diff --git a/SETUP.md b/SETUP.md
index fe36328..c2fda5b 100644
--- a/SETUP.md
+++ b/SETUP.md
@@ -54,7 +54,11 @@ Boot up the vector database and internal services:
 docker compose up -d
 ```
 > [!NOTE]
-> This spins up **Qdrant** (our semantic memory vector store) alongside necessary backend supporting services.
+> This spins up **Qdrant** (our semantic memory vector store) and the **Intelligence Layer**
+> (the Python NLP service that extracts entities/relationships into your knowledge graph).
+> The Intelligence Layer image is large on first build (bundled spaCy + sentence-transformers
+> models), so the initial `docker compose up` can take a few minutes. It's optional at runtime —
+> the agent calls it best-effort, so chat and memory keep working even while it's still starting.
 
 ### 4. Start Ollama
 Start your local inference engine:
diff --git a/docker-compose.yml b/docker-compose.yml
index 3da4005..baed4cb 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -44,6 +44,24 @@ services:
         wait $$OLLAMA_PID
       '
 
+  # Python Intelligence Layer (NLP): entity + relationship extraction that feeds
+  # the knowledge graph. The agent calls it best-effort, so the stack still works
+  # if this container is starting up or unavailable.
+  python-agent:
+    build:
+      context: ./services/python-agent
+      dockerfile: Dockerfile
+    container_name: torvaix-intelligence
+    ports:
+      - "8000:8000"
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 40s
+
   torvaix:
     build:
       context: .
@@ -57,6 +75,7 @@ services:
       - NODE_ENV=production
       - OLLAMA_URL=http://ollama:11434
       - QDRANT_URL=http://qdrant:6333
+      - PYTHON_SERVICE_URL=http://python-agent:8000
       - AGENT_PORT=3001
       - JWT_SECRET=${JWT_SECRET:-change-me-in-production}
       - TORVAIX_MODEL=${TORVAIX_MODEL:-llama3.2}
@@ -73,6 +92,8 @@ services:
         condition: service_healthy
       ollama:
         condition: service_started
+      python-agent:
+        condition: service_started
     restart: unless-stopped
     healthcheck:
       test: ["CMD", "curl", "-f", "http://localhost:3001/api/health"]
diff --git a/services/python-agent/.dockerignore b/services/python-agent/.dockerignore
new file mode 100644
index 0000000..d605235
--- /dev/null
+++ b/services/python-agent/.dockerignore
@@ -0,0 +1,8 @@
+__pycache__/
+*.pyc
+*.pyo
+.venv/
+venv/
+.pytest_cache/
+.mypy_cache/
+*.egg-info/
diff --git a/services/python-agent/Dockerfile b/services/python-agent/Dockerfile
new file mode 100644
index 0000000..2181f23
--- /dev/null
+++ b/services/python-agent/Dockerfile
@@ -0,0 +1,29 @@
+# Torvaix Intelligence Layer (NLP) — spaCy + sentence-transformers behind FastAPI
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# curl is used by the container healthcheck
+RUN apt-get update \
+  && apt-get install -y --no-install-recommends curl \
+  && rm -rf /var/lib/apt/lists/*
+
+# Install CPU-only PyTorch first so the requirements install below does NOT pull the
+# multi-GB CUDA build. This keeps the image ~800MB instead of ~2.5GB.
+RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Bake the models into the image so the first request is instant and works offline.
+RUN python -m spacy download en_core_web_sm \
+  && python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"
+
+COPY app ./app
+
+EXPOSE 8000
+
+HEALTHCHECK --interval=30s --timeout=10s --retries=5 --start-period=40s \
+  CMD curl -f http://localhost:8000/health || exit 1
+
+CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]