BGE vs OpenAI embeddings: is a small open model good enough?
BGE Base is a small open-source embedding model you can run almost anywhere. How it compares with OpenAI's text-embedding-3 models on English documents, where its 512-token limit and English-only training matter, and a benchmark.
BGE (BAAI General Embedding) models are among the most downloaded open-source embedding models. BGE Base EN v1.5 is small enough to run on a CPU, produces compact 768-dimension vectors, and costs almost nothing through a hosted API. OpenAI's text-embedding-3 models are the default choice for many teams. The question is simple: does the small open model give up enough quality to justify paying for OpenAI?
This post compares BGE Base EN v1.5 with OpenAI text-embedding-3-small and 3-large, then benchmarks them on English documentation.
At a glance
| BGE Base EN v1.5 | OpenAI 3-small | OpenAI 3-large | |
|---|---|---|---|
| Weights | Open source | API only | API only |
| Price per 1M tokens (hosted, in Truvec) | $0.008 | $0.02 | $0.13 |
| Dimensions | 768 | 1,536 | 3,072 |
| Max input | 512 tokens | 8,191 tokens | 8,191 tokens |
| Languages | English | Multilingual | Multilingual |
What to know about BGE
512 tokens is the hard limit. Anything past it is ignored. With 1,000-character chunks (about 250 tokens) that's fine. With long chunks, BGE is evaluated on truncated text, and its scores drop for reasons that have nothing to do with the model's understanding. If you use BGE, keep chunks short.
It's English-only. On French, Arabic or mixed-language content, it falls far behind multilingual models. Don't use it outside English.
It needs a query prefix. BGE was trained with an instruction prepended to queries ("Represent this sentence for searching relevant passages: "). Without it, retrieval gets worse. Many integrations forget it.
It's tiny to run. A base-size model embeds thousands of chunks per minute on modest hardware, and the vectors take a quarter of the space of OpenAI 3-large's. For edge deployments, on-premises installs or tight budgets, that matters more than a few points of quality.
When BGE is good enough
- English content, short chunks, and a budget or hosting constraint;
- a first version of a RAG system, where you need something that works today and can switch later;
- a large corpus that gets re-embedded often, where per-token cost adds up.
When to pay for OpenAI (or another API)
- Non-English or mixed-language content;
- long chunks you don't want to split further;
- a measurable quality gap on your documents that changes your users' answers.
The benchmark below compares the three models on English developer documentation, with chunks short enough for BGE to read in full. Look at the size of the quality gap, then at the cost column: whether the gap is worth the price is your call, but now it's an informed one. For what each metric in the tables means, see RAG evaluation metrics explained; for why a public leaderboard can't settle this for your documents, see why MTEB alone isn't enough.
Real benchmark - generated by Truvec
Benchmark: BGE Base vs OpenAI on English documentation
- Documents
- 261
- Chunks
- 1,944
- Chunking
- recursive, 1000 chars, 200 overlap
- Golden set
- 100 questions, human-reviewed
- Models
- 3 compared
- Top k
- 5
- Run
- 2026-11-14
Download the full report (PDF) Every question, with the passages each model retrieved.
Key results
- OpenAI text-embedding-3-large ranks first, with an MRR of 0.867 and a hit rate of 99.0% in the top 5.
- It leads the runner-up, BGE Base EN v1.5, by 0.094 MRR.
- Hit rate ranges from 95.0% to 99.0% across the 3 models.
- Fastest to embed a query: OpenAI text-embedding-3-small (6.3 ms per query, measured in batch from our servers).
Retrieval quality
| Model | Hit rate@5 | Recall@5 | MRR | nDCG@5 | Precision@5 |
|---|---|---|---|---|---|
| 1OpenAI text-embedding-3-large | 99.0% | 97.5% | 0.885 | 20.6% | |
| 2BGE Base EN v1.5 | 95.0% | 92.5% | 0.802 | 19.4% | |
| 3OpenAI text-embedding-3-small | 96.0% | 94.5% | 0.811 | 20.0% |
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 |
|---|---|---|---|---|---|
| 1OpenAI text-embedding-3-large | 79.0% | 88.0% | 94.0% | 98.0% | 99.0% |
| 2BGE Base EN v1.5 | 64.0% | 85.0% | 90.0% | 92.0% | 95.0% |
| 3OpenAI text-embedding-3-small | 63.0% | 84.0% | 92.0% | 94.0% | 96.0% |
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 |
|---|---|---|---|---|---|---|
| 1OpenAI text-embedding-3-large | $0.13 | $0.04 | $0.0026 | 12 ms | 57 ms | 22.8 MB |
| 2BGE Base EN v1.5 | $0.0080 | $0.0027 | $0.0002 | 22 ms | 40 ms | 5.7 MB |
| 3OpenAI text-embedding-3-small | $0.02 | $0.0067 | $0.0004 | 6.3 ms | 48 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 |
|---|---|---|---|
| 1OpenAI text-embedding-3-large | 0.687 | 0.650 | 0.037 |
| 2BGE Base EN v1.5 | 0.806 | 0.782 | 0.023 |
| 3OpenAI text-embedding-3-small | 0.715 | 0.687 | 0.027 |
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 | OpenAI text-embedding-3-large | BGE Base EN v1.5 | OpenAI text-embedding-3-small |
|---|---|---|---|
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 | 3 | 1 |
How can I import a custom MCP catalog into Docker Desktop, and where does it appear afterward?Expected passage### Custom catalogs with Dynamic MCP Custom catalogs work particularly well with Dynamic MCP, where agents discover and add MCP servers on-demand during conversations. When you run the gateway with a custom catalog, the `mcp-find` tool searches only within that catalog. If your c | 1 | 2 | 3 |
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 | 5 | 3 | miss |
Which Docker Desktop versions are supported under the Pro, Team, and Business plans?Expected passage### Supported versions Docker Desktop support is scoped by plan: | Plan | Supported versions | Machines per subscription | | :--- | :--- | :--- | | Docker Pro | Latest only | One | | Docker Team | Latest only | One per subscription seat | | Docker Business | Latest plus versions | 2 | 1 | 3 |
Where can I inspect all Docker Desktop settings that have been applied, including settings not visible in the GUI?Expected passageAs a workaround, you can check the `settings-store.json` file to view all applied settings: - Mac: `~/Library/Application Support/Docker/settings-store.json` - Windows: `%APPDATA%\Docker\settings-store.json` - Linux: `~/.docker/desktop/settings-store.json` The `settings-store.jso | 1 | 4 | 2 |
What must an organization’s only owner do before leaving, and where can they initiate the departure?Expected passageIf the account is the only owner of an organization or company, assign the owner role to another member first. Leave after that member can administer the organization or company. To leave an organization and its teams: 1. Sign in to Docker Home and select the organization. 1. Sel | 3 | 1 | 2 |
How can I use a value from a `.env` file inside a Docker Compose service's environment settings?Expected passageBasic example: ```console $ cat .env ## define COMPOSE_DEBUG based on DEV_MODE, defaults to false COMPOSE_DEBUG=${DEV_MODE:-false} $ cat compose.yaml services: webapp: image: my-webapp-image environment: - DEBUG=${COMPOSE_DEBUG} $ DEV_MODE=true docker compose config services: web | 1 | 5 | 2 |
What file and service configuration are required for Docker Hub automated tests, and how is the test result determined?Expected passage## Set up automated test files To set up your automated tests, create a `docker-compose.test.yml` file which defines a `sut` service that lists the tests to be run. The `docker-compose.test.yml` file should be located in the same directory that contains the Dockerfile used to bui | 2 | 5 | miss |
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.
- Chunks average about 170 tokens, well under the 380 tokens BGE Base EN v1.5 reads per chunk in Truvec (model limit 512), so BGE saw the chunks in full.
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
OpenAI text-embedding-3-large won here, on these documents. Find the model that wins on yours: run the same benchmark on your documents with Truvec.