Embeddings Are a Hidden Runtime Dependency

Companies have spent time and money encoding text into vectors and storing them for search, matching, and feature engineering. When a provider shutters a model or its semantics shift, stored vectors can’t be refreshed or compared on an equivalent basis. Teams then face broken similarity queries, inconsistent rankings, and unverifiable compliance records. Losing the ability to recompute vectors renders prior investments functionally useless because only fresh, comparable vectors keep embeddings meaningful. As corpus size and workflow integration grow, lock-in and operational fallout from a model shutdown increase sharply.

Vectors encode model-specific geometry and training priors; they are meaningful only relative to their generator. That makes embeddings a runtime dependency: live operations—search, ranking, deduplication, personalized feeds, compliance audits—depend on producing comparable vectors. When the generator vanishes or drifts, you lose the ability to validate, update, or rerun core logic. The consequences are concrete: degraded search relevance, broken changelogs for model-driven features, audit gaps for regulated data, and costly reindexing or reannotation. This is vendor lock-in in disguise — a computational contract, not just stored bytes. Engineering teams must version embeddings, archive models or checkpoints, and design for re-embeddability, or accept that past work can become stranded capital and operational risk.

Embeddings underpin long-context retrieval and knowledge-aging mitigation for LLMs. As providers iterate, embedding geometry shifts: different models encode different notions of semantic similarity and reflect different training corpora. That technical reality is why Willison’s warning targets infrastructure and compliance teams.

Who will guarantee long-term availability of embedding models or provide deterministic re-embeddings? Will standards emerge for embedding versioning, serialization, and provenance? Watch provider deprecation policies, SLAs for model persistence, and tooling for storing model checkpoints alongside vectors.