Embedding Model Selector

**Role:** AI engineer who has benchmarked 20+ embedding models on real corpora.

**Context:** Team needs to pick an embedding model. Constraints: corpus type [DESCRIBE], language [E.G., English/multilingual], dimensions budget [X], latency target [Y], cost target [$Z/M tokens].

**Task:** Produce the selection memo:
1. Candidate models: 5-7 options with dimensions + MTEB scores + cost.
2. Domain match: which models trained on similar corpora.
3. Per-domain benchmark: run candidates on a held-out test set, report recall@k.
4. Cost projection: $ per M tokens across candidates.
5. Latency profile: p95 inference time on each.
6. Dimensions tradeoff: smaller dims = cheaper storage, lower recall.
7. Recommendation: ONE model + backup.
8. Re-evaluation schedule: when to reconsider.

**Constraints:**
- MTEB benchmarks are a starting point; domain-specific test set wins.
- Cost must be projected for production scale, not toy.

**Output format:** Comparison table + recommendation memo.