Free tool · Runs in your browser

Which embedding model should you use?
Seven questions, a sourced shortlist.

Start with OpenAI text-embedding-3-small or voyage-4-lite at $0.02 per million tokens, move to voyage-4-large when retrieval quality is the bottleneck, use a code or multimodal model when that is what you search, and self-host BGE-M3 or Qwen3-Embedding when the data must stay on your servers.

The picker turns your answers into a ranked shortlist of two or three of the 14 models it covers, with the cost math, vector storage, a reranker and a 50-query test to confirm it. Every price and spec links to the vendor's own page, checked on 17 November 2026. Nothing you enter leaves your browser.

Answer the questions

Seven questions. The shortlist below updates as you answer, and your answers stay on this page.

1. What languages is your content in?
2. What are you searching?
3. Search queries per month
Who waits for them?
4. How big is the corpus?

40M tokens, about 100K chunks of 400 tokens.

Convert documents to tokens

2,000 x 10 pages x 800 tokens a page (OpenAI's rule of thumb) = 16,000,000 tokens.

New or changed each month
5. What matters more?
6. Where may the data go?
7. What do you already run on?

Your shortlist

Ranked by the scoring below. Costs are estimates from list prices checked on 17 November 2026.

  1. 1voyage-4-large

    Score 9.3

    Voyage AI (MongoDB) · voyage-4-large · queries via voyage-4-lite

    Start here: the best retrieval quality for general text, and queries can go through the cheaper voyage-4-lite with no reindex.

    Estimated cost

    Corpus, once40,000,000 tokens x $0.12 / 1M$4.80
    Changed docs, monthly4,000,000 tokens (10%) x $0.12 / 1M$0.48
    Queries via voyage-4-lite, monthly100,000 queries x 50 tokens x $0.02 / 1M$0.10

    $4.80 once, then $0.58 a month.

    • Through the batch API the one-off corpus run is $3.22 ($0.0804 / 1M).
    • First 200 million tokens free, which this estimate ignores.

    Storage for 100,000 vectors

    1,024 dims, float324,096 B each409.6 MB
    1,024 dims, int81,024 B each102.4 MB
    1,024 dims, binary128 B each12.8 MB

    Vectors x dimensions x bytes per dimension, before index overhead. The model returns int8 and binary vectors itself; binary is usually paired with a float rescore of the top hits.

    Context 32,000 tokens. Sizes: 256, 512, 1024, 2048.

    How the score adds up
    • Retrieval quality for general text (9.5/10)+9.5
    • Query latency: queries go through voyage-4-lite+0.4
    • Price: $0.12 per million tokens, 6.0x the cheapest API-0.6

    Quality basis: First on the RTEB retrieval board, ahead of Cohere Embed v4 and OpenAI 3-large in Voyage's own tests.

    Sources, checked 17 November 2026: Voyage text embedding docs, Voyage pricing, Voyage 4 announcement, MongoDB Voyage AI geographies, MongoDB Voyage AI model lifecycle, Voyage AI on AWS Marketplace.

  2. 2Cohere Embed 5 Pro

    Score 8.3

    Cohere · embed-v5.0-pro · queries via embed-v5.0-fast

    Strong alternative: a 128K context and image input, launched 30 September 2026, so independent results are still thin, and queries can go through the cheaper Cohere Embed 5 Fast with no reindex.

    Estimated cost

    Corpus, once40,000,000 tokens x $0.12 / 1M$4.80
    Changed docs, monthly4,000,000 tokens (10%) x $0.12 / 1M$0.48
    Queries via Cohere Embed 5 Fast, monthly100,000 queries x 50 tokens x $0.08 / 1M$0.40

    $4.80 once, then $0.88 a month.

    Storage for 100,000 vectors

    2,048 dims, float328,192 B each819.2 MB
    1,024 dims, int81,024 B each102.4 MB
    2,048 dims, binary256 B each25.6 MB

    Vectors x dimensions x bytes per dimension, before index overhead. The model returns int8 and binary vectors itself; binary is usually paired with a float rescore of the top hits.

    Context 128,000 tokens. Sizes: 256, 512, 768, 1024, 1536, 2048.

    How the score adds up
    • Retrieval quality for general text (8.5/10)+8.5
    • Query latency: queries go through Cohere Embed 5 Fast+0.4
    • Price: $0.12 per million tokens, 6.0x the cheapest API-0.6

    Quality basis: A 128K context and image input, launched 30 September 2026, so independent results are still thin.

    Sources, checked 17 November 2026: Cohere Embed 5 announcement, Cohere Embed docs, Microsoft Foundry models sold by Azure.

  3. 3Gemini Embedding 2

    Score 6.8

    Google · gemini-embedding-2

    Worth testing: successor to gemini-embedding-001, an MTEB top-five model, with image, video, audio and PDF input added.

    Estimated cost

    Corpus, once40,000,000 tokens x $0.20 / 1M$8.00
    Changed docs, monthly4,000,000 tokens (10%) x $0.20 / 1M$0.80
    Queries, monthly100,000 queries x 50 tokens x $0.20 / 1M$1.00

    $8.00 once, then $1.80 a month.

    • Through the batch API the one-off corpus run is $4.00 ($0.10 / 1M).
    • Free tier on the standard API, which this estimate ignores.

    Storage for 100,000 vectors

    3,072 dims, float3212,288 B each1.2 GB
    1,024 dims, float324,096 B each409.6 MB
    256 dims, float321,024 B each102.4 MB

    Vectors x dimensions x bytes per dimension, before index overhead. The API returns floats at the size you ask for; any int8 or binary step happens in your vector database.

    Context 8,192 tokens. Any size from 128 to 3072.

    How the score adds up
    • Retrieval quality for general text (8/10)+8.0
    • Query latency: large model on every query-0.4
    • Price: $0.20 per million tokens, 10x the cheapest API-0.8

    Quality basis: Successor to gemini-embedding-001, an MTEB top-five model, with image, video, audio and PDF input added.

    Sources, checked 17 November 2026: Gemini API embeddings docs, Gemini API pricing.

Reranker: add one

Voyage rerank-3-lite

Retrieve the top 50 with embeddings and rerank them down to the 5 to 10 chunks the prompt sees. It adds one model call to every query, so measure p95 latency with it switched on.

Voyage's reranker for latency-sensitive applications, served from its EU endpoint too. $0.02 per million tokens, first 200 million free. Context: 32,000 tokens. Alternative: Voyage rerank-3.

100,000 queries x (50 x 50 + 50 x 400) = 22,500 tokens each x $0.02 / 1M = $45.00 a month at list price.

Sources, checked 17 November 2026: Voyage reranker docs, Voyage pricing, MongoDB Voyage AI model lifecycle.

Full ranking (12 models scored)
Every eligible model and its score
ModelScoreOnceMonthly
voyage-4-large9.3$4.80$0.58
voyage-48.5$2.40$0.34
Cohere Embed 5 Pro8.3$4.80$0.88
voyage-4-lite7.9$0.80$0.18
Cohere Embed 5 Fast7.7$3.20$0.72
Gemini Embedding 26.8$8.00$1.80
OpenAI text-embedding-3-small6.4$0.80$0.18
OpenAI text-embedding-3-large6.3$5.20$1.17
Qwen3-Embedding-8B3.4$0$800
Qwen3-Embedding-0.6B2.7$0$400
voyage-4-nano2.7$0$400
BGE-M32.2$0$400

Confirm it on your own data: 50 queries

The ranking is a starting point. This afternoon-sized evaluation is what decides.

  1. 1Collect 50 real queriesPull them from search logs, support tickets or chat transcripts. Mix full questions, keyword fragments and typos the way people really type them.
  2. 2Label the right chunksMark the one to three chunks that answer each query. An LLM can propose labels; a person who knows the content approves them.
  3. 3Embed the same chunks with voyage-4-large, Cohere Embed 5 Pro, and Gemini Embedding 2Same chunking (about 400 tokens), same distance metric, same top-k. Store the model name and version with every vector so a later switch is a cutover, not an archaeology project.
  4. 4Score recall@10 and nDCG@10Recall@10 asks whether a right chunk made the shortlist at all; nDCG@10 also rewards ranking it near the top, which is what the prompt sees.
  5. 5Add Voyage rerank-3-lite on top of voyage-4-largeRerank the top 50 and score again. If the cheapest model plus a reranker matches the best model alone, that is usually the better buy.
  6. 6Time and price 1,000 queriesMeasure p95 latency from the region you deploy in, with the reranker on and off, and the cost per 1,000 queries. Pick the cheapest configuration within two points of the best recall@10.

Want a second pair of eyes before you commit to a reindex? That is part of AI integration, and the reasoning behind these picks is in the embedding models comparison.

What the picker assumes

Prices are list prices per million tokens, with no volume discounts. A page is about 800 tokens (OpenAI's own rule of thumb), a chunk 400 tokens, a query 50 tokens, and a reranker reads the top 50 chunks per query. Voyage counts rerank tokens as the query tokens times the number of documents plus all document tokens, and the estimate uses that rule.

The quality scores are the one judgement call. They come from the RTEB retrieval board, the MTEB English and Multilingual boards, and the vendors' published comparisons, and they set the order, not the verdict. Benchmarks average over dozens of datasets and can hide exactly the domain you care about, which is why every result ends with an evaluation on your own queries.

Open models show a GPU line instead of a token price: one A10 or L4 class GPU at about $400 a month, from my own deployments ($300 to $500), and twice that for an 8B model. Treat it as an assumption to replace with your own cloud bill.

FAQ

Which embedding model should I use?

It depends on four things: your languages, what you search, where the data may go and how much you embed. For general English text, voyage-4-large has the best published retrieval quality at $0.12 per million tokens, and OpenAI text-embedding-3-small or voyage-4-lite at $0.02 are the cheap defaults. For code, voyage-code-4. For PDFs and slides, a multimodal model such as Cohere Embed 5 or voyage-multimodal-3.5. When the data must stay on your servers, BGE-M3 or Qwen3-Embedding. Confirm any choice on 50 of your own queries.

How does the picker score the models?

Each model has a quality score from 0 to 10 per content type, set from the public RTEB and MTEB leaderboards and the vendors' own published comparisons, plus language, latency, cost, price and stack adjustments. Your budget answer sets the weights. Hard rules come first: a self-host answer keeps only open-weight models, and an EU answer keeps only models with a confirmed EU processing route. Every result shows how its score adds up, and the same answers always give the same ranking.

How is the monthly embedding cost estimated?

From list prices only: the corpus tokens times the price per million for the first run, then the changed share of the corpus and your queries (at 50 tokens each) every month. Where a model shares an embedding space with a cheaper one, queries are priced on the cheaper model. Batch prices and free allowances are shown next to the estimate, not taken off it. Open models show an assumed GPU cost instead, labelled as an assumption.

How much storage do the vectors need?

Vectors times dimensions times bytes per dimension, before index overhead. The picker assumes chunks of 400 tokens, so 40 million tokens are 100,000 vectors; at 1024 dimensions in float32 (4 bytes each) that is about 410 MB. int8 output cuts it by 4x and binary by 32x, on the models that return those types themselves, such as the Voyage 4 family and Cohere Embed 5.

Do I need a reranker?

Usually yes once quality matters: retrieve the top 50 with embeddings, then let a reranker order them and keep the 5 to 10 that go into the prompt. The picker recommends one for your answers (Voyage rerank-3 or rerank-3-lite, Cohere Rerank 5 for many languages, bge-reranker-v2-m3 or Qwen3-Reranker to self-host) and prices it, because at high query volume the reranker can cost more than the embeddings.

Which embedding models keep data in the EU?

OpenAI processes embeddings in the EEA for projects created in its Europe region (eu.api.openai.com). MongoDB serves the Voyage 4 models, voyage-code-4 and voyage-multimodal-3.5 from a Europe endpoint (eu.ai.mongodb.com). Cohere Embed 5 can run in an EU region on Microsoft Foundry or Amazon SageMaker, or as a private deployment. Open models such as BGE-M3 and Qwen3-Embedding stay wherever you host them. The picker found no EU-only route confirmed for Gemini Embedding 2.

Is anything I enter sent anywhere?

No. The scoring runs in your browser and makes no network calls. A share link keeps your answers after the # in the URL, the part browsers never send to a server, and the site's analytics record visits, never your answers.