Semantic search

Press / anywhere in the inbox to open semantic search.

Unlike Gmail keyword search, semantic search finds threads by meaning — e.g. searching "contract renewal" can surface emails that discuss extending an agreement without using those exact words.

How it works

1. Your query is embedded with the active AI provider's embedding model
2. pgvector compares your query vector to stored thread embeddings
3. Results rank by cosine similarity
4. Only threads you have classified (backfill complete + webhooks) appear

Using semantic search

1. Press /
2. Type a natural-language query (1–500 characters)
3. Results show subject, sender, snippet, lane, priority, and relevance score
4. Click a result to open the thread

Results return empty if:

• Backfill has not completed yet
• No threads match above the similarity threshold
• You switched embedding provider and re-embed is still running

Provider switcher interaction

Embeddings are provider-specific. Each classification row stores which provider created its vector (openai or gemini).

When you switch providers in the header:

1. A re-embed job starts in the background
2. An activity banner shows progress
3. Semantic search uses the new provider's embeddings after re-embed completes

You can also trigger re-embed via POST /api/inbox/reembed (see API routes).

Semantic vs advanced search

Use both — they complement each other.

Tips

• Keep queries focused: "Q3 board meeting reschedule" beats "meeting"
• Check you're on the provider you embedded with (see header switcher)
• If results seem stale after new mail, wait a few seconds for webhook classification

Related

• Advanced search
• AI configuration