diff --git a/embedding_service/README.md b/embedding_service/README.md
new file mode 100644
index 0000000..2a0ced3
--- /dev/null
+++ b/embedding_service/README.md
@@ -0,0 +1,104 @@
+# Embedding service
+
+A small FastAPI service that turns texts into 768-dimensional vectors with
+[`intfloat/multilingual-e5-base`](https://huggingface.co/intfloat/multilingual-e5-base).
+The portfolio assistant uses it to index the knowledge base and to embed
+visitor questions.
+
+It runs as its own process so the model is loaded once. Inside Django, every
+Gunicorn worker would load its own copy.
+
+## Setup
+
+The service has its own virtual environment, separate from the Django one:
+
+```bash
+cd embedding_service
+python -m venv .venv
+```
+
+Activate it. **Windows (PowerShell):**
+
+```powershell
+.\.venv\Scripts\Activate.ps1
+```
+
+**macOS / Linux:**
+
+```bash
+source .venv/bin/activate
+```
+
+Then install the dependencies:
+
+```bash
+python -m pip install --upgrade pip
+python -m pip install -r requirements.txt
+```
+
+`requirements.txt` pulls the CPU-only PyTorch build from the PyTorch package
+index. The default Linux wheel on PyPI would add several gigabytes of CUDA
+libraries.
+
+## Running
+
+```bash
+uvicorn main:app --host 127.0.0.1 --port 8002
+```
+
+The first start downloads the model (about 1.1 GB) into the Hugging Face
+cache. Uvicorn only accepts connections once the model is loaded. The process
+uses about 1.8 GB of memory after the first request.
+
+## API
+
+Interactive documentation is available at `/docs` while the service runs.
+
+### `GET /health`
+
+```json
+{"status": "ok", "model": "intfloat/multilingual-e5-base"}
+```
+
+### `POST /embed`
+
+```json
+{"kind": "query", "texts": ["Wie deployt Benjamin seine Projekte?"]}
+```
+
+- `kind` is `query` for search questions and `passage` for knowledge base
+  sections. The service adds the prefix e5 was trained with. Without it the
+  model still returns vectors, but retrieval quality drops without any error.
+- `texts` holds 1 to 64 non-blank strings.
+
+The response contains one vector per text, in input order:
+
+```json
+{"model": "intfloat/multilingual-e5-base", "dimensions": 768, "vectors": [[0.026, 0.063, ...]]}
+```
+
+Vectors are normalized to length 1, so their dot product is the cosine
+similarity. e5 scores cluster between roughly 0.7 and 1.0 even for unrelated
+texts. Only their order is meaningful, not their absolute value.
+
+Invalid input returns `422`. So does any text longer than the model's limit
+of 512 tokens, which the model would otherwise cut off silently.
+`detail.indexes` lists the positions of the affected texts.
+
+## Smoke test
+
+With the service running:
+
+```bash
+python smoke_test.py
+```
+
+It checks the vector size and that related passages, German and English,
+score higher than an unrelated one. It exits with `0` on success and `1` on
+failure, and needs nothing beyond the Python standard library.
+
+## Why this folder is not a Python package
+
+There is deliberately no `__init__.py`. Without it, Django's test runner and
+coverage ignore this folder, which matters because FastAPI and PyTorch are not
+installed in the Django environment or in CI.
diff --git a/embedding_service/main.py b/embedding_service/main.py
new file mode 100644
index 0000000..633a6a0
--- /dev/null
+++ b/embedding_service/main.py
@@ -0,0 +1,85 @@
+"""HTTP service that turns texts into multilingual-e5-base embeddings."""
+
+from contextlib import asynccontextmanager
+from typing import Annotated, Literal
+
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel, Field, StringConstraints
+from sentence_transformers import SentenceTransformer
+
+MODEL_NAME = "intfloat/multilingual-e5-base"
+MAX_TEXTS_PER_REQUEST = 64
+
+ml_models = {}
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Load the model once at startup instead of once per request."""
+    ml_models["embedder"] = SentenceTransformer(MODEL_NAME, device="cpu")
+    yield
+    ml_models.clear()
+
+
+app = FastAPI(title="Embedding service", lifespan=lifespan)
+
+Text = Annotated[str, StringConstraints(strip_whitespace=True, min_length=1)]
+
+
+class EmbedRequest(BaseModel):
+    """Texts to embed, marked as search queries or knowledge passages."""
+
+    kind: Literal["query", "passage"]
+    texts: Annotated[
+        list[Text], Field(min_length=1, max_length=MAX_TEXTS_PER_REQUEST)
+    ]
+
+
+class EmbedResponse(BaseModel):
+    """One normalized vector per input text, in input order."""
+
+    model: str
+    dimensions: int
+    vectors: list[list[float]]
+
+
+@app.get("/health")
+def health():
+    """Report readiness. Uvicorn only answers once the model is loaded."""
+    return {"status": "ok", "model": MODEL_NAME}
+
+
+def _reject_truncated(embedder, texts):
+    """Refuse texts the model would silently cut off at its token limit."""
+    token_counts = [
+        len(ids) for ids in embedder.tokenizer(texts)["input_ids"]
+    ]
+    too_long = [
+        index for index, count in enumerate(token_counts)
+        if count > embedder.max_seq_length
+    ]
+    if too_long:
+        raise HTTPException(
+            status_code=422,
+            detail={
+                "error": "Text exceeds the model's token limit.",
+                "limit": embedder.max_seq_length,
+                "indexes": too_long,
+            },
+        )
+
+
+@app.post("/embed")
+def embed(request: EmbedRequest) -> EmbedResponse:
+    """Return one normalized embedding per text."""
+    embedder = ml_models["embedder"]
+    # e5 was trained with these prefixes. Without them it still returns
+    # vectors, but retrieval quality drops without any error.
+    texts = [f"{request.kind}: {text}" for text in request.texts]
+    _reject_truncated(embedder, texts)
+    vectors = embedder.encode(texts, normalize_embeddings=True)
+    return EmbedResponse(
+        model=MODEL_NAME,
+        dimensions=vectors.shape[1],
+        vectors=vectors.tolist(),
+    )
diff --git a/embedding_service/requirements.txt b/embedding_service/requirements.txt
new file mode 100644
index 0000000..331269c
--- /dev/null
+++ b/embedding_service/requirements.txt
@@ -0,0 +1,53 @@
+# CPU-only PyTorch. The +cpu build only exists on this index; the default
+# Linux wheel on PyPI pulls in several gigabytes of CUDA libraries.
+--extra-index-url https://download.pytorch.org/whl/cpu
+
+annotated-doc==0.0.5
+annotated-types==0.8.0
+anyio==4.15.1
+certifi==2026.7.22
+click==8.5.0
+cloudpickle==3.1.2
+colorama==0.4.6
+fastapi==0.141.1
+filelock==3.32.3
+fsspec==2026.7.0
+h11==0.16.0
+hf-xet==1.6.0
+httpcore==1.0.9
+httpx==0.28.1
+huggingface_hub==1.33.0
+idna==3.20
+Jinja2==3.1.6
+joblib==1.6.0
+markdown-it-py==4.2.0
+MarkupSafe==3.0.3
+mdurl==0.1.2
+mpmath==1.3.0
+narwhals==2.26.0
+networkx==3.6.1
+numpy==2.5.3
+packaging==26.3
+pydantic==2.13.5
+pydantic_core==2.46.5
+Pygments==2.21.0
+PyYAML==6.0.3
+regex==2026.9.10
+rich==15.0.0
+safetensors==0.8.0
+scikit-learn==1.9.1
+scipy==1.18.1
+sentence-transformers==6.1.0
+setuptools==84.0.0
+shellingham==1.5.4
+starlette==1.7.0
+sympy==1.14.0
+threadpoolctl==3.7.0
+tokenizers==0.23.2
+torch==2.14.0+cpu
+tqdm==4.70.1
+transformers==5.17.0
+typer==0.27.2
+typing-inspection==0.4.4
+typing_extensions==4.16.0
+uvicorn==0.53.0
diff --git a/embedding_service/smoke_test.py b/embedding_service/smoke_test.py
new file mode 100644
index 0000000..29fc13f
--- /dev/null
+++ b/embedding_service/smoke_test.py
@@ -0,0 +1,75 @@
+"""Smoke test for a running embedding service.
+
+Checks the vector size and that related passages, in German and in English,
+score higher than an unrelated one. Uses only the standard library, so it
+runs with any Python 3, including on the server after a deploy.
+"""
+
+import json
+import sys
+from urllib.request import Request, urlopen
+
+BASE_URL = "http://127.0.0.1:8002"
+EXPECTED_DIMENSIONS = 768
+
+QUERY = "Wie bringt Benjamin seine Projekte auf den Server?"
+PASSAGES = {
+    "related, German": (
+        "Das Deployment läuft automatisch über GitHub Actions "
+        "auf einen eigenen VPS."
+    ),
+    "related, English": (
+        "Deployment runs automatically through GitHub Actions "
+        "to a private VPS."
+    ),
+    "unrelated": (
+        "Der Pokedex zeigt Pokémon mit ihren Typen und Werten "
+        "in einer Kartenansicht."
+    ),
+}
+
+
+def embed(kind, texts):
+    """Send texts to the service and return their vectors."""
+    body = json.dumps({"kind": kind, "texts": texts}).encode("utf-8")
+    request = Request(
+        f"{BASE_URL}/embed",
+        data=body,
+        headers={"Content-Type": "application/json"},
+    )
+    with urlopen(request, timeout=30) as response:
+        return json.load(response)["vectors"]
+
+
+def cosine(a, b):
+    """Cosine similarity of two normalized vectors, their dot product."""
+    return sum(x * y for x, y in zip(a, b))
+
+
+def main():
+    """Print the similarity scores and return 1 if a check fails."""
+    [query] = embed("query", [QUERY])
+    vectors = embed("passage", list(PASSAGES.values()))
+    scores = {
+        label: cosine(query, vector)
+        for label, vector in zip(PASSAGES, vectors)
+    }
+
+    for label, score in scores.items():
+        print(f"{score:.4f}  {label}")
+
+    failures = []
+    if len(query) != EXPECTED_DIMENSIONS:
+        failures.append(f"expected {EXPECTED_DIMENSIONS} dimensions")
+    for label in ("related, German", "related, English"):
+        if scores[label] <= scores["unrelated"]:
+            failures.append(f"'{label}' does not beat 'unrelated'")
+
+    for failure in failures:
+        print(f"FAIL: {failure}")
+    print("FAILED" if failures else "OK")
+    return 1 if failures else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())