A dense vector is a numerical representation of meaning produced by an embedding model. In semantic retrieval, document chunks and user queries are encoded into the same vector space so the system can find nearby matches. section 4.0
For retrieval systems, keep these implementation details in view:
- Embedding quality affects whether relevant queries and documents align; weak domain fit can return unrelated passages. section 4.0
- Match the vector database index dimension to the embedding model output; health-care embeddings are typically 768 dimensions in the example. section 4.5.4
Health-care policy search
A policy assistant encodes policy chunks and a question such as “Is physical therapy covered?” as vectors, then uses similarity search to retrieve relevant passages. section 4.0 section 4.5.4
