A team currently runs OpenAI text-embedding-3-small with a relevance threshold of cosine ≥ 0.78. They want to switch to Voyage-3 for quality gains. Explain why 0.78 is meaningless on the new model and what they need to re-derive.
Free · 2 AI evals / day
Cosine similarity measures the angle between two vectors. The vectors come from an embedding model, and every embedding model has its own way of arranging meaning in space. Two models can both output 1536-dimensional vectors and still place 'dog' and 'puppy' at different absolute cosines, because they spread their space differently.
So when the team trained on text-embedding-3-small and landed on 0.78 as the cutoff that gave them their target precision and recall, that 0.78 is a property of THAT model's distribution. On Voyage-3 the same labeled pairs will sit at different absolute cosine numbers. The cutoff that gave 0.78 on the old model could mean anything on the new one.
The practical step is to re-derive. Collect labeled (query, doc, relevant?) pairs from production logs, embed both sides with Voyage-3, sweep the cosine cutoff, and pick the value that hits the precision and recall targets. An alternative is to drop the threshold entirely and use top-K plus a reranker, which is model-agnostic because it depends on rank order, not absolute score.
Cosine similarity is a measurement inside one model's vector space. Each embedding model is a different learned function and induces a different score distribution. The typical cosine between random pairs might be 0.6 on text-embedding-3-small and 0.4 on Voyage-3 just because of how each model was trained. So 0.78 on the old model and 0.78 on the new one are not the same evidence of relevance.
The team has three options. First, re-derive the cutoff. Collect a labeled set of (query, doc, relevant?) tuples that reflects production traffic, embed both sides with Voyage-3, sweep the cosine cutoff against precision and recall. Pick the operating point.
Second, switch to rank-based filtering. Top-K retrieval is model-agnostic because it uses ordering, not absolute score, and is the dominant production pattern at scale.
Third, calibrate. Fit an isotonic regression or Platt scaler on a labeled set that maps raw Voyage-3 cosine onto a probability of relevance. Now the score is comparable to whatever your downstream uses, and re-fit only when the model changes again.
Quick reference
TL;DR
Cosine 0.78 is a per-model artifact. Switching to Voyage-3 means the score distribution moves, so the team must re-derive the cutoff against a labeled set or move to a rank-based approach.
Think of two photographers grading photo brightness on the same zero to one scale. One shoots outdoors, so most of his photos read 0.7 to 0.9. The other shoots indoors, so hers read 0.3 to 0.6. If both call a photo a 0.78, they mean wildly different things. Switching photographers means relearning what 0.78 implies on the new style. Search-similarity scores behave like this. Each search tool is a different photographer of meaning. To keep using a fixed cutoff you have to re-measure what that number now corresponds to on the new one, or just pick the brightest three photos per shoot and stop arguing about exact numbers.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
A fixed cutoff filter survives an embedding-model migration without any recalibration step The new model's score distribution will silently shift the operating point and break precision or recall.
Hot-swapping vendors via the unversioned alias in production Vendors sometimes pin new snapshots under the same alias; the cutoff will drift without warning.
Treating MTEB rank as a proxy for absolute cosine on your corpus MTEB measures rank-order quality on benchmark corpora; absolute cosine on your data is decoupled.
Quick scan · 2–4 min
Everything important, quickly.
3 min: cosine as per-model geometry + score distribution shift + labeled-set cutoff sweep + rank-based and calibration escape hatches + production detection signals.
Real-world examples
Real products, models, and research that use this idea.
1Voyage AI's voyage-3 release notes (2026) explicitly state that cosine values are not comparable to voyage-2 and recommend sweeping the cutoff on a labeled set before reusing pipelines.
2Cohere's embed-v4 migration guide pairs cosine retrieval with Rerank v3 specifically so the cross-encoder absorbs the calibration question.
3OpenAI's 2026 embedding docs warn that absolute cosine is a model-specific quantity and recommend dated snapshot aliases for stability.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow large should the labeled set be for a credible threshold sweep?
Sign in or go Pro to see the approach for this follow-up.
QWhat if MTEB shows Voyage-3 beats text-embedding-3-small but your retrieval gets worse?
Sign in or go Pro to see the approach for this follow-up.
QWhen would you skip a cutoff and go top-K plus reranker?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Lifting a tuned cutoff like 0.78 directly to the new vendor without sweeping it against a labeled set on the new model's score distribution.