OpenAI Embeddings, retrieval, and reranking

Capability boundary

The OpenAI Embeddings, retrieval, and reranking catalog describes the jobs this class of model is intended to solve. It does not mean that every listed model shares request fields, input limits, or output formats. Read each model card together with its lifecycle status, origin, and official source. A platform provider may host models from several original vendors, so both the displayed ID and origin matter.

Selection criteria

Embeddings, retrieval, and reranking occupy different stages of a RAG pipeline. Embedding models map queries and documents into a comparable space, while rerankers rescore an existing candidate set; their scores are not interchangeable. Use a domain query set to measure recall, nDCG, Top-K hits, truncation on long documents, multilingual behavior, and terminology. Pin chunking, normalization, distance metric, and index version so that vector dimension or one attractive example does not become the selection criterion.

Integration and protocol boundary

Confirm the exact model ID, authentication method, endpoint, and request schema in the OpenAI official documentation, then look for the same exact ID and supported endpoint in this site's model marketplace. Changing only the base URL does not guarantee an identical response contract. Streaming events, asynchronous jobs, file uploads, tool calls, and error payloads may require separate adapters.

Production checklist

Build a repeatable evaluation from representative inputs and record the model ID, parameters, input version, and result. Measure success rate, tail latency, throttling, timeout behavior, cost per completed task, and safety outcomes. Prepare a replacement path for preview or legacy models. The official catalog explains product boundaries; callable availability, channel status, and current pricing remain the responsibility of the model marketplace.