Semantic Search

Semantic Search uses AI embeddings to find similar transcriptions by meaning, not just keywords. It's available in the History window alongside text-based search.

How It Works

1. Transcriptions are embedded in batches using Gemini's gemini-embedding-001 model
2. Each transcription becomes a 768-dimensional vector stored locally in MongoDB
3. When you search, your query is also embedded
4. Results are ranked by cosine similarity — how closely the meaning matches

This means searching for "project deadline" can find a transcription that says "we need to finish by Friday" even though the words don't overlap.

Using Semantic Search

1. Open the History window (from the History tab or menu)
2. Switch to the Semantic Search tab
3. Type a search query describing what you're looking for
4. Optionally filter by date range
5. Results appear ranked by similarity score (0-100%)

Similarity Scores

• 80-100%: Very strong match — the transcription closely relates to your query
• 60-80%: Good match — relevant content
• 40-60%: Moderate match — loosely related
• Below 40%: Weak match — may not be relevant

Embedding Coverage

The Search tab shows embedding coverage: e.g., "Embeddings: 150 / 200 (75% coverage)"

Embeddings are generated in the background when you accumulate 100+ unembedded transcripts. This batch approach:

• Minimizes API calls
• Doesn't block the UI
• Processes oldest transcripts first

Configuration

Configure in Settings:

Setting	Default	Description
embedding_enabled	True	Enable/disable semantic search
embedding_model	gemini-embedding-001	Embedding model
embedding_dimensions	768	Vector size
embedding_batch_size	100	Transcripts per batch

Cost

Gemini's gemini-embedding-001 model is free (1,500 requests/minute). Semantic search adds no cost to your usage.