Configuration reference

Set backend variables in .env or the process environment before startup. Restart the backend to apply startup settings. Process variables take precedence over load_dotenv() defaults. Protect .env and never treat an example file as proof a setting is implemented.

Backend settings

Variable	Default	Consumer / behavior
API_KEY	Must be provided	Shared bearer authentication; placeholders are not rejected consistently
ALLOWED_ORIGINS	localhost/127.0.0.1 on ports 3000 and 8000	CORS middleware; not network access control
OLLAMA_HOST	http://localhost:11434	Model client and generation HTTP endpoint
OLLAMA_MODEL	qwen3.5:2b	Model when routing is disabled or initialization fails
OLLAMA_TIMEOUT	900 seconds	Generation HTTP timeout
OLLAMA_THINK	false	Generation payload think option
OLLAMA_TEMPERATURE	0.7	Generation option
OLLAMA_CONTEXT_SIZE	8192	Generation context option
OLLAMA_MAX_TOKENS	2048	Generation prediction limit
ENABLE_MODEL_ROUTING	true	Initialize and use router
ENABLE_CONVERSATION_MEMORY	true	Store/retrieve conversation state; disable for first-use lab
CONVERSATION_RETENTION_DAYS	7	Expiry assigned to new query-created threads
CONVERSATION_DB_PATH	data/conversations.db relative to project code	SQLite path; module reads this before app calls load_dotenv()

For a custom database path, export it into the process environment before starting the service. Setting it only in .env can be too late because the database module is imported first.

Settings that do not do what their names suggest

Setting	Actual status
DEFAULT_MODEL	Read/logged; router rules still hard-code their default
RATE_LIMIT_*	Not consumed by backend rate limiting
CHROMA_DB_PATH, CHROMA_COLLECTION_NAME	Backend uses fixed relative path and collection names
API_HOST, API_PORT	Not wired into the checked-in startup commands; configure Uvicorn/service arguments
LOG_LEVEL	Backend logging is configured at INFO in code
BACKUP_RETENTION_DAYS, BACKUP_SCHEDULE	Do not assume the backup script reads these or installs a schedule

Resource limits belong to the actual container/service configuration. An environment variable alone does not enforce them. Script-specific settings should be checked in the script you invoke.

First-use lab choices

Use a unique API_KEY, ENABLE_CONVERSATION_MEMORY=false, and a localhost Ollama host. Keep routing enabled after installing both models. See installation for the explicit localhost server command.

Next: Model behavior and storage paths.