This pilot tests how semantic search can improve access to the collection using locally hosted AI models.
The work builds on an earlier prototype, where the National Museum was an early adopter of AI-based search. In this version, we test a fully local approach using our own models and infrastructure.
The project is developed at the National Museum as part of ongoing work with digital innovation and artificial intelligence in collections, led by Tord Nilsen.
Why we test semantic search
The pilot evaluates whether semantic search provides more relevant access to the collection than traditional keyword-based search.
Traditional search depends on exact terms and predefined metadata. This limits discovery, especially for users who do not know how objects are described internally.
Semantic search focuses on meaning rather than exact wording. It allows users to search in natural language and retrieve results based on similarity.
We test:
- whether relevance improves
- whether search becomes easier to use
- whether local AI models provide sufficient quality
What changed from version 1
This pilot introduces a different technical and strategic approach.
- from external AI services to locally hosted models
- from sending images out to processing them internally
- from dependency on third-party infrastructure to full control
In the previous prototype, images were analysed using external services.
In this version, all processing takes place within the museum’s own infrastructure.
How semantic search works
The system connects images, text, and queries through a shared representation of meaning.
Images are processed with a locally hosted vision–language model, which generates textual descriptions of visible content. These descriptions, together with existing metadata, are converted into vector representations (embeddings).
When a user searches, the query is processed in the same way. The system compares vectors and retrieves results based on similarity.
→ Read more about how semantic search works
Technology and local AI
The system is implemented as a fully local solution.
- locally hosted vision–language model (Qwen 3.5, 9B)
- locally hosted embedding model (BAAI BGE-M3)
- processing of images and metadata within the museum’s infrastructure
- use of the museum’s own API
No external AI services are used.
→ Read more about the technology
Responsible use of AI
The pilot follows defined principles for responsible AI use, including transparency, human control, inclusion, and data security.
All processing is done locally, and images and metadata are not used to train external models.
→ Read more about responsible use of AI
What we test in the pilot
The pilot evaluates:
- relevance of search results
- quality of generated descriptions
- user understanding of how the system works
- performance of local AI models
What we’re NOT testing
Due to the famous scope creep, we’re not testing
- Combination of semantic and metadata search
- onDisplay functionality
- Filtering
What happens next
The pilot is limited in scope and used for evaluation.
Results will inform further development, integration with existing systems, and requirements for quality, governance, and infrastructure.
The aim is to improve access to the collection while maintaining control over data and technology.
