A blazing fast inference solution for text embeddings models.
Benchmark for BAAI/bge-base-en-v1.5 on an Nvidia A10 with a sequence length of 512 tokens:
Text Embeddings Inference (TEI) is a toolkit for deploying and serving open source text embeddings and sequence classification models. TEI enables high-performance extraction for the most popular models, including FlagEmbedding, Ember, GTE and E5. TEI implements many features such as:
- No model graph compilation step
- Metal support for local execution on Macs
- Small docker images and fast boot times. Get ready for true serverless!
- Token based dynamic batching
- Optimized transformers code for inference using Flash Attention, Candle and cuBLASLt
- Safetensors weight loading
- Production ready (distributed tracing with Open Telemetry, Prometheus metrics)
Text Embeddings Inference currently supports Nomic, BERT, CamemBERT, XLM-RoBERTa models with absolute positions, JinaBERT model with Alibi positions and Mistral, Alibaba GTE and Qwen2 models with Rope positions.
Below are some examples of the currently supported models:
| MTEB Rank | Model Size | Model Type | Model ID |
|---|---|---|---|
| 1 | 7B (Very Expensive) | Mistral | Salesforce/SFR-Embedding-2_R |
| 2 | 7B (Very Expensive) | Qwen2 | Alibaba-NLP/gte-Qwen2-7B-instruct |
| 9 | 1.5B (Expensive) | Qwen2 | Alibaba-NLP/gte-Qwen2-1.5B-instruct |
| 15 | 0.4B | Alibaba GTE | Alibaba-NLP/gte-large-en-v1.5 |
| 20 | 0.3B | Bert | WhereIsAI/UAE-Large-V1 |
| 24 | 0.5B | XLM-RoBERTa | intfloat/multilingual-e5-large-instruct |
| N/A | 0.1B | NomicBert | nomic-ai/nomic-embed-text-v1 |
| N/A | 0.1B | NomicBert | nomic-ai/nomic-embed-text-v1.5 |
| N/A | 0.1B | JinaBERT | jinaai/jina-embeddings-v2-base-en |
| N/A | 0.1B | JinaBERT | jinaai/jina-embeddings-v2-base-code |
To explore the list of best performing text embeddings models, visit the Massive Text Embedding Benchmark (MTEB) Leaderboard.
Text Embeddings Inference currently supports CamemBERT, and XLM-RoBERTa Sequence Classification models with absolute positions.
Below are some examples of the currently supported models:
| Task | Model Type | Model ID |
|---|---|---|
| Re-Ranking | XLM-RoBERTa | BAAI/bge-reranker-large |
| Re-Ranking | XLM-RoBERTa | BAAI/bge-reranker-base |
| Sentiment Analysis | RoBERTa | SamLowe/roberta-base-go_emotions |
model=BAAI/bge-large-en-v1.5
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id $modelAnd then you can make requests like
curl 127.0.0.1:8080/embed \
-X POST \
-d '{"inputs":"What is Deep Learning?"}' \
-H 'Content-Type: application/json'Note: To use GPUs, you need to install the NVIDIA Container Toolkit. NVIDIA drivers on your machine need to be compatible with CUDA version 12.2 or higher.
To see all options to serve your models:
text-embeddings-router --helpUsage: text-embeddings-router [OPTIONS]
Options:
--model-id <MODEL_ID>
The name of the model to load. Can be a MODEL_ID as listed on <https://hf.co/models> like `thenlper/gte-base`.
Or it can be a local directory containing the necessary files as saved by `save_pretrained(...)` methods of
transformers
[env: MODEL_ID=]
[default: thenlper/gte-base]
--revision <REVISION>
The actual revision of the model if you're referring to a model on the hub. You can use a specific commit id
or a branch like `refs/pr/2`
[env: REVISION=]
--tokenization-workers <TOKENIZATION_WORKERS>
Optionally control the number of tokenizer workers used for payload tokenization, validation and truncation.
Default to the number of CPU cores on the machine
[env: TOKENIZATION_WORKERS=]
--dtype <DTYPE>
The dtype to be forced upon the model
[env: DTYPE=]
[possible values: float16, float32]
--pooling <POOLING>
Optionally control the pooling method for embedding models.
If `pooling` is not set, the pooling configuration will be parsed from the model `1_Pooling/config.json` configuration.
If `pooling` is set, it will override the model pooling configuration
[env: POOLING=]
Possible values:
- cls: Select the CLS token as embedding
- mean: Apply Mean pooling to the model embeddings
- splade: Apply SPLADE (Sparse Lexical and Expansion) to the model embeddings. This option is only
available if the loaded model is a `ForMaskedLM` Transformer model
- last-token: Select the last token as embedding
--max-concurrent-requests <MAX_CONCURRENT_REQUESTS>
The maximum amount of concurrent requests for this particular deployment.
Having a low limit will refuse clients requests instead of having them wait for too long and is usually good
to handle backpressure correctly
[env: MAX_CONCURRENT_REQUESTS=]
[default: 512]
--max-batch-tokens <MAX_BATCH_TOKENS>
**IMPORTANT** This is one critical control to allow maximum usage of the available hardware.
This represents the total amount of potential tokens within a batch.
For `max_batch_tokens=1000`, you could fit `10` queries of `total_tokens=100` or a single query of `1000` tokens.
Overall this number should be the largest possible until the model is compute bound. Since the actual memory
overhead depends on the model implementation, text-embeddings-inference cannot infer this number automatically.
[env: MAX_BATCH_TOKENS=]
[default: 16384]
--max-batch-requests <MAX_BATCH_REQUESTS>
Optionally control the maximum number of individual requests in a batch
[env: MAX_BATCH_REQUESTS=]
--max-client-batch-size <MAX_CLIENT_BATCH_SIZE>
Control the maximum number of inputs that a client can send in a single request
[env: MAX_CLIENT_BATCH_SIZE=]
[default: 32]
--auto-truncate
Automatically truncate inputs that are longer than the maximum supported size
Unused for gRPC servers
[env: AUTO_TRUNCATE=]
--default-prompt-name <DEFAULT_PROMPT_NAME>
The name of the prompt that should be used by default for encoding. If not set, no prompt will be applied.
Must be a key in the `sentence-transformers` configuration `prompts` dictionary.
For example if ``default_prompt_name`` is "query" and the ``prompts`` is {"query": "query: ", ...}, then the
sentence "What is the capital of France?" will be encoded as "query: What is the capital of France?" because
the prompt text will be prepended before any text to encode.
The argument '--default-prompt-name <DEFAULT_PROMPT_NAME>' cannot be used with '--default-prompt <DEFAULT_PROMPT>`
[env: DEFAULT_PROMPT_NAME=]
--default-prompt <DEFAULT_PROMPT>
The prompt that should be used by default for encoding. If not set, no prompt will be applied.
For example if ``default_prompt`` is "query: " then the sentence "What is the capital of France?" will be
encoded as "query: What is the capital of France?" because the prompt text will be prepended before any text
to encode.
The argument '--default-prompt <DEFAULT_PROMPT>' cannot be used with '--default-prompt-name <DEFAULT_PROMPT_NAME>`
[env: DEFAULT_PROMPT=]
--hf-api-token <HF_API_TOKEN>
Your HuggingFace hub token
[env: HF_API_TOKEN=]
--hostname <HOSTNAME>
The IP address to listen on
[env: HOSTNAME=]
[default: 0.0.0.0]
-p, --port <PORT>
The port to listen on
[env: PORT=]
[default: 3000]
--uds-path <UDS_PATH>
The name of the unix socket some text-embeddings-inference backends will use as they communicate internally
with gRPC
[env: UDS_PATH=]
[default: /tmp/text-embeddings-inference-server]
--huggingface-hub-cache <HUGGINGFACE_HUB_CACHE>
The location of the huggingface hub cache. Used to override the location if you want to provide a mounted disk
for instance
[env: HUGGINGFACE_HUB_CACHE=]
--payload-limit <PAYLOAD_LIMIT>
Payload size limit in bytes
Default is 2MB
[env: PAYLOAD_LIMIT=]
[default: 2000000]
--api-key <API_KEY>
Set an api key for request authorization.
By default the server responds to every request. With an api key set, the requests must have the Authorization
header set with the api key as Bearer token.
[env: API_KEY=]
--json-output
Outputs the logs in JSON format (useful for telemetry)
[env: JSON_OUTPUT=]
--otlp-endpoint <OTLP_ENDPOINT>
The grpc endpoint for opentelemetry. Telemetry is sent to this endpoint as OTLP over gRPC. e.g. `http://localhost:4317`
[env: OTLP_ENDPOINT=]
--otlp-service-name <OTLP_SERVICE_NAME>
The service name for opentelemetry. e.g. `text-embeddings-inference.server`
[env: OTLP_SERVICE_NAME=]
[default: text-embeddings-inference.server]
--cors-allow-origin <CORS_ALLOW_ORIGIN>
Unused for gRPC servers
[env: CORS_ALLOW_ORIGIN=]
Text Embeddings Inference ships with multiple Docker images that you can use to target a specific backend:
| Architecture | Image |
|---|---|
| CPU | ghcr.io/huggingface/text-embeddings-inference:cpu-1.5 |
| Volta | NOT SUPPORTED |
| Turing (T4, RTX 2000 series, ...) | ghcr.io/huggingface/text-embeddings-inference:turing-1.5 (experimental) |
| Ampere 80 (A100, A30) | ghcr.io/huggingface/text-embeddings-inference:1.5 |
| Ampere 86 (A10, A40, ...) | ghcr.io/huggingface/text-embeddings-inference:86-1.5 |
| Ada Lovelace (RTX 4000 series, ...) | ghcr.io/huggingface/text-embeddings-inference:89-1.5 |
| Hopper (H100) | ghcr.io/huggingface/text-embeddings-inference:hopper-1.5 (experimental) |
Warning: Flash Attention is turned off by default for the Turing image as it suffers from precision issues.
You can turn Flash Attention v1 ON by using the USE_FLASH_ATTENTION=True environment variable.
You can consult the OpenAPI documentation of the text-embeddings-inference REST API using the /docs route.
The Swagger UI is also available
at: https://huggingface.github.io/text-embeddings-inference.
You have the option to utilize the HF_API_TOKEN environment variable for configuring the token employed by
text-embeddings-inference. This allows you to gain access to protected resources.
For example:
- Go to https://huggingface.co/settings/tokens
- Copy your cli READ token
- Export
HF_API_TOKEN=<your cli READ token>
or with Docker:
model=<your private model>
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
token=<your cli READ token>
docker run --gpus all -e HF_API_TOKEN=$token -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id $modelTo deploy Text Embeddings Inference in an air-gapped environment, first download the weights and then mount them inside the container using a volume.
For example:
# (Optional) create a `models` directory
mkdir models
cd models
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/Alibaba-NLP/gte-base-en-v1.5
# Set the models directory as the volume path
volume=$PWD
# Mount the models directory inside the container with a volume and set the model ID
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id /data/gte-base-en-v1.5text-embeddings-inference v0.4.0 added support for CamemBERT, RoBERTa and XLM-RoBERTa Sequence Classification models.
Re-rankers models are Sequence Classification cross-encoders models with a single class that scores the similarity
between a query and a text.
See this blogpost by the LlamaIndex team to understand how you can use re-rankers models in your RAG pipeline to improve downstream performance.
model=BAAI/bge-reranker-large
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id $modelAnd then you can rank the similarity between a query and a list of texts with:
curl 127.0.0.1:8080/rerank \
-X POST \
-d '{"query":"What is Deep Learning?", "texts": ["Deep Learning is not...", "Deep learning is..."]}' \
-H 'Content-Type: application/json'You can also use classic Sequence Classification models like SamLowe/roberta-base-go_emotions:
model=SamLowe/roberta-base-go_emotions
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id $modelOnce you have deployed the model you can use the predict endpoint to get the emotions most associated with an input:
curl 127.0.0.1:8080/predict \
-X POST \
-d '{"inputs":"I like you."}' \
-H 'Content-Type: application/json'You can choose to activate SPLADE pooling for Bert and Distilbert MaskedLM architectures:
model=naver/efficient-splade-VI-BT-large-query
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5 --model-id $model --pooling spladeOnce you have deployed the model you can use the /embed_sparse endpoint to get the sparse embedding:
curl 127.0.0.1:8080/embed_sparse \
-X POST \
-d '{"inputs":"I like you."}' \
-H 'Content-Type: application/json'text-embeddings-inference is instrumented with distributed tracing using OpenTelemetry. You can use this feature
by setting the address to an OTLP collector with the --otlp-endpoint argument.
text-embeddings-inference offers a gRPC API as an alternative to the default HTTP API for high performance
deployments. The API protobuf definition can be
found here.
You can use the gRPC API by adding the -grpc tag to any TEI Docker image. For example:
model=BAAI/bge-large-en-v1.5
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.5-grpc --model-id $modelgrpcurl -d '{"inputs": "What is Deep Learning"}' -plaintext 0.0.0.0:8080 tei.v1.Embed/EmbedYou can also opt to install text-embeddings-inference locally.
First install Rust:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | shThen run:
# On x86
cargo install --path router -F mkl
# On M1 or M2
cargo install --path router -F metalYou can now launch Text Embeddings Inference on CPU with:
model=BAAI/bge-large-en-v1.5
text-embeddings-router --model-id $model --port 8080Note: on some machines, you may also need the OpenSSL libraries and gcc. On Linux machines, run:
sudo apt-get install libssl-dev gcc -yGPUs with Cuda compute capabilities < 7.5 are not supported (V100, Titan V, GTX 1000 series, ...).
Make sure you have Cuda and the nvidia drivers installed. NVIDIA drivers on your device need to be compatible with CUDA version 12.2 or higher. You also need to add the nvidia binaries to your path:
export PATH=$PATH:/usr/local/cuda/binThen run:
# This can take a while as we need to compile a lot of cuda kernels
# On Turing GPUs (T4, RTX 2000 series ... )
cargo install --path router -F candle-cuda-turing -F http --no-default-features
# On Ampere and Hopper
cargo install --path router -F candle-cuda -F http --no-default-featuresYou can now launch Text Embeddings Inference on GPU with:
model=BAAI/bge-large-en-v1.5
text-embeddings-router --model-id $model --port 8080You can build the CPU container with:
docker build .To build the Cuda containers, you need to know the compute cap of the GPU you will be using at runtime.
Then you can build the container with:
# Example for Turing (T4, RTX 2000 series, ...)
runtime_compute_cap=75
# Example for A100
runtime_compute_cap=80
# Example for A10
runtime_compute_cap=86
# Example for Ada Lovelace (RTX 4000 series, ...)
runtime_compute_cap=89
# Example for H100
runtime_compute_cap=90
docker build . -f Dockerfile-cuda --build-arg CUDA_COMPUTE_CAP=$runtime_compute_capAs explained here MPS-Ready, ARM64 Docker Image, Metal / MPS is not supported via Docker. As such inference will be CPU bound and most likely pretty slow when using this docker image on an M1/M2 ARM CPU.
docker build . -f Dockerfile --platform=linux/arm64