Embeddings
How can different words express a similar idea?
An embedding model represents an input as a vector of numbers. A retrieval system compares query and document vectors to find useful semantic matches, including passages without the exact query words. The representation depends on the model and its training; similarity does not establish that a passage is correct.
HotelGuesthouseApartmentRestaurantCaféBakery
Hotel, guesthouse and apartment share meaning with a place to stay.
A 2D sketch of similarity, not real embedding coordinates.
What to understand
- Use compatible query and document embeddings, including the model's required query or document input settings.
- Evaluate the languages, document types, and specialist vocabulary in your own corpus before choosing a model.
- Treat model and dimension changes as index changes. Compatibility must be documented; equal vector length alone is not enough.
Go to the source
Primary documentation for the ideas in this explainer.
Follow the next part of the system.
Keyword, vector & hybrid search