example.env

# IMPORTANT: Values must not be wrapped in quotes. This file is passed through to the
# containers as is, and 'docker stack deploy' treats quotes as part of the value.

# The environment setting.
# Use 'development' for local development, and 'production' in production.
ENVIRONMENT=development

# Ports that will be mapped to the host during development.
WEB_DEV_PORT=8000
POSTGRES_DEV_PORT=5432

# Ports that will be mapped to the host during production.
WEB_HTTP_PORT=80
WEB_HTTPS_PORT=443

# The Docker image to use for the app services.
# Defaults to the official GHCR image. Override this to use a locally built image
# (e.g. RADIS_IMAGE=radis-staging:latest for a staging deployment).
# RADIS_IMAGE=ghcr.io/openradx/radis:latest

# Override the Docker Swarm stack name. Defaults to {project_name}_{environment}
# (e.g. radis_dev, radis_prod). Also used to derive unique session and CSRF cookie names
# to prevent cookie conflicts when running multiple stacks on the same host.
# Set explicitly for additional deployments like staging.
# STACK_NAME=radis_staging

# Django debug settings (only used in development).
FORCE_DEBUG_TOOLBAR=true
REMOTE_DEBUGGING_ENABLED=false
REMOTE_DEBUGGING_PORT=5678

# The Django secret key used for cryptographic signing.
# IMPORTANT: Use a unique and secure key in production!
DJANGO_SECRET_KEY=your_django_secret_key_here

# The Postgres database password (only used in production).
POSTGRES_PASSWORD=your_postgres_password_here

# Miscellaneous Django security settings.
DJANGO_ALLOWED_HOSTS=localhost,127.0.0.1
DJANGO_CSRF_TRUSTED_ORIGINS=
DJANGO_INTERNAL_IPS=127.0.0.1

# Redirect all HTTP requests to HTTPS (only used in production).
DJANGO_SECURE_SSL_REDIRECT=true

# The salt that is used for hashing tokens in the token authentication app.
# Cave, changing the salt after some tokens were already generated makes them all invalid!
TOKEN_AUTHENTICATION_SALT=your_token_authentication_salt_here

# Email configuration.
# The email address that is used for sending emails to the users and critical errors
# to the admins. The smtp server is only used in production. In development the emails
# are just logged to the console.
DJANGO_SERVER_EMAIL=server@example-project.example
DJANGO_EMAIL_URL=smtp://localhost:25

# The Django server admins that will receive critical error notifications.
# Also used by django-registration-redux to send account approval emails to.
DJANGO_ADMIN_EMAIL=admin@radis.example
DJANGO_ADMIN_FULL_NAME=RADIS Admin

# A support Email address that is presented to the users where they can get support.
SUPPORT_EMAIL=support@radis.example

# A superuser that will have access to the Django admin interface.
# Optionally with a provided auth token for the API.
SUPERUSER_USERNAME=superuser
SUPERUSER_EMAIL=superuser@radis.example
SUPERUSER_PASSWORD=your_superuser_password_here
SUPERUSER_AUTH_TOKEN=your_superuser_auth_token_here

# Location of the backup folder.
BACKUP_DIR=./.docker-data/backups

# Enable the daily database backup periodic task.
# Set to "false" to no-op the shared backup_db task (e.g. in test environments).
BACKUP_ENABLED=true

# Cron schedule for the shared backup_db periodic task.
BACKUP_CRON=0 3 * * *

# Site information that is synced to the database and used by the sites framework.
SITE_NAME=RADIS
SITE_DOMAIN=localhost

# Settings for SSL encryption (only used in production).
# SSL_HOSTNAME and SSL_IP_ADDRESSES are used to generate self-signed certificates
# with 'uv run cli generate-certificate-files', but you can also provide both files
# on your own.
SSL_HOSTNAME=localhost
SSL_IP_ADDRESSES=127.0.0.1
SSL_SERVER_CERT_FILE=./cert.pem
SSL_SERVER_KEY_FILE=./key.pem
SSL_SERVER_CHAIN_FILE=./chain.pem

# The timezone used by the server.
TIME_ZONE=Europe/Berlin

# LLM configuration.
# RADIS runs no inference server of its own. All inference is sent to the OpenAI-compatible
# endpoint below, which must support the OpenAI API (beta) client and structured outputs.
#
# The default expects a provider running on the Docker host (e.g. Ollama, LM Studio,
# llama.cpp). Use 'uv run cli get-host-ip' if host.docker.internal cannot be resolved.
# On Linux, that provider must listen on more than loopback, otherwise the containers get a
# connection refused. Ollama has no authentication, so prefer binding it to the Docker bridge
# (OLLAMA_HOST=172.17.0.1 for the default docker0) over every interface (0.0.0.0). See
# docs/dev-docs/contributing.md.
#
# IMPORTANT: host.docker.internal is a development convenience and is NOT injected into the
# production stack. A production deployment must point this at an endpoint its nodes can
# actually reach, or every LLM job fails with a connection error.
# See docs/dev-docs/contributing.md for how to run Ollama on your machine.
LLM_BASE_URL=http://host.docker.internal:11434/v1
# Not all providers really use that key, but the OpenAI API client needs it to be set.
LLM_API_KEY=whatever
# The model every feature uses unless it has its own setting below. Must name a model the
# endpoint above serves (an Ollama tag when using Ollama). The default is small and fast:
# enough to check that everything works, not to judge the results by.
#
# Request parameters can be appended as a query string. They are merged into the request
# body, so both standard OpenAI fields (temperature, top_p, seed) and provider extensions
# work, and a dotted key becomes a nested object:
#   LLM_DEFAULT_MODEL=qwen3:8b?temperature=0&chat_template_kwargs.enable_thinking=false
# Values are read as JSON where possible, so temperature=0 sends a number while
# reasoning_effort=none sends the string "none".
#
# 'reasoning_effort=none' below turns off "thinking", which makes reasoning models much
# faster and their structured output cleaner. That spelling is Ollama's and matches the
# default endpoint; vLLM and SGLang want chat_template_kwargs.enable_thinking=false
# instead, as each server silently ignores the other's. Drop it for providers that reject
# parameters they don't know, such as OpenAI and Azure OpenAI.
LLM_DEFAULT_MODEL=qwen3.5:0.8b?reasoning_effort=none

# Per-feature overrides. Leave blank to use LLM_DEFAULT_MODEL. Useful to spend a stronger
# (or cheaper) model where it matters, e.g. a large model for labeling and a fast one for
# chat. Each takes the same 'model?param=value' form.
LLM_CHATS_MODEL=
LLM_QUERY_GENERATION_MODEL=
LLM_EXTRACTIONS_MODEL=
LLM_SUBSCRIPTIONS_MODEL=
LLM_LABELING_MODEL=
#
# The LLM request timeout, rate-limit gate, and transient-retry knobs have sensible defaults
# in settings (LLM_REQUEST_TIMEOUT_SECONDS, LLM_RATE_LIMIT_*, LLM_TRANSIENT_RETRY_*);
# override them here only if needed.

# The language of the example reports that will be seeded to the development database.
# Possible values are 'en' or 'de'.
EXAMPLE_REPORTS_LANGUAGE=en

# Dedicated LLM configuration for generating sample reports via the CLI. These do not
# impact inference. The provider must support the OpenAI API client.
REPORT_LLM_MODEL_NAME=gemma3
REPORT_LLM_PROVIDER_URL=http://host.docker.internal:11434/v1
# API key (if required by the generation provider). Only used with
# 'cli generate-example-reports'.
REPORT_LLM_PROVIDER_API_KEY=ollama

# Embedding service (OpenAI-compatible /v1/embeddings), used by hybrid search.
#
# Leave EMBEDDINGS_MODEL empty to run full-text search only — reports stay fully
# searchable, no embedding jobs are queued and the service is never called.
#
# The endpoint and key default to LLM_BASE_URL and LLM_API_KEY, which is what you want
# when one provider serves both (OpenAI, Ollama, an LLM gateway). Set them only when
# embeddings live somewhere else — a self-hosted vLLM or SGLang serves one model per
# process, so there the embedding model is a second server.
#EMBEDDINGS_BASE_URL=
#EMBEDDINGS_API_KEY=

# The model, as 'model[?param=value&...]' — the same spec the LLM_*_MODEL settings take.
# Parameters are merged into the request body, so a provider that supports OpenAI's
# 'dimensions' can be asked for the width directly:
#   EMBEDDINGS_MODEL=Qwen/Qwen3-Embedding-4B
#   EMBEDDINGS_MODEL=text-embedding-3-large?dimensions=1024
# For Ollama in dev: ollama pull dengcao/Qwen3-Embedding-4B:Q5_K_M
EMBEDDINGS_MODEL=

# Vector dimension. Schema-coupled: changing this after deploy requires dropping the
# embedding column, re-migrating, and running `./manage.py embed_pending`. When the
# model spec also sets 'dimensions', the two must agree (checked at startup).
EMBEDDINGS_DIM=1024

# Instruction prefix prepended to search queries before embedding. Model-specific:
# Qwen3-Embedding wants one, text-embedding-3 wants none. Not a request parameter, so
# it is not part of the model spec. The default contains a literal newline
# ("...reports.\nQuery: "); to reproduce a multi-line instruction here, wrap the value
# in double quotes so the \n is escape-processed (a single-quoted or unquoted value
# keeps the literal backslash-n instead of a newline).
#EMBEDDINGS_QUERY_INSTRUCTION=

# Throughput tuning (all optional).
# EMBEDDINGS_REQUEST_TIMEOUT_SECONDS defaults to LLM_REQUEST_TIMEOUT_SECONDS (itself 60
# by default), not to a fixed 60 — raising the LLM timeout for a slow endpoint raises
# this one too unless set here explicitly.
#EMBEDDINGS_REQUEST_TIMEOUT_SECONDS=
#EMBEDDINGS_BATCH_SIZE=200
#EMBEDDINGS_SUBJOB_SIZE=1000
#EMBEDDINGS_WORKER_CONCURRENCY=2

# How long a search query's embedding stays cached, in seconds (default 900 = 15 min).
# This is the knob to reach for right after a model or provider swap: cached query
# vectors from the old model can otherwise keep serving stale results for up to this
# long. Lower it temporarily, or clear the cache, if you need the swap to take effect
# immediately.
#EMBEDDINGS_QUERY_CACHE_TIMEOUT_SECONDS=900

# Auto-labeling (radis.labels)
# Both prompts have sensible built-in defaults; override only to customize.
# LABELING_SYSTEM_PROMPT=...        # generic per-label prompt; only $report is substituted
# LABELING_GATE_SYSTEM_PROMPT=...   # generic group gate (Yes/No) prompt
LABELING_JOB_PRIORITY=1
LABELING_TASK_BATCH_SIZE=100
LABELING_LLM_CONCURRENCY_LIMIT=2
LABELING_GATE_BATCH_SIZE=10
LABELING_SCAN_CRON=0 2 * * *

# Recovery of analysis tasks left IN_PROGRESS by a killed worker.
# The grace period must never be set below 30 (Procrastinate's own stall threshold).
ANALYSIS_STALLED_WORKER_GRACE_SECONDS=30
ANALYSIS_SWEEP_CRON=* * * * *

# OpenTelemetry Configuration
# Set this to the OTLP HTTP endpoint of the centralized openradx-observability stack.
# See https://github.com/openradx/openradx-observability for setup instructions.
OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector.local:4318

# Docker swarm mode does not respect the Docker Proxy client configuration
# (see https://docs.docker.com/network/proxy/#configure-the-docker-client),
# but we can set those environment variables manually.
# Malke sure to use .local in NO_PROXY as otherwise the communication with
# the other services will not work.
# HTTP_PROXY="http://user:pass@myproxy.net:8080"
# HTTPS_PROXY="http://user:pass@myproxy.net:8080"
# NO_PROXY="localhost,.local"