Embedding recipes and data shapes

Planned entries are structurally configured, not qualified or supported. See Model recipe qualification for exact evidence states, preflight gates, and first-wave blockers.

Loss chooser

Data shape	Objective	Distillation meaning
query + positive document	MNRL/InfoNCE	relative in-batch retrieval
query + positive + negative	triplet/contrastive	explicit separation
pair + teacher score	MarginMSE	graded relevance margin
text + teacher vector	cosine/MSE + projection	geometry; projection must be trained on train only
candidate list + teacher order/scores	pairwise/listwise KL	ranking distribution
sentence pair + similarity	CoSENT/cosine	STS
text + class/cluster	supervised classification/clustering	label or grouping structure

Never coerce scores to Boolean silently, truncate arbitrary teacher vectors, or mix evaluation IDs into generation/mining. Hard-negative mining must use train-only corpora, global IDs, deduplication, and false-negative filtering against positives, same-document groups, and known relevance.

Synthetic multilingual queries must record language, generator/provider, model revision, prompt hash, usage/cost, rights, and source document. Provider generation is opt-in and resumable; the local fixture is synthetic and offline.

Model conventions are lock-driven:

• Arctic: exact asymmetric query/document prompts; evaluate 768 and the lock-published 256 dimension separately.
• BGE-M3: dense-only MVP. Sparse and ColBERT/hybrid fusion are later gated work, not implied by dense export.
• Nomic: preserve search_query:, search_document:, classification:, and clustering: prefixes. MoE requires expert/router coverage, utilization and save/reload gates; active parameters are not optimizer memory.
• GTE: trust_remote_code is never implicit. It requires an explicit opt-in to a pinned reviewed commit and a clean offline reload check.
• Qwen: query instruction normally applies only to queries; last-token/EOS pooling, left padding, normalization, and allowed dimensions come from the lock.

For air-gapped operation, pre-stage the wheel, NPM tarball, exact model/tokenizer revisions, schemas, locks, licenses/NOTICE, and datasets; set offline modes in the model stack; verify hashes; and reject any missing cache entry. An offline claim is invalid if installation or reload contacts a registry or model hub.