Skip to content

MIGRATION.md

Latest commit

 

History

History
55 lines (43 loc) · 3.28 KB

MIGRATION.md

File metadata and controls

55 lines (43 loc) · 3.28 KB
Raw

Migration Plan — 0.1.x → 0.2.0 (AI workflow framework)

This is the working migration plan for turning Entropia Riko from a node-editor prototype into a reproducible AI workflow framework. The UI is frozen — all changes land in the backend (core/runtime/nodes/server), examples, and docs.

1. Current state (0.1.x)

Already present and reusable:

  • Node contract — BaseNode + NodeInput/NodeOutput/Parameter + @register; ~200 nodes.
  • IR — TensorValue with kinds scalar | tensor | image_tensor | model | text | json; GraphDocument (.riko/.ric).
  • Runtime — registry, executor (validate → topo sort → execute), codegen (PyTorch/Keras export), trainer (train + save state_dict), subgraph (run/import).
  • Model I/O — save_model / model_loader (safetensors + torch), /api/train?save_path=.
  • Server — modular routers (health/plugins/graph/files/train/project/fs), project mini file-manager with a working-directory tree.
  • UI — decoupled into entropia-template-ui (npm) + thin frontend/ entry.

2. Gaps to close

Area Gap
Port types No strong typing: only 6 kinds, no connection compatibility check in validate.
Project system No project.riko manifest, no templates, no project browser APIs (create/open/scan/validate/migrate).
Data/model mgmt No DATASET/CHECKPOINT nodes, no asset registry, paths are absolute.
Reproducibility No experiment records, no unified bake/cache system.
Workflow mgmt Single-file graphs only; no multi-workflow workspace, versioning, or dependency graph.
Examples Mostly toy; need real AI pipelines (CV / diffusion / audio / video).

3. Phases (priority order)

  1. Port type system — extend TensorValue kinds + typed connection validation (foundation).
  2. Project system — project.riko manifest, requirements.riko, project templates, project browser APIs.
  3. Professional nodes — file_input, dataset, checkpoint_save/load (model/training/inference exist).
  4. Bake/Cache + Experiment — artifact ledger + experiment records for reproducibility.
  5. Workflow management — multi-workflow workspace + dependency/version metadata.
  6. Examples — replace toy graphs with real multimodal pipelines.
  7. Docs + release — README, docs/, changelog; publish PyPI/npm/GitHub 0.2.0.

4. Key file map

Change Files
Port kinds + compatibility core/tensor.py, runtime/executor.py (validate), new core/types.py (compat table)
Project manifest + templates new entropia_riko/project/ (manifest.py, templates.py, workspace.py), server/routers/project.py
Professional nodes nodes/torch_ops/ (dataset.py, checkpoint.py, file_io.py)
Bake/cache + experiment new entropia_riko/project/cache.py, entropia_riko/project/experiment.py
Examples examples/** (rewrite), templates/project/**
Docs README.md, docs/*.md, CHANGELOG.md

5. Acceptance criteria (0.2.0)

A new user can: create a project → pick an AI template → get a full folder structure + default workflows → load dataset → load model → train → save checkpoint → run inference → export results → reopen the project and continue, without manually managing files.