Collections And Configuration

Agent creates a collection via PUT /collections/products with vectors={size:1536, distance:'Cosine'} to store OpenAI text-embedding-3-small vectors.

Set vectors.size to exactly the embedder's output dimension (1536) and distance to the metric the model was trained for (Cosine for normalized OpenAI embeddings). Both are fixed at create time and immutable — verify the embedder dimension before create rather than guessing.

A product needs both an image embedding and a text embedding per item. Agent configures vectors as a map {image:{size:512,distance:'Cosine'}, text:{size:768,distance:'Cosine'}}.

Define named vectors as a map at create time; every upsert must supply vectors as a dict keyed by the named-vector names; search/query must specify which named vector via the using parameter. Do not mix a default unnamed vector with named vectors in the same collection.

A 50M-point collection exceeds RAM budget. Agent sets vectors.on_disk=true and leaves the HNSW graph in memory.

Set vectors.on_disk=true to store raw vectors via memmap while keeping the HNSW index resident; expect higher tail latency on rescore reads. Pair with quantization always_ram if low-latency first-stage search is required. Verify memory accounting against the documented storage layout.