Settings Reference

Every AI_SDK_* Django setting, with its default and purpose. Most are optional: the SDK works with sensible defaults out of the box.

Settings are read via getattr(settings, ...) at call time (cached where noted), so none of them require a restart of the dev server beyond Django’s normal settings reload.

Agent Registration and Runtime Agents

Models

RAG and Vector Stores

Files and Uploads

MCP Servers

Integrations

Permissions

Workflows

Streaming and Titles

Tracing

Requires the opt-in django_ai_sdk.tracing app in INSTALLED_APPS.

One Haystack environment variable shapes what the tracer records. Unlike the settings above, it is read once, when haystack.tracing is first imported: it must be a real environment variable set before the process starts, and changing it needs a restart. Assigning it in settings.py does nothing.

Logging

Background Tasks

Document processing (memories/tasks.py) and workflow runs (workflows/tasks.py) are enqueued with django-tasks, which the SDK depends on directly. It is not an AI_SDK_* setting: the backend is Django’s TASKS setting, and which one you pick is a deployment decision the SDK does not make for you.

With no TASKS setting at all, django-tasks defaults to django_tasks.backends.immediate.ImmediateBackend - tasks run inline, in the calling request. That works out of the box and needs no extra package, but it means a document upload processes synchronously while the user waits.

For durable processing with a separate worker, install a backend that provides one. The database backend ships separately:

pip install django-tasks-db

INSTALLED_APPS = [
    # ...
    "django_tasks",
    "django_tasks_db",
]

TASKS = {
    "default": {
        "BACKEND": (
            "django_tasks.backends.immediate.ImmediateBackend"
            if DEBUG
            else "django_tasks_db.DatabaseBackend"
        ),
    }
}

Then run the worker: ./manage.py db_worker. The demo project uses exactly this split - immediate in DEBUG, the database backend otherwise.

Provider Credentials

Not AI_SDK_* prefixed, but part of the same surface: the generator factories read them at call time. Every key is optional - when one is unset the factory omits it, so Haystack falls back to its own environment variable of the same name.