Drag each answer to line up with its matching prompt
Multilingual legal corpus spanning English, French, Spanish, and Mandarin documents
Per call latency and hosted vs self hosted throughput; self-hosted small models or local ONNX quantized models beat any API on tight latency
Highly specialized biomedical corpus with technical jargon not seen in general training data
Domain specific fine-tuned embeddings (BioBERT derived, MedCPT) or fine-tuning a base model on labeled in domain pairs
Corpus of long form legal contracts where individual clauses run 4K+ tokens
Max input length of the embedding model (Cohere Embed v4 and Voyage Embed v3 support 32K+ tokens; many older models cap at 512 or 8K)
Mobile / on device retrieval where the index has to fit in <500MB of RAM
Multilingual coverage benchmarks (Voyage Embed v3 multilingual, Cohere Embed v4 multilingual, BGE-M3) over English only models
Hard P99 latency target of 30ms per embedding call at 10K QPS
Strongest off the shelf model on MTEB (e.g. Voyage Embed v3, text-embedding-3-large, Cohere Embed v4) as the safe default, replaced after building an eval set
Greenfield English language enterprise search with no labeled eval data yet and a tight launch deadline
Embedding dimensionality (smaller dims = smaller storage; e.g. 256-dim or quantized 3072-dim from text-embedding-3-large)
No universal best embedding model. Each deployment has one dominant constraint: language coverage, domain fit, input length, dimension, latency, or off the shelf strength. Pick the model that resolves it.
Imagine choosing a translator for a job. If the documents are in four languages, you hire someone who actually speaks all four, not the smartest English-only translator in town. If the documents are medical reports full of jargon, you hire someone trained in medicine, not a generalist with more credentials. If each document is the size of a novel, you need someone with the stamina to read the whole thing before summarizing, not someone who can only handle short notes. If you need ten thousand translations per second, you build a small fast team rather than calling one expensive consultant. Picking the right text to numbers model is the same. Each scenario has one constraint that dominates everything else, and the trick is naming that constraint before evaluating candidates.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Embedding model selection is one of the most asked and most mis answered RAG interview questions. The mistake is treating it as a leaderboard problem: look up the top MTEB model, pick it, move on. That works exactly when your corpus matches the benchmark's assumptions, English, general purpose, short documents, and breaks the moment any assumption shifts.
This deep dive frames the choice as constraint resolution. Six common constraints, six distinct dominant signals, six different correct models. By the end you should be able to listen to a scenario description and name the constraint before naming the model.
Multilingual coverage as the dominant constraint
If your corpus contains documents in more than one language, language coverage trumps everything else. An English-only model trained on Common Crawl English will produce embeddings whose French and Mandarin regions are sparse and noisy. Two documents that say the same thing in different languages will land in different vector regions, so French queries fail to retrieve Spanish documents.
Multilingual models train on parallel corpora and use techniques like contrastive alignment across languages to produce a shared embedding space. The same sentence translated into ten languages lands near the same point. Cross-lingual retrieval works because the geometry was explicitly designed for it.
The 2026 references are Voyage Embed v3 multilingual, Cohere Embed v4 multilingual, and BGE-M3. All three benchmark within a few points of each other on cross-lingual retrieval and significantly above English-only models on multilingual corpora. BGE-M3 is the open weight option for teams that need to self-host.
The trap is multilingual models that fail closed: a model labeled multilingual that supports 20 languages well and your 21st poorly. Always test on at least one query per language before deciding.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
# Six-constraint selector for embedding models in a RAG project.
# Name the dominant constraint, then pick.
def pick_embedding_model(corpus_profile, ops_profile):
langs = corpus_profile.languages # e.g. {'en', 'fr', 'es', 'zh'}
domain = corpus_profile.domain # 'general' | 'biomed' | 'legal' | 'code'
max_tokens = corpus_profile.max_doc_tokens
dim_budget = ops_profile.max_dim # None or int
p99_ms = ops_profile.p99_latency_ms
if len(langs) > 1:
return "voyage-embed-v3-multilingual" # or Cohere Embed v4 multilingual, BGE-M3
if domain in ("biomed", "legal", "finance"):
return f"fine-tuned-{domain}-embed" # or domain pretrained
if max_tokens > 8000:
return "voyage-embed-v3" # 32K input; Cohere Embed v4 alt
if dim_budget and dim_budget < 512:
return "text-embedding-3-large@256" # Matryoshka truncate
if p99_ms < 50:
return "self hosted bge small onnx" # local, no network hop
return "text-embedding-3-large" # safe MTEB default| Constraint | Primary check | Reference models (2026) |
|---|---|---|
| Multilingual coverage | Trained on the target languages with shared space | Voyage Embed v3 multilingual, Cohere Embed v4, BGE-M3 |
| Specialized domain | Domain pretraining or in-domain fine-tuning | BioBERT-derived, MedCPT, Voyage-finance, fine-tuned base |
| Long inputs (4K+ tokens) | Max input length and truncation behavior | Voyage Embed v3 (32K), Cohere Embed v4 (32K) |
| On-device / tight RAM | Vector dimension and quantization support | text-embedding-3-large@256 (Matryoshka), BGE-small |
| Tight P99 latency | Self-host versus API; ONNX or TensorRT | BGE-small-en ONNX, gte-small local, MiniLM quantized |
| Greenfield English default | MTEB rank as a placeholder until eval exists | Voyage Embed v3, text-embedding-3-large, Cohere Embed v4 |
Real products, models, and research that use this idea.
- Cohere Embed v4 (multilingual, 32K input) and Voyage Embed v3 (multilingual variant, 32K input) are the 2026 references for cross-lingual retrieval in enterprise RAG.
- OpenAI text-embedding-3-large supports Matryoshka representation learning: a 3072-dim vector can be truncated to 256 dim with graceful quality loss, ideal for on-device indexes.
What an interviewer would ask next. Try answering before peeking at the approach.
QYou ran MTEB and your in-house eval and they disagree on the top model. Which do you trust?
Always trust your in-house eval if the labels are clean and the workload is representative. MTEB is a leaderboard, your eval is your production. Investigate the divergence to learn whether MTEB is leaking corpus characteristics that don't match yours.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Picking the top MTEB model without checking whether your corpus matches the benchmark's languages, domain, or input length distribution. MTEB rank is a default, not a decision.
60 second bullets to scan on the way to the call.
Naming the dominant constraint for a given RAG deployment
Why MTEB rank is a default, not a decision rule
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.