The best embedding model for RAG: 16 models compared on one knowledge base
There is no single best embedding model for RAG, but there is a best one for your documents, budget and constraints. How the main model families differ, which trade-offs matter, and a 16-model benchmark on a real knowledge base.
"Which embedding model is best for RAG?" has an honest answer: the one that retrieves the right passages from your documents, at a cost and latency you can live with. That sounds like a dodge, but it's a practical rule. The ranking of embedding models changes from one corpus to another, and the gap between the top models is usually smaller than the gap created by your domain, your chunking and your users' questions.
This post gives you a map of the options, the trade-offs that decide between them, and a starting shortlist per situation. The last section benchmarks all 16 models Truvec runs on the same knowledge base.
The model families
Embedding models for retrieval fall into a few groups:
| Family | Models in Truvec | What sets them apart |
|---|---|---|
| OpenAI | text-embedding-3-small, text-embedding-3-large, ada-002 | The default in most tutorials. 3-small is very cheap; both v3 models can shorten their vectors |
| Voyage AI | Voyage 4 Large, Voyage 4, Voyage 4 Lite | Retrieval-focused models in three price tiers, with separate query and document modes |
| Cohere | Embed v4, Embed English v3, Embed Multilingual v3 | Strong multilingual line-up; v3 models read up to 512 tokens per text |
| Gemini Embedding 2, Gemini Embedding 001 | Large 3,072-dimension vectors and long inputs | |
| Open source | Qwen3 Embedding 8B, BGE Base, Nomic, mxbai, GTE Large | Open weights: you can self-host them. Cheapest per token through a hosted API |
Older models such as text-embedding-ada-002 are still in many production systems. Including them in a comparison tells you what a migration would gain.
The trade-offs that decide
Retrieval quality on your data. Measure it with hit rate and MRR at the number of chunks you actually send to the LLM. How to choose an embedding model for RAG explains the metrics and how to build the test set.
Cost. Embedding prices differ by more than 10× between the cheapest and the most expensive models. You pay to embed the whole corpus once, again on every re-embedding (new chunking, new model), and a little per query. For most RAG systems the corpus is the bigger bill.
Vector size. Dimensions range from 768 to 4,096. Storage and memory grow linearly with them: at a million chunks, 4,096 dimensions take about 16 GB of raw float32 vectors, against 3 GB at 768. If your vector database keeps vectors in RAM, this is a monthly cost.
Input length. Some models read only 512 tokens per chunk and silently ignore the rest; others read thousands. If your chunks are long, a short-input model is evaluated on truncated text.
Languages. English-only models fall apart on other languages. If any part of your corpus or your users isn't English, test multilingual models.
Hosting. API models are simple to run but send your text to a provider. Open-weight models can run in your own infrastructure, which may be a hard requirement for regulated data.
A starting shortlist
Use this to decide what to test, not what to pick:
- Tight budget, English content: OpenAI text-embedding-3-small, Voyage 4 Lite, and an open model such as BGE Base or Nomic.
- Quality first: the large models from each provider (Voyage 4 Large, OpenAI 3-large, Cohere Embed v4, Gemini Embedding 2) plus Qwen3 Embedding 8B.
- Multilingual: Cohere Embed v4 or Multilingual v3, Gemini, Qwen3, and OpenAI 3-large as a reference.
- Must self-host: Qwen3 Embedding, BGE, Nomic, mxbai and GTE, compared with an API model to see what you give up.
- Migrating off ada-002: ada-002 itself, 3-small and 3-large, so the comparison shows the gain on your own data.
Why the winner changes between corpora
Run the same 16 models on product docs, contracts and support tickets, and you'll get three different orderings. Models differ in how they handle rare vocabulary (product names, legal terms, error codes), long passages, and short or vague questions. The differences that look large on a leaderboard often vanish on your data, and the ones that matter for you rarely show up on a leaderboard.
Two habits keep the decision honest:
- Treat close results as ties. With 100 questions, a difference of a couple of points is within the noise. Pick between tied models on price, latency and storage.
- Re-run when things change. New models ship several times a year, and your corpus grows. A comparison that takes minutes is worth repeating.
The benchmark below runs all 16 models on one knowledge base. Use it as an example of what a comparison shows, then run it on your own documents.
Real benchmark - generated by Truvec
Benchmark: All 16 Truvec models on a product knowledge base
- Documents
- 261
- Chunks
- 1,944
- Chunking
- recursive, 1000 chars, 200 overlap
- Golden set
- 100 questions, human-reviewed
- Models
- 16 compared
- Top k
- 5
- Run
- 2026-11-14
Download the full report (PDF) Every question, with the passages each model retrieved.
Key results
- Voyage 4 Large ranks first, with an MRR of 0.922 and a hit rate of 99.0% in the top 5.
- One model is within 0.02 MRR of the winner (Voyage 4): on this corpus they are practically tied.
- Hit rate ranges from 90.0% to 100.0% across the 16 models.
- Cheapest model within 2 points of the winner's hit rate: Voyage 4 Lite, $0.0067 to embed this corpus against $0.04.
- Fastest to embed a query: OpenAI text-embedding-ada-002 (5.5 ms per query, measured in batch from our servers).
Retrieval quality
| Model | Hit rate@5 | Recall@5 | MRR | nDCG@5 | Precision@5 |
|---|---|---|---|---|---|
| 1Voyage 4 Large | 99.0% | 97.5% | 0.924 | 20.6% | |
| 2Voyage 4 | 100.0% | 99.5% | 0.929 | 21.2% | |
| 3Voyage 4 Lite | 99.0% | 97.5% | 0.907 | 20.6% | |
| 4Gemini Embedding 2 | 99.0% | 98.5% | 0.897 | 21.0% | |
| 5Gemini Embedding 001 | 98.0% | 97.5% | 0.890 | 20.8% | |
| 6OpenAI text-embedding-3-large | 99.0% | 97.5% | 0.885 | 20.6% | |
| 7Cohere Embed v4 | 98.0% | 96.5% | 0.875 | 20.4% | |
| 8Qwen3 Embedding 8B | 99.0% | 98.0% | 0.883 | 20.8% | |
| 9Cohere Embed Multilingual v3 | 94.0% | 92.5% | 0.861 | 19.6% | |
| 10mxbai Embed Large v1 | 96.0% | 93.5% | 0.837 | 19.6% | |
| 11Cohere Embed English v3 | 97.0% | 95.0% | 0.838 | 20.0% | |
| 12GTE Large | 93.0% | 90.5% | 0.810 | 19.0% | |
| 13BGE Base EN v1.5 | 95.0% | 92.5% | 0.802 | 19.4% | |
| 14OpenAI text-embedding-3-small | 95.0% | 93.5% | 0.807 | 19.8% | |
| 15Nomic Embed Text v1.5 | 90.0% | 89.0% | 0.789 | 18.8% | |
| 16OpenAI text-embedding-ada-002 | 93.0% | 91.5% | 0.758 | 19.4% |
Hit rate by position
How many questions the first results already answer. A model that finds the answer at position 1 lets you send fewer chunks to the LLM.
| Model | Top 1 | Top 2 | Top 3 | Top 4 | Top 5 |
|---|---|---|---|---|---|
| 1Voyage 4 Large | 87.0% | 94.0% | 99.0% | 99.0% | 99.0% |
| 2Voyage 4 | 85.0% | 95.0% | 99.0% | 99.0% | 100.0% |
| 3Voyage 4 Lite | 81.0% | 95.0% | 99.0% | 99.0% | 99.0% |
| 4Gemini Embedding 2 | 80.0% | 90.0% | 95.0% | 98.0% | 99.0% |
| 5Gemini Embedding 001 | 78.0% | 92.0% | 97.0% | 98.0% | 98.0% |
| 6OpenAI text-embedding-3-large | 79.0% | 88.0% | 94.0% | 98.0% | 99.0% |
| 7Cohere Embed v4 | 76.0% | 93.0% | 95.0% | 97.0% | 98.0% |
| 8Qwen3 Embedding 8B | 77.0% | 88.0% | 95.0% | 99.0% | 99.0% |
| 9Cohere Embed Multilingual v3 | 78.0% | 88.0% | 92.0% | 94.0% | 94.0% |
| 10mxbai Embed Large v1 | 73.0% | 84.0% | 88.0% | 94.0% | 96.0% |
| 11Cohere Embed English v3 | 71.0% | 84.0% | 92.0% | 94.0% | 97.0% |
| 12GTE Large | 68.0% | 86.0% | 91.0% | 93.0% | 93.0% |
| 13BGE Base EN v1.5 | 64.0% | 85.0% | 90.0% | 92.0% | 95.0% |
| 14OpenAI text-embedding-3-small | 63.0% | 84.0% | 92.0% | 94.0% | 95.0% |
| 15Nomic Embed Text v1.5 | 67.0% | 80.0% | 86.0% | 89.0% | 90.0% |
| 16OpenAI text-embedding-ada-002 | 55.0% | 79.0% | 88.0% | 93.0% | 93.0% |
Cost, speed and size
Embedding prices are the providers' list prices. Corpus cost is a one-time cost per model; the vector size is raw float32 storage, before the vector database's own overhead.
| Model | Price / 1M tokens | Embed the corpus | 1,000 queries | Query embedding | Search p95 | Vectors |
|---|---|---|---|---|---|---|
| 1Voyage 4 Large | $0.12 | $0.04 | $0.0024 | 17 ms | 16 ms | 7.6 MB |
| 2Voyage 4 | $0.06 | $0.02 | $0.0012 | 11 ms | 24 ms | 7.6 MB |
| 3Voyage 4 Lite | $0.02 | $0.0067 | $0.0004 | 17 ms | 20 ms | 7.6 MB |
| 4Gemini Embedding 2 | $0.20 | $0.07 | $0.0040 | 20 ms | 27 ms | 22.8 MB |
| 5Gemini Embedding 001 | $0.15 | $0.05 | $0.0030 | 16 ms | 23 ms | 22.8 MB |
| 6OpenAI text-embedding-3-large | $0.13 | $0.04 | $0.0026 | 8.0 ms | 30 ms | 22.8 MB |
| 7Cohere Embed v4 | $0.12 | $0.04 | $0.0024 | 14 ms | 54 ms | 11.4 MB |
| 8Qwen3 Embedding 8B | $0.10 | $0.03 | $0.0020 | 10 ms | 33 ms | 30.4 MB |
| 9Cohere Embed Multilingual v3 | $0.10 | $0.03 | $0.0020 | 11 ms | 17 ms | 7.6 MB |
| 10mxbai Embed Large v1 | $0.02 | $0.0054 | $0.0003 | 18 ms | 16 ms | 7.6 MB |
| 11Cohere Embed English v3 | $0.10 | $0.03 | $0.0020 | 8.9 ms | 16 ms | 7.6 MB |
| 12GTE Large | $0.02 | $0.0054 | $0.0003 | 18 ms | 24 ms | 7.6 MB |
| 13BGE Base EN v1.5 | $0.0080 | $0.0027 | $0.0002 | 14 ms | 15 ms | 5.7 MB |
| 14OpenAI text-embedding-3-small | $0.02 | $0.0067 | $0.0004 | 9.4 ms | 41 ms | 11.4 MB |
| 15Nomic Embed Text v1.5 | $0.0080 | $0.0027 | $0.0002 | 13 ms | 19 ms | 5.7 MB |
| 16OpenAI text-embedding-ada-002 | $0.10 | $0.03 | $0.0020 | 5.5 ms | 27 ms | 11.4 MB |
Score separation
Average similarity of the best relevant chunk and of the best irrelevant one. A wide gap makes a similarity threshold reliable; a narrow one means right and wrong chunks look alike to the model.
| Model | Relevant chunk | Best distractor | Gap |
|---|---|---|---|
| 1Voyage 4 Large | 0.653 | 0.607 | 0.046 |
| 2Voyage 4 | 0.692 | 0.646 | 0.046 |
| 3Voyage 4 Lite | 0.665 | 0.622 | 0.044 |
| 4Gemini Embedding 2 | 0.816 | 0.792 | 0.025 |
| 5Gemini Embedding 001 | 0.796 | 0.781 | 0.015 |
| 6OpenAI text-embedding-3-large | 0.687 | 0.650 | 0.037 |
| 7Cohere Embed v4 | 0.684 | 0.638 | 0.046 |
| 8Qwen3 Embedding 8B | 0.827 | 0.785 | 0.042 |
| 9Cohere Embed Multilingual v3 | 0.788 | 0.747 | 0.041 |
| 10mxbai Embed Large v1 | 0.842 | 0.817 | 0.026 |
| 11Cohere Embed English v3 | 0.777 | 0.741 | 0.037 |
| 12GTE Large | 0.929 | 0.917 | 0.012 |
| 13BGE Base EN v1.5 | 0.806 | 0.782 | 0.023 |
| 14OpenAI text-embedding-3-small | 0.715 | 0.687 | 0.027 |
| 15Nomic Embed Text v1.5 | 0.842 | 0.815 | 0.027 |
| 16OpenAI text-embedding-ada-002 | 0.871 | 0.861 | 0.009 |
Example questions
Rank of the first relevant chunk for each model; miss means it wasn't in the top 5. These are the questions where the models disagree most.
| Question | Voyage 4 Large | Voyage 4 | Voyage 4 Lite | Gemini Embedding 2 | Gemini Embedding 001 | OpenAI text-embedding-3-large | Cohere Embed v4 | Qwen3 Embedding 8B | Cohere Embed Multilingual v3 | mxbai Embed Large v1 | Cohere Embed English v3 | GTE Large | BGE Base EN v1.5 | OpenAI text-embedding-3-small | Nomic Embed Text v1.5 | OpenAI text-embedding-ada-002 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
How do I push an image to Docker Hub, and what tagging restriction applies?Expected passage## Push an image to Docker Hub Select the image from the list, select the **More options** button and select **Push to Hub**. > [!NOTE] > > You can only push an image to Docker Hub if the image belongs to your Docker ID or your organization. That is, the image must contain the co | miss | 5 | miss | 5 | miss | miss | miss | 4 | 4 | 4 | 2 | 4 | 3 | 1 | 4 | miss |
When both JIT and SCIM are enabled, how are organization roles and account deactivation handled?Expected passageRoles don't move back and forth. JIT sets the organization role only when it first adds the user. A later SCIM update can change that role, and the next sign-in leaves the SCIM role in place. When a user isn't assigned to the Docker application in the IdP, the next synchronizatio | 1 | 1 | 2 | 1 | 1 | 5 | 3 | miss | 3 | 5 | 2 | 4 | 3 | miss | 3 | 4 |
How do I create, publish, and share a custom MCP catalog with my team?Expected passageCreate a catalog and specify which servers to include: ```console $ docker mcp catalog create registry.example.com/mcp/data-tools:latest \ --title "Data Analysis Tools" \ --server catalog://mcp/docker-mcp-catalog/sequentialthinking \ --server catalog://mcp/docker-mcp-catalog/brav | 1 | 1 | 1 | 1 | 1 | 1 | 5 | 3 | miss | 4 | 5 | 2 | 2 | 2 | 3 | 3 |
How do I search Docker Desktop logs, move between matching entries, and clear the terminal?Expected passage### Logs Select **Logs** to view output from the container in real time. While viewing logs, you can: - Use `Cmd + f`/`Ctrl + f` to open the search bar and find specific entries. Search matches are highlighted in yellow. - Press `Enter` or `Shift + Enter` to jump to the next or p | 1 | 1 | 1 | 2 | 3 | 4 | 2 | 3 | miss | 1 | 5 | 2 | 1 | 1 | 2 | 1 |
When multiple Compose files are supplied with `-f`, how are their service settings combined, and which file takes precedence?Expected passageCompose merges files in the order they're specified on the command line. Subsequent files may merge, override, or add to their predecessors. For example: ```console $ docker compose -f compose.yaml -f compose.admin.yaml run backup_db ``` The `compose.yaml` file might specify a `w | 1 | 2 | 3 | 1 | 1 | 1 | 4 | 2 | miss | 5 | 3 | miss | miss | 4 | miss | miss |
Which organizational roles and subscription plans are needed to create organization access tokens in Docker?Expected passage> [!WARNING] > > OATs don't work with Docker Desktop or Image Access Management. For > those features, use > PATs instead. ## Best practices - Set an expiration date and rotate tokens on a schedule - Grant only the repositories and scopes each job needs - Review the **Last used** | 1 | 1 | 1 | 1 | 1 | 4 | 1 | 2 | 3 | miss | miss | miss | miss | miss | miss | miss |
How do I make a container service accessible from my host computer using Docker port publishing?Expected passage# Publishing and exposing ports ## Explanation If you've been following the guides so far, you understand that containers provide isolated processes for each component of your application. Each component - a React frontend, a Python API, and a Postgres database - runs in its own | 3 | 3 | 3 | 4 | miss | 1 | 1 | 3 | miss | miss | 1 | 2 | 1 | 3 | miss | 2 |
How can I confirm that an image migration preserved the same digest, including for multi-platform images?Expected passage```bash #!/usr/bin/env bash set -euo pipefail SRC_ORG="oldorg" DEST_ORG="neworg" REPO="myapp" TAG="1.2.3" SRC_IMAGE="${SRC_ORG}/${REPO}:${TAG}" DEST_IMAGE="${DEST_ORG}/${REPO}:${TAG}" # Using crane SRC_DIGEST=$(crane digest "${SRC_IMAGE}") DEST_DIGEST=$(crane digest "${DEST_IMAGE | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 3 | 5 | miss | 3 | 1 | miss | 1 |
Notes
- The corpus is a subset of the public Docker documentation (Apache 2.0), taken from github.com/docker/docs at commit 553482c: the Desktop, Docker Hub, Compose, accounts, security and subscription and billing manuals, the FAQs, the support pages and the getting-started guides. Release notes and pages shorter than 600 characters were left out, and pages were cut at 40,000 characters.
- Several posts on this blog use this corpus and the same 100 questions, each with its own selection of models. A model can differ by a question between two of these runs: questions are embedded again at each run, and vector search is approximate.
- With 100 questions, a gap of a couple of points between two models is within the noise.
- BGE Base, mxbai Embed Large and GTE Large read at most 380 tokens per chunk in Truvec, and the two Cohere v3 models 512. Chunks average about 170 tokens, so little or nothing was truncated.
Methodology
Truvec split the documents into chunks, generated a golden set of questions with gpt-5.6-luna (each question is linked to the chunks that answer it), embedded the corpus and the questions with every model, and retrieved the top 5 chunks for each question by cosine similarity. Every model saw the same chunks and the same questions. Scores are averaged over all 100 questions.
- Hit rate
- Share of questions with at least one relevant chunk in the top k results. The most direct answer to "does the right passage reach the LLM?".
- Recall
- Share of the relevant chunks found in the top k, averaged over questions. Matters when an answer spans several chunks.
- MRR
- Mean reciprocal rank: 1 when the first relevant chunk is ranked first, 1/2 when second, 0 when missing. Rewards putting the answer at the top.
- nDCG
- Normalized discounted cumulative gain: how close the ranking of all relevant chunks is to the ideal ranking.
- Precision
- Share of the top k results that are relevant. Low precision means more noise sent to the LLM.
Downloads: full report (PDF) - benchmark.json - benchmark.csv - this post as Markdown
Voyage 4 Large won here, on these documents. Find the model that wins on yours: run the same benchmark on your documents with Truvec.