@volter/twin-cohere
v1.0.0
Published
Cohere Protocol 3 twin: LibreChat web-search rerank v2 with labeled deterministic content.
Readme
@volter/twin-cohere
Protocol 3 Cohere API twin derived from the vendored OpenAPI document. It serves
POST /v2/rerank on api.cohere.com and api.cohere.ai for LibreChat's shipped web-search feature with the end user's own in-app key. Demand records the user flow; the customer life uses the
request body sent by its pinned @librechat/[email protected] adapter. Other API
operations answer the vendor-shaped gap.
The twin runs no model. It ranks Unicode-word overlap with the query, keeps
original document indices and stable ties, and labels that stand-in in
meta.warnings. top_n limits results; document and query budgets truncate
Unicode words, an explicitly labeled approximation of Cohere tokens. Valid
priority values are accepted and bounded; this immediate compute path has
no inference queue or load. Semantic quality, model tokenization, latency and
load-dependent scheduling are not simulated. The documented model catalogue
and release dates gate requests on the World clock.
Bearer credentials are checked through the kernel's held-key mechanism.
POST /_twin/app-credentials stands for dashboard key creation and returns
the same World application key on resume, filling COHERE_API_KEY.
The adapter's COHERE_API_URL points at /v2/rerank. Keys are held by SHA-256;
no vendor resource is stored by reranking, so there is no vendor-backed refresh
or deploy. The live-call rate budget uses the documented trial tier, 10/minute.
Cohere's core job here is an API call; the console provides incidental credential and billing setup, so this pack has no screens. Reference, overview and release requests, plus applicable stated rules, are in doc-examples.json and replayed by the examples journey. The generated score and coverage report the current walks; the final independent judging and conformance belong to the repository procedure.
