User Guide
Image Search
Search images across all data sources by description or visual similarity.
Search modes#
- Text search — describe what you're looking for in natural language. Example: "architecture diagram", "team photo", "product screenshot".
- Image upload — upload a reference image to find visually similar images (reverse image search).
- Combined — use both text and an image together for more precise results.
AI image analysis#
Every indexed image is automatically analyzed by AI to generate rich metadata that improves search accuracy:
- Captions — a one-sentence description of what the image shows, generated by a vision LLM.
- Classification — each image is tagged as photo, chart, screenshot, diagram, illustration, icon, or other.
- Semantic tags — 5–10 tags describing the image content (e.g. "indoor", "meeting", "whiteboard", "architecture").
- OCR text — visible text extracted from screenshots, charts, and diagrams.
This metadata is indexed alongside the visual embeddings, enabling hybrid search: a query like "people in a meeting" matches both by visual similarity and by caption/tag text.
Embedding providers#
| Provider | Best for | Requirements |
|---|---|---|
| OpenCLIP (ViT-B-32) | Local, free, no API keys | ~1 GB RAM for the model |
| Azure AI Vision | Cloud-hosted, higher accuracy, dense captions | Azure subscription, endpoint, API key |
Similarity threshold#
Results below the minimum score threshold are filtered out. The default is 0.25. Adjust it in Settings → AI Assistant → Image Search:
- Lower values (0.10–0.20) — more results, may include less relevant images.
- Higher values (0.30–0.50) — stricter filtering, only highly relevant matches.
Tip
Image search combines visual similarity with caption and tag matching. You can describe what the image looks like ("red bar chart") or what it contains ("Q3 revenue data") — both work.