Open items after the Dagster / storage / inference cut. Jobs and sensors: Dagster. HTTP shapes: Inference. Inbox rules: Inbox.
| # | Status | Item | Issue | Doc |
|---|---|---|---|---|
| 1 | Done | First inbox run: sensors on, one finished zip, MinIO has materials only | — | Inbox |
| 2 | Done | Switch EMBEDDER / RERANKER to http and match the live API |
— | Inference |
| 3 | If the drop is a network share | Ignore zips still being copied (settle window) | #1 | Inbox |
| 4 | Later | Fetch from FTP / object store if you will not mount the share | #2 | Inbox |
Not bugs in the current cut. Do not undo per-zip jobs, Postgres/MinIO, or “zip stays on the inbox.”
| Fact | Why it matters |
|---|---|
| Local BGE still loads in the Dagster worker | default_limit: 4 caps ingest and extract together. Tighten in the UI if BGE RAM is tight. |
| HTTP embed is dense-only | Sparse slot is a placeholder. Hybrid lexical retrieval is weaker on EMBEDDER=http until the API returns sparse. #3 |
INGEST_DIRECTORY is not a Settings field |
Only Dagster reads it. mp ingest still wants an explicit path. |
$DAGSTER_HOME/dagster.yaml is a copy |
Recopy after template changes. pools may only have default_limit / granularity. |