@daihum/ai-model-contracts

Pure, serializable model-operation payload contracts for DAIHUM AI. The stable vocabulary leaf: zero dependencies, no behavior, no environment assumptions.

Subpaths: ./chat, ./embedding, ./candidate-rank, ./tools, ./usage, ./embedding-compat. Prefer subpath imports over the root barrel.

What this package is NOT

By rule (boundary gate R1/R7/R8) it carries only model I/O. It must never grow:

• routePlan or any routing/provider-selection structure — that lives in @daihum/ai-routing, which lowers a route plan into resolved providerId/modelId/routeId + a provider call context.
• execution/authority context (workspaceId, actorId, dataClass, …) — that travels in the provider call context (@daihum/ai-provider-contracts), never in a payload.
• provider registries/interfaces/errors — @daihum/ai-provider-contracts.
• secrets, catalog, pricing, cost, audit, config, runtime code, fetch/fs, RAG hits.

Test: these types should be copy-pasteable into another repo as plain TypeScript without bringing behavior, environment, provider knowledge, or product assumptions.

candidate-rank vs rerank

candidate-rank is the model-level ranking op (query + candidate texts → ids/ order/scores), a peer of chat/embedding. The word rerank is reserved for the RAG pipeline stage in @daihum/ai-rag. Candidates are { candidateId, text } only (no metadata backdoor); result score is optional (scoreKind documents its origin).