Reflex

Reflex is a low-latency search API for ranking document strings supplied directly in a request. It combines semantic relevance with a bounded lexical boost, helping exact terms, close spellings, query coverage, and term proximity without allowing lexical evidence to replace the semantic score.

Unlike collection-based RAG, Reflex does not require you to create, populate, and manage a permanent collection before searching. Send the query and the current document strings; Reflex processes them immediately and automatically reuses cached document embeddings when available.

Use Reflex for dynamic or application-owned document sets that must be searchable immediately. Use Semantic Search when documents should live in managed RAG collections, be shared by multiple integrations, or be attached to an AI Gateway.

Request

Parameter	Type	Required	Description
query	string	Yes	Non-empty text used to rank the documents.
documents	string[]	Yes	One to 1,000 document strings to search. Exact duplicate strings are removed before processing.
top_n	number	No	Maximum number of results to return. The default is 5, or all distinct documents when fewer than five are supplied. The maximum is 200 or the number of distinct documents, whichever is lower.
min_score	number	No	Minimum semantic similarity score from 0 to 1, applied before the lexical boost. The default is 0.

The response identifies the request and Reflex model, then returns the ranked results and document-token usage. Each result contains the document text, a relevance_score from 0 to 1, and its zero-based index. Because exact duplicates are removed first, index refers to the deduplicated document sequence.

Cache behavior

Reflex caches document processing automatically within your account. An exact document string can be reused across requests and across different document sets. Changing the text creates a different cache entry, while changing only its position in the request does not.

Cache availability is not permanent. Use the response's usage object to see how the current request was billed:

Field	Meaning
input_tokens	Document tokens processed as cache misses.
cached_input_tokens	Document tokens served from cache.
total_tokens	Total tokens across the distinct documents in the request.

The query is not included in these document-token counters.

Pricing

See Pricing for current Reflex input-token prices and billing details.