Sentence Transformers

By Hugging Face

Run embedding models locally (e.g. bge-m3) to turn text into vectors for RAG and memory — no API.

Website ↗ Repository ↗ Docs ↗

v5.6.02026-06-16PythonApache-2.0

#embeddings #rag #local #python

18.9k Completely free Open source

Doc version 1.0 · Updated June 28, 2026

Related tools

Overview

Sentence Transformers (SBERT) runs embedding models — like the multilingual bge-m3 — directly on your machine, turning text into vectors with no API call.
It is the local, provider-independent half of a RAG pipeline: generate embeddings here, then store and search them in a vector store.

flowchart LR
  text[text] --> st[SentenceTransformer]
  st --> vec[vector]
  vec --> store[(vector store)]

The Code samples tab shows encoding text and a built-in similarity search — pick from the selector to compare.

When to use it

Choose Sentence Transformers when you want embeddings computed locally — for cost, privacy, or offline use — instead of calling a hosted embedding API.

Example

Python · encode

When to use

Reach for it to compute embeddings locally — no provider key, works offline.

How it flows

encode loads the Hugging Face model and returns a NumPy array of vectors you can store or compare.

Structure

flowchart LR
  text[text] --> st[SentenceTransformer]
  st --> vec[vector]
  vec --> use[store / compare]

from sentence_transformers import SentenceTransformer

# bge-m3 is multilingual and runs on CPU — no API key needed.
model = SentenceTransformer("BAAI/bge-m3")

vectors = model.encode(["how do I reset my password?", "비밀번호 재설정 방법"])
print(vectors.shape)  # (2, 1024)

When to use

Use the built-in similarity for quick semantic search without a vector database.

How it flows

Embed the corpus and the query, then rank documents by cosine similarity.

Structure

flowchart LR
  docs[documents] --> emb[embeddings]
  q[query] --> qe[query vector]
  emb --> sim[cosine similarity]
  qe --> sim
  sim --> top[best match]

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("BAAI/bge-m3")
docs = ["cats are great pets", "the stock market fell", "kittens love to play"]
doc_emb = model.encode(docs)

query = model.encode("tell me about cats")
scores = model.similarity(query, doc_emb)[0]  # cosine similarity
print(docs[int(scores.argmax())])  # -> a cat-related sentence

Example

Sentence Transformers — local embeddings + similarity

A real, runnable mini-project. Download it and run with Docker.

View samples/sentence-transformers_1 on GitHub ↗

June 28, 2026

Sentence Transformers — local embeddings + similarity

A tiny Sentence Transformers demo: it embeds a few documents and a query on CPU, then ranks the documents by cosine similarity. No API key — the model runs locally.

Heavy image. Pulls PyTorch (~several GB), so the first build is slow.

Run with Docker

cd samples/sentence-transformers_1
docker build -t aas-sentence-transformers .
docker run --rm aas-sentence-transformers "tell me about young cats"

The first run downloads the model (all-MiniLM-L6-v2, ~80 MB), so it needs internet once. Swap in BAAI/bge-m3 in app.py for a multilingual model.

Run with Docker (in a devcontainer with DooD)

In a dev container that talks to the host Docker daemon (Docker-outside-of-Docker), the attached docker run may print nothing even though it ran — the live attached stream drops output over the VM boundary. Run detached and read the logs:

cd samples/sentence-transformers_1
docker build -t aas-sentence-transformers .
docker logs -f "$(docker run -d aas-sentence-transformers "tell me about young cats")"

Run locally

cd samples/sentence-transformers_1
pip install -r requirements.txt
python app.py "tell me about young cats"

Example run

Embeddings are deterministic (fixed model + input), so the ranking is the same every run — only the first run downloads the model.

nearest to 'tell me about young cats' by cosine similarity:
  0.5793  A kitten is a young cat that loves to play and nap.
  0.4883  Cats are independent pets that purr and chase laser pointers.
  0.0734  A rocket is propelled by ejecting exhaust gas at high speed.

Files

app.py

"""Sentence Transformers: local embeddings + similarity search, no API key.

Embeds a few documents and a query with a model running on CPU, then ranks the
documents by cosine similarity. This is the local, provider-independent half of
a RAG pipeline.

Run:
    docker build -t aas-sentence-transformers .
    docker run --rm aas-sentence-transformers "tell me about young cats"
"""

import sys

from sentence_transformers import SentenceTransformer

DOCS = [
    "Cats are independent pets that purr and chase laser pointers.",
    "The stock market fell sharply today.",
    "A kitten is a young cat that loves to play and nap.",
    "A rocket is propelled by ejecting exhaust gas at high speed.",
]


def main() -> None:
    query = " ".join(sys.argv[1:]) or "tell me about young cats"

    # all-MiniLM-L6-v2 is small and fast; swap in BAAI/bge-m3 for multilingual.
    model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

    doc_emb = model.encode(DOCS)
    query_emb = model.encode(query)
    scores = model.similarity(query_emb, doc_emb)[0]  # cosine similarity

    print(f"nearest to {query!r} by cosine similarity:")
    for doc, score in sorted(zip(DOCS, scores.tolist()), key=lambda x: -x[1])[:3]:
        print(f"  {score:.4f}  {doc}")


if __name__ == "__main__":
    main()

Dockerfile

# Image for the Sentence Transformers sample. Pulls PyTorch, so it is several GB
# and the first build is slow. No API key — the model runs on CPU.
FROM python:3.13-slim

ENV PYTHONUNBUFFERED=1

WORKDIR /app

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY app.py .

ENTRYPOINT ["python", "app.py"]

Related tools

Overview

When to use it

Pricing

When to use

How it flows

When to use

How it flows

Run with Docker

Run with Docker (in a devcontainer with DooD)

Run locally

Example run