Agentic RAG

Autumn is the Python RAG backend separate from the Aglaea/Cyrene stack. It is built around document collections, partitions, uploaded files, retrieval tools, and chat endpoints, with FastAPI as the application layer and LlamaIndex doing the retrieval orchestration.

Backend shape

The service keeps application state in PostgreSQL through SQLAlchemy and Alembic migrations, stores vectors in Qdrant, and can run with object storage and Redis-compatible infrastructure for the parts that need them. The API is organized around collections, files, tools, and health checks, which makes it easier to debug the pipeline than a single "chat" endpoint would be.

Retrieval work

Autumn exposes retrieval as explicit tools instead of hiding it behind one opaque prompt. It can retrieve semantically from collection content, use hybrid search where the source data supports it, and keep summaries or metadata available when a short answer needs more than nearest-neighbor chunks.

Deployment notes

• Containerized the API and backing services so the stack can run on a VPS instead of depending on a managed prototype environment.
• Cut operating costs by 60% through the lower-cost deployment path.
• Scoped Autumn to the Python/FastAPI backend while Cyrene owns the Go/Kafka/Qdrant indexing path used by Aglaea.