Semantic Search

Semantic search finds content based on meaning rather than exact keyword matches. It understands intent, synonyms, and conceptual relationships.

Keyword search: Query: "car repair" Finds: Documents containing "car" AND "repair" Misses: "automobile maintenance", "vehicle fix", "fixing my sedan"

Semantic search: Query: "car repair" Finds: All documents about fixing vehicles, regardless of exact words used Includes: "automobile maintenance", "vehicle service", "mechanic work"

Semantic search answers the question: "What documents are about the same thing as my query?" rather than "What documents contain these words?"

1. Create embeddings Convert all documents in your corpus to vector embeddings—numerical representations that capture meaning.

2. Index vectors Store embeddings in a vector database optimized for similarity search.

3. Process query When a user searches, convert their query to an embedding using the same model.

4. Find similar vectors Search the database for document embeddings most similar to the query embedding.

5. Return results Return the documents corresponding to the most similar vectors.

Why it works: Embedding models learn that similar concepts have similar vectors. "Dog" and "puppy" are close in vector space, so a search for one finds documents about the other.

Keyword search wins when:

• Users search for exact phrases or codes
• Speed is critical at massive scale
• Queries are simple and specific

Semantic search wins when:

• Users describe what they want in natural language
• Vocabulary varies across documents
• Finding conceptually related content matters

Many production systems combine both approaches:

How hybrid search works:

1. Run keyword search (BM25) on the query
2. Run semantic search on the query
3. Combine and rerank results

Why hybrid?

• Keyword search catches exact matches semantic might miss
• Semantic search finds conceptually related content
• Combined results are more comprehensive

Reranking: After combining results, use a more sophisticated model to reorder by relevance. Cross-encoder rerankers consider query and document together for better ranking.

Implementation: Most vector databases support hybrid search:

• Weaviate: Built-in hybrid
• Qdrant: Sparse + dense vectors
• Pinecone: Metadata filtering + semantic

Customer support Find relevant help articles even when customers describe problems differently than documentation.

E-commerce Search products by description: "something to keep coffee hot" finds thermoses, travel mugs, insulated cups.

Document retrieval Find relevant contracts, policies, or reports based on natural language queries.

Code search Find functions by describing what they do: "function that validates email addresses"

Knowledge bases Power internal wikis where employees can ask questions naturally.

RAG systems Retrieve relevant context for LLMs to use in generating responses.

Choose the right embedding model Match model to your domain. General models work well for most cases; specialized models exist for code, legal, medical text.

Chunk documents appropriately Split long documents into meaningful chunks. Too small loses context; too large reduces precision.

Include metadata Store document metadata (title, date, category) for filtering and display.

Set relevance thresholds Don't return results below a similarity threshold. Poor results are worse than no results.

Monitor and iterate Track which searches succeed and fail. Use failures to improve chunking, embedding choices, or add keywords.

Consider latency Semantic search is slower than keyword search. Optimize index settings, use caching, consider approximate nearest neighbor algorithms.

Test with real queries Before deploying, test with actual user queries. Academic benchmarks don't predict real-world performance.