From content to coordinates
An embedding model converts text, images, audio, products, or users into fixed-length arrays of floating-point numbers. Similar items are placed near one another in that high-dimensional space. The vector database stores those arrays together with stable IDs and metadata.
- The embedding model determines vector dimensions
- Cosine distance compares direction
- Euclidean distance compares geometric separation
- Metadata narrows the candidate set before or during ranking
What happens during a query
The application embeds the query with the same model used for indexed records. The database searches for nearest neighbors, applies metadata filters, and returns IDs, scores, and metadata. Approximate nearest-neighbor techniques trade a small amount of recall for much faster searches over large collections.
Dense, sparse, and hybrid search
Dense vectors capture semantic similarity. Sparse retrieval such as BM25 remains excellent for exact names, identifiers, and rare terms. Hybrid search combines both result lists so a query can benefit from meaning and lexical precision.
Talqora accepts dense values and sparse_text in the same write request, then exposes dense, sparse, and hybrid query modes through one endpoint.
What belongs outside the vector database
Keep original documents, business transactions, and authoritative records in their source systems. Store stable references and the metadata needed for retrieval. A vector database is a search index and retrieval service, not a replacement for every operational database.