Storage

Manor AI uses PostgreSQL, Redis, MinIO, and JuiceFS together. Each holds a different kind of state; understanding the split is what makes backups, scaling, and debugging tractable.

PostgreSQL

PostgreSQL (16, with pgvector) stores all relational application data: users, entities, workspaces, tasks, goals, documents metadata, conversations, workflow definitions and runs, and the document_chunks table that holds chunk embeddings for semantic search.

• The stack ships pgvector/pgvector:pg16; a vanilla PostgreSQL without the pgvector extension will not work.
• Back up before every upgrade — see Backup and Restore.
• Watch connection counts if you raise API_WORKERS or pool sizes; see the sizing note in Configuration.

Redis

One Redis instance serves four separate roles on separate logical databases:

DB	Role
/0	Application cache, rate limits, presence (REDIS_URL)
/1	JuiceFS metadata — durable, must be backed up
/2	Celery broker
/3	Celery result backend

Because /1 holds filesystem metadata, this Redis is not a disposable cache. REDIS_MAXMEMORY defaults to 0 (unlimited) with noeviction — if you bound memory, never use an eviction policy that could evict JuiceFS keys; move JuiceFS metadata to its own Redis first.

MinIO

MinIO provides S3-compatible object storage for uploaded files, document assets, generated media artifacts, and JuiceFS data blocks (bucket configured via MINIO_BUCKET / JUICEFS_BUCKET).

JuiceFS

JuiceFS gives Manor AI a POSIX filesystem per entity — the surface agent file tools read and write — backed by Redis metadata (/1) and MinIO object blocks. The juicefs-init one-shot service formats and mounts it on stack start; the API and worker containers mount it at MANOR_FS_ROOT (default /mnt/manor) over FUSE.

Consequences:

• Files are not readable directly from the MinIO bucket — blocks only make sense through JuiceFS with its metadata.
• Metadata (Redis /1) and blocks (MinIO) must be backed up as a pair.
• The workers stop with a 30-second grace period so the FUSE layer can flush its write cache to MinIO; avoid SIGKILL.

Embeddings (Ollama)

Document RAG uses a local Ollama service with the mxbai-embed-large model (1024 dimensions), preloaded by ollama-init into the ollama_data volume on first boot. First startup downloads the model, so allow extra time. Without the model, embedding calls fail and search degrades to text-only.

Production Notes

• Use persistent volumes for every stateful service (pg_data, redis_data, minio_data, ollama_data).
• Monitor disk usage — generated media and document uploads accumulate in MinIO.
• Back up PostgreSQL, MinIO, and Redis /1 together; keep filesystem and database snapshots from drifting apart.