Skip to content

Latest commit

 

History

History
40 lines (32 loc) · 2.19 KB

File metadata and controls

40 lines (32 loc) · 2.19 KB

Development Plan

Current Focus

  1. Keep the local-first document automation workflow stable.
  2. Improve Workflow Builder creation, validation, and execution visibility.
  3. Improve schema editing and AI-assisted draft generation.
  4. Keep batch and workflow execution reliable at large document counts.
  5. Keep public docs and examples aligned with runnable code.

Recently Landed

  • Workflow Builder save fallback for stale drafts.
  • Compact execution status panel and removal of the canvas run banner.
  • AI workflow draft generation from up to 10 sample images.
  • Workflow Builder node asset editing now covers schema fields, classifier classes, and required-field checklists with module-style table editors.
  • Blank KIE nodes can now start a new schema draft directly inside Workflow Builder.
  • Draft materialization for schema/checklist assets during workflow save.
  • VLM settings now support OpenAI-compatible base_url, local no-key endpoints, structured inference parameters with reasoning/thinking off by default, request timeout tuning, and KIE field group tuning.
  • KIE result review marks low-confidence model output as needs_review and keeps frontend polling open for long local VLM inference.
  • Local OpenAI-compatible VLM calls use a raw JSON fallback path because many local servers accept json_schema parameters without reliably enforcing valid structured output.
  • README workflow screenshots are generated from the PoC UI smoke flow.
  • Public repository scope cleaned to focus on product features.

Next Product Tasks

  • Add stronger draft preview before applying AI-generated workflows.
  • Add fixture-based UI checks for AI workflow draft application.
  • Improve export preset editing inside result review.
  • Add more sample document templates with clear source notes.
  • Add explicit cancel/retry controls for long-running single KIE jobs.
  • Add an operator-visible diagnostic panel for local VLM raw response parse failures.

Working Rules

  • Read code and docs before changing behavior.
  • Keep changes narrowly scoped.
  • Prefer existing helpers and patterns.
  • Verify with backend tests, frontend build, and targeted smoke checks.
  • Keep public-facing docs free of service-only implementation details.