Streamlit form schema

ui.streamlit describes a generic launch form. DrtmlStepLauncher renders it, reads defaults from merged params, and returns a params overlay. The step never imports Streamlit.

Declared title / submit_label may exist in DRTML for future shell use; the current Streamlit shell still owns its own page chrome and submit buttons. Rely on fields and generic control blocks for live behavior.

Complete example

ui:
  streamlit:
    enabled: true

    fields:
      corpus_uri:
        type: text
        label: Corpus URI
        param: domain.corpus_uri
        required: true
        help: Local path or configured object URI

      upstream_runs:
        type: runs_picker
        label: Upstream run IDs
        param: mapping.upstream_runs

      max_docs:
        type: int
        label: Maximum documents
        param: runtime.max_docs
        min: 0

      threshold:
        type: float
        label: Score threshold
        param: domain.score_threshold
        min: 0.0
        step: 0.05

      normalize:
        type: checkbox
        label: Normalize text
        param: domain.normalize

      language:
        type: select
        label: Language
        param: domain.language
        options: [en, de, fr]

    io_controls:
      enabled: true

    runtime_controls:
      enabled: true

    parallelism_controls:
      enabled: true
      title: Partition workers
      backend_options: [inprocess, processes, ray]
      default_backend: inprocess
      backend_label: Dispatch backend
      workers_label: Worker count
      ray_address_label: Ray Client address

    flush_controls:
      from_dataset_output: true
      mode_param_template: flush.{dataset}.mode
      value_param_template: flush.{dataset}.value

    compute_controls:
      enabled: false

    embed_train_controls:
      enabled: false

Field types

text

Rendered with st.text_input; returns stripped text.

runs_picker

Rendered with st.text_area. List defaults are serialized as comma-separated text. Use it for upstream run selection fields interpreted by the corresponding runtime parser.

int / integer

Rendered with integer number_input. Supports min, default, label, help, and param.

float / number

Rendered with floating number_input. Supports min and step.

bool / boolean / checkbox

Rendered with st.checkbox.

select

Rendered with st.selectbox; requires non-empty options. An invalid default falls back to the first option.

Unknown types deliberately fall back to a text input.

Common field properties

• param: merged params key; defaults to the field id.
• label: visible label; defaults to field id.
• help: Streamlit help text.
• default: fallback only; params_defaults takes precedence.
• required: keep empty values in the overlay and let validation/runtime reject them. Optional empty values are omitted.

Use only namespaces allowed by DRTML validation. A control is valid only if a runtime/domain reader consumes its param.

Generated control groups

io_controls

Shown when the manifest has outputs. Currently exposes global parquet compression (storage.parquet_compression) from the framework codec catalog.

runtime_controls

Produces:

• runtime.resume_enabled
• runtime.progress_log_every_n_docs
• runtime.finalize_flush_every_n_docs
• runtime.document_batch_size
• runtime.corpus_batch_max_chars

Use it for partition-loop/corpus steps only. Disable it for modes that do not consume these params.

parallelism_controls

Produces:

• runtime.parallelism.backend
• runtime.parallelism.workers
• optional runtime.parallelism.ray_address

Configuration may customize labels/help/backend options and expose embed-specific inflight/feed-cap controls.

flush_controls

When from_dataset_output: true, the renderer enumerates output datasets and creates flush.<dataset>.mode / .value controls seeded from each DRTML flush block. Runtime applies them as StoragePlan overrides.

compute_controls

Current builtin renders SGNS settings: embedding method, dimension, epochs, learning rate, negative samples, seed, threads, minimum token count, pair radius/window, pair batch size, and pair-weight damping.

embed_train_controls

Renders one descriptor/material kind per run from configured kinds and writes the list-valued compute.embed_kinds (or configured kind_param).

Launch flow

manifest + params_defaults
  → streamlit_ui_spec
  → DrtmlStepLauncher.render_form
  → params overlay
  → Airflow/local launch request
  → merged_params_from_manifest
  → framework dispatch

The renderer does not execute the step and does not write datasets. It only constructs the same namespaced params overlay that any other integration can submit.

Extending form capabilities

For a generic field type:

1. define the YAML shape and namespace behavior;
2. add rendering to _render_field in integration/ui/drtml_form.py;
3. add manifest/UI validation;
4. ensure a runtime/domain consumer exists;
5. add form tests for defaults, empty/required behavior, and overlay type;
6. document the field here.

For a reusable group, create a focused renderer module, call it from DrtmlStepLauncher, and keep all step-specific names in DRTML configuration. Do not add custom forms to step packages.