General English RAG with strong default quality, cheap, hosted, Matryoshka-truncatable
Code-symbol-aware retrieval over a large source-code corpus
Cross-lingual retrieval across 30+ languages, including non-Latin scripts, with sparse+dense+multi-vector outputs
Production RAG where the embedding vendor also offers a tightly-paired cross-encoder reranker
On-prem / self-hosted deployment with no API dependency, open weights, strong MTEB
Embedding vendors in 2026 specialize. There is no single best model. The right choice depends on what kind of content you are embedding and what constraints you have on cost, hosting, and language coverage.
For general English text (support tickets, product descriptions, blog posts), OpenAI text-embedding-3-small or 3-large is the standard pick. They are cheap, hosted, score high on the MTEB benchmark, and support Matryoshka truncation.
For source code, voyage-code-3 from Voyage AI is the specialist. It is trained with code-aware tokenization and on a code-heavy corpus, beating general models by 10 to 30 percent on code-retrieval benchmarks.
For multilingual workloads, BGE-M3 is the open-source standard. It covers 100+ languages and produces multiple output types (dense, sparse, and multi-vector) from a single call.
For RAG systems that also need reranking, Cohere bundles embed-v3 with rerank-3 in a way that is easy to plug into one stack.
For on-prem or no-API deployments, Nomic Embed v1.5 and Snowflake Arctic Embed L 2.0 are the open-weights options with strong MTEB scores.
Embedding model selection in 2026 is workload-driven. Five buckets cover most cases.
**General English RAG.** OpenAI text-embedding-3-large is the high-quality default; 3-small is the cost-optimized variant. Both are Matryoshka-truncatable. The text-embedding-3-large at 1024 dimensions is the sweet spot for most teams.
**Code retrieval.** voyage-code-3 from Voyage AI is the specialist. Trained on a code-heavy corpus with code-aware tokenization, it beats general models by 10 to 30 percent on benchmarks like CodeSearchNet. Cursor, Sourcegraph, and Continue all use code-specialist embedders.
**Multilingual retrieval.** BGE-M3 from BAAI is the open-weights leader, covering 100+ languages with shared embedding space across languages. It also produces dense, sparse, and multi-vector outputs from one call, useful for hybrid retrieval.
**Paired reranker.** Cohere ships embed-v3 (or v4) with rerank-3, a cross-encoder tuned to the same domain. Picking both from one vendor simplifies the stack and gives you a fast first-stage plus a precision second-stage.
On-prem / open weights. Nomic Embed v1.5 and Snowflake Arctic Embed L 2.0 are the strong open-weights choices for self-hosted deployments. Both rival OpenAI on MTEB and run on a single H100.
Quick reference
TL;DR
Embedding vendor choice is workload-driven: OpenAI 3-large for general English, Voyage voyage-code-3 for code, BGE-M3 for multilingual, Cohere for paired reranker, Nomic or Arctic for open-weights on-prem.
Picture a hardware store with five aisles, each selling a different kind of saw. There is the all-purpose handsaw most people grab without thinking. There is a tile saw for cutting ceramic, a multi-tool that works in many materials, a power saw that comes paired with a matching sander as a bundle, and a hand-built saw you can buy as a kit and assemble yourself at home. Picking the right search tool is the same exercise. There is a general-purpose option, a code specialist, a multi-language specialist, a paired-ranker option, and a self-assemble kit you can run in your own basement.
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.
Tiny corpus under 10k documents where any modern embedder is fast enough and the differences are imperceptible Selection effort outweighs the quality gain at small scale; pick a default and move on
Single-language English-only support workload with no code content Specialists buy nothing here; OpenAI text-embedding-3-large is the right starting point
Strict zero-API regulatory environment where managed APIs are forbidden Hosted vendors are eliminated by policy; the choice narrows to open-weights regardless of quality preference
Pure semantic-similarity tasks like clustering where any reasonably modern embedder works Retrieval-quality differences mostly manifest in ranking; clustering tolerates weaker embedders
Quick scan · 2–4 min
Everything important, quickly.
7 minutes: the five workload buckets, vendor specializations, MTEB vs labeled eval, switching cost, and decision rule.
Side-by-side
Vendor / Model
Best for
Hosting
Standout feature
OpenAI text-embedding-3-large
general English RAG
hosted API
Matryoshka truncation, cheap
Voyage voyage-code-3
code retrieval
hosted API
code-aware tokenizer, +10-30% on CSN
BGE-M3
multilingual, hybrid
open weights
dense + sparse + multi-vector from one call
Cohere embed-v3 / v4 + rerank-3.5
paired-reranker stacks
hosted API
tight embedder-reranker pairing
Nomic v1.5 / Arctic Embed L 2.0
on-prem, no API
open weights
single-H100 deployable, strong MTEB
Real-world examples
Real products, models, and research that use this idea.
1Cursor and Continue use code-specialized embedding models (voyage-code-3 or jina-code) for repository-level retrieval over user codebases.
2Notion AI and Linear AI rely on OpenAI text-embedding-3-large for general English content retrieval.
3Many enterprise RAG deployments running on Databricks or Snowflake pick BGE-M3 or Arctic Embed for in-warehouse embedding without API egress.
4Cohere is the canonical vendor for stacks that pair an embedder with a tightly-coupled cross-encoder reranker like rerank-3.5.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhen does it make sense to fine-tune an open-weights embedder rather than pick a specialist vendor?
Sign in or go Pro to see the approach for this follow-up.
QHow do you safely evaluate two embedding vendors against your real workload?
Sign in or go Pro to see the approach for this follow-up.
QWhat is the operational cost of running BGE-M3 self-hosted versus paying OpenAI?
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
Picking text-embedding-3-large for everything because it is the default, then losing 10 to 30 percent recall on a code-retrieval or multilingual workload that needed a specialist.