wiki/public/raw/docs-semantic-search.md

Semantic Search

---

Type: page Date: 2026-01-06

Semantic Search

Semantic search finds content by meaning, not exact words. Ask questions naturally and find conceptually related content.

Press Cmd+K then Tab to switch to Semantic mode. For exact word matching, see Keyword Search.

---

When to use each mode

Use case	Mode
"authentication error" (exact term)	Keyword
"login problems" (conceptual)	Semantic
Find specific code or commands	Keyword
"how do I deploy?" (question)	Semantic
Need matches highlighted on page	Keyword
Not sure of exact terminology	Semantic

How semantic search works

┌─────────────────────────────────────────────────────────────────────────┐
│                     SEMANTIC SEARCH FLOW                                │
└─────────────────────────────────────────────────────────────────────────┘

  ┌──────────────┐    ┌─────────────────┐    ┌──────────────────┐
  │ User query:  │───▶│ OpenAI API      │───▶│ Query embedding  │
  │ "how to      │    │ text-embedding- │    │ [0.12, -0.45,    │
  │  deploy"     │    │ ada-002         │    │  0.78, ...]      │
  └──────────────┘    └─────────────────┘    └────────┬─────────┘
                                                      │
                                                      ▼
                                           ┌─────────────────────┐
                                           │ Convex vectorSearch │
                                           │ Compare to stored   │
                                           │ post/page embeddings│
                                           └──────────┬──────────┘
                                                      │
                                                      ▼
                                           ┌─────────────────────┐
                                           │ Results sorted by   │
                                           │ similarity score    │
                                           │ (0-100%)            │
                                           └─────────────────────┘

1. Your query is converted to a vector (1536 numbers) using OpenAI's embedding model
2. Convex compares this vector to stored embeddings for all posts and pages
3. Results are ranked by similarity score (higher = more similar meaning)
4. Top 15 results returned

Technical comparison

Aspect	Keyword	Semantic
Speed	Instant	~300ms
Cost	Free	~$0.0001/query
Highlighting	Yes	No
API required	No	OpenAI

Configuration

Semantic search requires an OpenAI API key:

npx convex env set OPENAI_API_KEY sk-your-key-here

If the key is not configured:

• Semantic search returns empty results
• Keyword search continues to work normally
• Sync script skips embedding generation

How embeddings are generated

When you run npm run sync:

1. Content syncs to Convex (posts and pages)
2. Script checks for posts/pages without embeddings
3. For each, combines title + content into text
4. Calls OpenAI to generate 1536-dimension embedding
5. Stores embedding in Convex database

Embeddings are generated once per post/page. If content changes, a new embedding is generated on the next sync.

Files involved

File	Purpose
convex/schema.ts	embedding field and vectorIndex on posts/pages
convex/embeddings.ts	Embedding generation actions
convex/embeddingsQueries.ts	Queries for posts/pages without embeddings
convex/semanticSearch.ts	Vector search action
convex/semanticSearchQueries.ts	Queries for hydrating search results
src/components/SearchModal.tsx	Mode toggle (Tab to switch)
scripts/sync-posts.ts	Triggers embedding generation after sync

Limitations

• No highlighting: Semantic search finds meaning, not exact words, so matches can't be highlighted
• API cost: Each search query costs ~$0.0001 (embedding generation)
• Latency: ~300ms vs instant for keyword search (API round-trip)
• Requires OpenAI key: Won't work without OPENAI_API_KEY configured
• Token limit: Content is truncated to ~8000 characters for embedding

Similarity scores

Results show a percentage score (0-100%):

• 90%+: Very similar meaning
• 70-90%: Related content
• 50-70%: Loosely related
• <50%: Weak match (may not be relevant)

Resources

• Convex Vector Search
• OpenAI Embeddings
• Keyword Search - Full-text search documentation