This article explains the hybrid search, which combines data-driven and AI-based search methods.
The hybrid search integrates full-text search, tag search, AI search and visual image analysis (e.g. person detection) in a single search process. This leads to more precise and context-sensitive search results – even if the underlying metadata is not fully maintained.
No additional settings are required to use the new search. By default, each search entry takes into account all available information of an image file. This is divided into two main areas.
Data-driven information:
- Tags (incl. person tags)
- File, folder and album names
- Metadata (technical, descriptive or self-defined)
Visual information:
- visual similarity vectors (for recognising faces, objects, logos, etc.)
- semantic similarity vectors (for interpreting the image content)
All search results are summarised in a central view and sorted according to relevance. The strength of the respective match influences this sorting.
The degree of hybridization can be adjusted as a percentage using a slider if required (right-hand filter bar, visual search section). The default setting is 50/50.
Search behaviour
- For short or unspecific entries (e.g. "machine"), the system delivers a mixture of text-based, semantic and visual hits, which are sorted according to relevance.
- For long formulations (e.g. "press photo of an employee in the laboratory calibrating a sensor"), the semantic component dominates and improves the accuracy of hits.
- When searching for image files without tags or metadata, meaningful hits are still displayed, as the AI works exclusively on the basis of the image content.
- All filters listed here are compatible in combination with filters.
Best Practices
- Search queries can be formulated in full sentences; the AI interprets the context.
- Filters are used to narrow down large numbers of hits.
- The tag search should also be used if special project or event tags are available.
💡 Tip: The similarity search is still recommended for searching for visually similar images using image content (and not using natural language).