How semantic search works
Semantic search relies on embeddings: lists of numbers produced by a language model, in which texts with similar meanings get similar numbers.
- Indexing. For each asset, the system collects its text (title, description, keywords, tags, captions, text read from the image) and converts it into an embedding, which is stored with the asset.
- Querying. The search query is converted into an embedding with the same model.
- Ranking. The system returns the assets whose embeddings are closest to the query's, usually measured by cosine similarity.
Because the comparison works on meaning, "woman working from home" can find assets tagged "remote work" or "laptop on sofa". Results depend on how much descriptive text each asset has. An image with no title, no tags and no caption gives the model nothing to compare, which is why semantic search is usually paired with automatic tagging and OCR.
Semantic search vs keyword search
| Keyword search | Semantic search | |
|---|---|---|
| Matches | The words in the query, with variants ("chairs" finds "chair") | The meaning of the query |
| Best for | Product codes, SKUs, file names, campaign names, exact phrases | Descriptions, concepts and themes ("calm office", "winter at home") |
| Misses | Assets described with other words | Exact identifiers; it can also return loosely related results |
| Ranking | How often and where the words appear (for example BM25) | Distance between embeddings |
Most systems that offer semantic search combine both in a hybrid search: each method produces a ranked list, and the two lists are merged into one. A search for a SKU still returns that product first, and a descriptive query still returns images whose tags use other words. Filters then handle structured conditions such as file type, date or usage rights.
When semantic search helps
- A sales manager needs "engineers on a factory floor" for a deck, and the shoot was tagged "production line, safety helmets".
- An agency asks for "warm autumn product shots", and the library's tags say "orange", "leaves" and "knitwear".
- A regional marketer looks for a partner kit image using their own wording rather than the brand team's internal terms.
- A new colleague searches for a concept before they know the names of the campaigns and product lines.
It helps least when the searcher already knows an exact identifier. For that, keyword search is faster and more precise.
Semantic search in Freedam
- Every search in the asset library runs keyword ranking and semantic similarity together and merges them into one ranked list. There is no mode to switch.
- An asset's embedding is built from its title, description, vocabulary terms, metadata, keywords, AI tags, AI caption, collections, file name, extracted document text and the text read by OCR.
- Automatic tagging adds tags and a caption to every image on upload, so an image nobody described still has text for semantic search to match.
- A query in double quotes runs an exact-phrase search without semantic matching.
- Misspelled queries get a "did you mean" suggestion.
Details: Searching by meaning in the Freedam docs, and the search product page.
Frequently asked questions
What is the difference between semantic search and vector search?
Vector search is the technique: finding the nearest vectors in a database. Semantic search is one use of it, where the vectors are embeddings of text meaning. Vector search is also used for other purposes, such as finding images that look alike (visual search).
Does semantic search replace tagging?
No. It needs text to compare against, and it cannot know your product names, campaigns or rights. Tags, controlled vocabularies and metadata fields still provide that information. Semantic search makes it findable with other words.
Does semantic search work across languages?
It depends on the embedding model. Multilingual models place texts with the same meaning close together whatever their language, so a query in French can match an asset described in English. Results are usually best when queries and descriptions share a language.
Related terms
Search by what you mean
Type a description such as "a cosy living room in winter" into Nordi's library and see photos that match it, whatever words they were tagged with.
You enter Nordi’s workspace as an administrator. No card, no setup.