Skip to Content
We are live but in Staging 🎉
ActionsCatalog

Actions — Catalog

The action set is server-driven — it is the Scriptum template catalog filtered per selection, resolved live by ListAvailableActions. This page snapshots what ships today; the authoritative list is always what dodil data actions list -b <bucket> -o json returns. Don’t hard-code these ids — enumerate them.

At the time of writing, dodil data actions list returns 13 actions. Each carries an action_id, an invoke mode (sync inline vs async streamed), and a params_form describing its typed form inputs.

The 13 shipping actions

Action idPurposeSelectionInvoke
object_qaAsk a question grounded in one document’s contentsObjectsync
summarizationMulti-level document summarization with keyword extractionObjectsync
classificationClassify documents by type, topic, language, and sensitivityObjectsync
entity_pii_extractionExtract structured entities and detect PII from any documentObjectsync
invoice_parsingExtract structured invoice fields (vendor, buyer, line items, totals, taxes)Objectsync
financial_statement_extractionExtract aggregate financial KPIs (revenue, net income, balance-sheet items)Objectsync
translationTranslate a document to a target language (Seed-X PPO 7B)Objectsync
objects_compareCompare N objects (N ≥ 2) on a chosen focusObjects (N)async
objects_synthesisSynthesize a single report across N selected objectsObjects (N)async
collection_topic_mapCluster a sample from a vector collection into named topicsCollectionasync
chunks_synthesize_answerGenerate a cited answer grounded in N selected chunksChunks (N)async
table_qaAsk a question in natural language against one K3 tableTableasync
plan_and_runFree-form agentic entry — the model writes and runs a Scriptum planNone (agentic)async

Selection is the selection shape the action accepts — and what ListAvailableActions filters on. Verified live: a --table selection offers only table_qa + plan_and_run; a single --key selection offers the seven Object actions + plan_and_run; objects_compare / objects_synthesis need N ≥ 2 keys; and plan_and_run is offered for every selection, since it is agentic.

Params and models

Each action’s params_form declares typed inputs, enum choices, defaults, and required flags. Common shapes seen in the live catalog:

  • llm_model — an LlmModel enum offering kimi-k2.5, kimi-k2.6, kimi-k3; the default is kimi-k2.5. Q&A / synthesis / topic-map actions expose it.
  • max_tokens — output cap (e.g. default 4096).
  • question — required on object_qa, table_qa, chunks_synthesize_answer.
  • translation runs on a dedicated Seed-X PPO 7B model, not the LlmModel enum.
  • Object actions on media declare a whisper-large-v3-turbo transcribe model and max_input_chars / chunk_size bounds.

Object-selection actions also carry accepted_content_types / accepted_extensions — the file kinds they run on (text, HTML, markdown, PDF, docx, xlsx, plus audio/* and video/* for the media-aware ones). The menu greys out actions whose accept-list doesn’t match the selected object.

Read the exact form for any action from the live catalog:

dodil data actions list -b kb-prod -o json \ | jq '.actions[] | select(.actionId == "summarization") | .paramsForm'

See also