)	Medium (generate embeddings)
Query cost	Minimal	Medium (embedding the query)
Explainable	✅ Yes (it shows the matched words)	⚠️ Hard (the score is abstract)
Scalability	✅ Millions of docs	✅ Millions of docs

---

Successes and failures

Case 1: A query with synonyms

Query: "car"

Documents:

• Doc A: "The automobile needs gasoline"
• Doc B: "The vehicle has 4 wheels"

Keyword:

• Doc A: ❌ No match (it doesn't contain "car")
• Doc B: ❌ No match

Semantic:

• Doc A: ✅ Match (the embedding of "automobile" is similar to "car")
• Doc B: ✅ Match (the embedding of "vehicle" is similar to "car")

Winner: Semantic search ✅

---

Case 2: A query with an exact code

Query: "error ABC-12345"

Documents:

• Doc A: "Error ABC-12345 in the system"
• Doc B: "A similar error ABC-12346"

Keyword:

• Doc A: ✅ Exact match (it contains "ABC-12345")
• Doc B: ❌ No match (a different code)

Semantic:

• Doc A: ✅ Match (a similar embedding)
• Doc B: ⚠️ It can get a high score (the embeddings of similar codes are close together)

Winner: Keyword search ✅ (exact precision is critical)

---

Case 3: A conceptual query

Query: "how to improve my app's speed"

Documents:

• Doc A: "A performance optimization guide"
• Doc B: "Tips to make your application faster"

Keyword:

• Doc A: ⚠️ Partial match ("speed" doesn't appear)
• Doc B: ⚠️ Partial match ("improve" doesn't appear as such)

Semantic:

• Doc A: ✅ Match (the embeddings of "improve speed" ~ "optimize performance")
• Doc B: ✅ Match (the embeddings of "speed" ~ "faster")

Winner: Semantic search ✅ (it understands synonyms and intent)

---

Decision matrix

Query type	Best method	Example
Exact (codes, names, numbers)	Keyword	"error E404", "article 42"
Conceptual (topics, questions)	Semantic	"How do I optimize?", "animals"
Exploratory (discovering related items)	Semantic	"topics about AI"
Legal/medical (precision is critical)	Keyword + filters	"diagnosis ICD-10 J44"
Ambiguous (multiple meanings)	Hybrid	"bank" (filter by context)

---

Summary

Key points:

• Keyword: Precise for exact matches, doesn't understand synonyms
• Semantic: Understands meaning, doesn't guarantee exact matches
• They aren't mutually exclusive: Combine them in hybrid search
• The decision: It depends on the type of query

---

Next capsule: 05-hybrid-search.md — Combining keyword + semantic.