Agentic-Additive-Manufacturing-Process-Optimization

Process optimization of Inova Mk1 from SLS4All

Data layout (for analysts and downstream consumers)

The recorder writes to a Postgres database in the agentic-sls-postgres container and to image files under data/frames/. The canonical schema is server/src/db/schema.sql — read that file for the authoritative table definitions, column types, and indexes.

Tables

Table	Cardinality	Source
builds	one per print run	recorder, on session start/end
telemetry	~10 Hz per sensor	recorder polling the Inova plugin REST API
position_hf	~1 kHz	plugin WebSocket /movement/position/stream (firmware-native rate)
frames	~5 Hz (chamber)	recorder polling firmware HTTP endpoints; image bytes on disk
events	sparse	recorder, on state transitions

Sample rates are configured in .env (TELEMETRY_HZ, CAMERA_HZ, THERMAL_HZ). position_hf is event-driven from the firmware and not rate-limited by the recorder.

Sensor identifiers and units

telemetry.sensor_id and telemetry.kind are free-form strings populated by the upstream Inova plugin. The exporter (see below) writes a data/exports/sensors.csv enumerating every distinct (sensor_id, kind) pair observed in the DB, with empty unit and description columns for human annotation. Treat that CSV as the authoritative units glossary once filled in.

Frames on disk

Image files live at data/frames/{build_id}/{ts_ms}_{kind}.{ext} where kind is one of chamber, galvo, or thermal. The frames table holds one row per file with (build_id, ts, kind, path); the path column points back to the file on disk.

Flat-file exports

sls-export (agentic_sls/pipeline/export.py) dumps the Postgres tables to flat files under data/exports/ so downstream consumers can work without a running container. Layout:

data/exports/
  builds.jsonl
  events.jsonl                  (embedding column dropped)
  frames.jsonl                  (embedding column dropped)
  build_to_inova_session.csv    (sidecar; manual mapping to Agentic-SLS-Database)
  sensors.csv                   (sidecar; (sensor_id, kind) glossary)
  telemetry/{build_id}.parquet
  position_hf/{build_id}.parquet

The exporter is idempotent: existing parquet files are skipped unless --force, and sidecar CSVs preserve hand-filled values across re-runs. By default only completed builds (ended_at IS NOT NULL) are exported; pass --include-active to also export the open build.

uv run sls-export
uv run sls-export --builds 1,2
uv run sls-export --out /tmp/exports --force

Dependencies come from the root pyproject.toml (uv sync).