docker-compose.llm.yml

Rama main · réplica de lectura del repositorio original.

# OPTIONAL local LLM backend for soc-ai — Ollama serving the model, LiteLLM
# proxying it under the stable alias `soc-ai-analyst`.
#
# For the local route when you have no model server yet. Joins the main
# stack's Docker network so the app reaches it as http://litellm:4000 with
# nothing published to the host at all.
#
# Usage (from the repo root, main stack already up — it owns the network):
#   docker compose -f docker-compose.llm.yml up -d
#   docker compose -f docker-compose.llm.yml exec ollama ollama pull qwen3:14b
#   # then in .env:  LITELLM_BASE_URL=http://litellm:4000
#   # and qualify it:  docker compose exec soc-ai python -m soc_ai model-probe
#
# Ordering note: on a FRESH install, run ./setup.sh before this file and enter
# http://litellm:4000 at its gateway-URL prompt (accept the model-list warning
# — the host can't resolve that name yet); it seeds .env for you. Editing
# .env by hand afterward instead? Apply it with `docker compose up -d`, not
# `restart` — restart does not re-read .env.
#
# Model choice: qwen3:14b is a reasonable tool-calling default on 16 GB+ of
# RAM/VRAM; docs/LESSER_MODELS.md covers qualifying ANY candidate with
# model-probe before pointing triage at it. Swap the pull + litellm.llm.yaml
# entry freely.
#
# Network: `soc-ai_default` below is Compose's own default project network
# name, derived from the repo directory name (docker-compose.yml declares no
# explicit `networks:`, so Compose falls back to `<project>_default`, and the
# project name defaults to the directory basename — `soc-ai` for a normal
# clone). Cloned into a differently named directory, or already running the
# main stack under a custom COMPOSE_PROJECT_NAME? Export that same
# COMPOSE_PROJECT_NAME before running this file too, so `<project>_default`
# still resolves to the network the main stack actually created.
#
# Tags are PINNED (verified against the registries 2026-08-19). Never run
# :latest — it's an unaudited moving target.

services:

  ollama:
    image: ollama/ollama:0.32.14
    container_name: soc-ai-ollama
    volumes:
      - ollama_models:/root/.ollama
    restart: unless-stopped
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"
    # cap_drop ALL untested against ollama's runtime needs (GPU passthrough
    # uses device requests, not caps) — hardened to no-new-privileges;
    # tighten further if your deployment proves it out.
    security_opt:
      - no-new-privileges:true
    # No memory limit on purpose: a cap below the model's working set turns a
    # slow load into an OOM kill — size the host instead (14b Q4 ≈ 10 GB
    # working set).

  litellm:
    image: ghcr.io/berriai/litellm:v1.97.0
    container_name: soc-ai-litellm
    command: ["--config", "/etc/litellm/config.yaml", "--port", "4000"]
    volumes:
      - ./litellm.llm.yaml:/etc/litellm/config.yaml:ro,Z  # see docker-compose.yml's mounts note: SELinux relabel, no-op elsewhere
    depends_on:
      - ollama
    restart: unless-stopped
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"
    security_opt:
      - no-new-privileges:true
    cap_drop:
      - ALL

volumes:
  ollama_models:

networks:
  default:
    name: soc-ai_default
    external: true