NODES
DATA SOURCES
BigQuery Cost Gate Estimate a BigQuery read's cost and pause for human approval before the extract scans. Auto-approves under the cap when told to.
BigQuery Extract Pull a BigQuery table (cheap Storage Read API backfill + cheap incremental delta). Transform downstream in SQL.
HTTP Request Fetch a JSON document from an external URL (GET).
Workflow Artifact Consume an artifact produced by another workflow run.
Sync Data Bring the latest version of one or more data sources into the workflow.
DATA PREP
JSON Parse Convert a JSON document into a tabular dataset.
Profile Forecastability Compute the raw per-SKU forecastability profile (ADI, CV², SBC class, spectral entropy, omega, etc.) without bucketing — meant to be read by the composer agent.
Data Transformation Run a pandas transformation in a sandbox. Pick a template or write code.
Select Frequency Choose the working frequency (daily / weekly / monthly / …) for the dataset.
SQL Transformation Run a SQL query over the wired datasets (DuckDB in-memory). Uses an approved, versioned SQL when one is pinned.
Validate Schema Assert column presence, dtypes, and row-count thresholds.
Filter Products Filter SKUs by category, tags, or stock thresholds.
Data Quality Report Statistical quality assessment: gaps, outliers, NaN%, schema sanity. Surfaces issues before modeling.
HIERARCHY
Build Hierarchy Define + materialize the aggregation tree (total → … → sku, plus side aggregates).
FORECAST
Compare Runs (A/B) Rank N forecasts by an objective function so the composer agent can decide whether iteration v_{n+1} actually improved over v_n.
Forecast Generate demand forecasts (uses the production multi-model pipeline).
Forecast Accuracy Tracking Grade forecast vintages against realized sales (closed periods) and persist accuracy rollups + the FVA degradation signal.
Fetch Plot Image Render a forecast as PNG (base64) so Claude vision can read the chart. Selective by SKU + horizon; not for batch use.
Daily Forecast Inference Infer today's forecast from the latest training cycle without re-fitting (ML .update() / stats .forward() / neural reload / foundation re-inference).
Hierarchical Reconciliation Reconcile a multi-level forecast (bottom-up / top-down / MinT). Produces a coherent forecast that plugs into Update Control Tower and Stocking nodes identically to a fresh forecast.
Avisos del plan Calcula las líneas del plan que piden una decisión y escribe el resumen que abre el Plan baseline.
LLM Briefing Escribe el briefing por serie del plan de demanda (el texto de las tarjetas Revisar/Observar) para el forecast conectado.
ABASTECIMIENTO
Backtest de políticas Prueba políticas de stock sobre las semanas del backtest del forecast y elige la mejor de cada segmento.
Ejecutar política Calcula los pedidos de esta semana y proyecta el stock, con la política que fijes o la que ganó el backtest.
LOGIC
Delay Wait N seconds before continuing (max 600s).
For Each Iterate over groups (e.g. per-channel) and emit a marker per group.
Approval Gate Pause until a user approves before continuing.
If Branch on a predicate evaluated against the input artifact.
Correr workflow Start another workflow of the org and wait for it to finish.
Switch Route the input to one of N output ports based on a string key.
OUTPUTS
Email Send a short email + optional CSV/Excel attachment via Resend.
Export CSV Save the input dataset as a CSV in org storage.
Publish Artifact Promote an artifact under a stable alias (e.g. `latest_forecast`).
Save as Data Source Write the input dataset as the org's selected file of a data type.
Update Control Tower Promote the forecast + hierarchy as the canonical Control Tower results.
Webhook POST a JSON summary of the input artifact to an external URL.
Send Orders Send orders to ERP / webhook (mock).
AI
Ask Strategist Ask the LLM a small decision question and capture the answer.
Data Engineering Agent Defer a transformation goal to the data-engineering agent.
Explain Artifact Narrate what an upstream artifact contains and how it should be read.
AI Step Bounded LLM reasoning step: prompt + tool subset + output schema + budget. The composer agent's escape hatch for in-DAG reasoning.
DISAGGREGATION
Distribute (top-down) Disaggregate a parent-level forecast down the tree using historical share + similar_to references. Defaults to SKUs; set `target_level` to stop at an intermediate level. Use when the target level forecasts noisily on its own (sparse demand, high product churn).
Stock sucursales y CD, arribos y parámetros
Drop to add node
100%
1. Último forecast Workflow Artifact
11s
Status COMPLETED
Started 12:09:30
Completed 12:09:41
2. Stock sucursales y CD, arribos y parámetros Sync Data
52s
Status COMPLETED
Started 12:09:30
Completed 12:10:22
3. Backtest de políticas Backtest de políticas
Status RUNNING
Started 12:10:23
4. Pedidos a proveedores Ejecutar política