Worked example: extract once, check daily

A contract register built from two workflows — one that writes it, one that watches it.

This is the pattern most Data Tables end up being: extract once, check daily. Somebody uploads a contract, a run reads it, a person confirms it, and the result becomes a row. From then on a schedule watches the register and emails whoever needs to act.

Two workflows, one table. The table is what connects them — neither workflow knows the other exists.

The table

Create it in Studio → Data Tables, or with a Table · Create table node that has Return existing if it exists ticked so the workflow is safe to re-run.

Column	Type	Holds
`Contract`	`string`	The contract reference. The key.
`Counterparty`	`string`	Who it is with.
`Deadline`	`date`	The date something has to happen by.
`Owner`	`string`	The email address that gets the reminder.
`Status`	`string`	`open` while it still matters.

Contract is what makes the whole pattern work: it is the value the upsert matches on, so the same contract uploaded twice updates one row instead of making two.

Workflow 1 — write the register

Form Trigger — the upload

A Form trigger with a file field named contract and a text field named owner. The run starts when somebody submits it, and the uploaded file arrives as {{ trigger.body.contract }}.

AI Prompt — the extraction

An AI Prompt node with the uploaded file attached, asked for the three facts the register needs as JSON:

Read the attached contract and answer with JSON only:
{"contract": "<reference>", "counterparty": "<name>", "deadline": "<YYYY-MM-DD>"}

Ask for the deadline as YYYY-MM-DD. A date column accepts ISO 8601 and nothing else, so "next Tuesday" is a failed node rather than a wrong row.

Human Review — the confirmation

A Human Review node, assigned to whoever owns contracts, with three response fields: contract (text), deadline (date) and counterparty (text). Put what the model read into the node's message, so the reviewer is checking a claim rather than starting from a blank form:

Extracted from the upload: {{ upstream.extract-1.output.text }}
Correct anything that is wrong below.

This is the step that makes an extraction trustworthy enough to keep. What the reviewer submits is what gets written: {{ upstream.<review node id>.output.response.deadline }}.

Data Table — the upsert

A Row · Upsert row node on the register. Its conditions are the key, and its mapping is the row:

node data — Row · Upsert into the register

{
  "resource": "row",
  "operation": "upsert",
  "tableId": "8a9f0e42-1d7c-4a3a-9c2e-6b4d1f77e001",
  "tableName": "Contract register",
  "conditions": [
    {
      "id": "cond-1",
      "column": "Contract",
      "operator": "eq",
      "value": "{{ upstream.review-1.output.response.contract }}",
      "valueMode": "expression"
    }
  ],
  "conditionMatch": "all",
  "mapping": "manual",
  "mappingRows": [
    {
      "id": "map-1",
      "column": "Contract",
      "value": "{{ upstream.review-1.output.response.contract }}",
      "valueMode": "expression"
    },
    {
      "id": "map-2",
      "column": "Counterparty",
      "value": "{{ upstream.review-1.output.response.counterparty }}",
      "valueMode": "expression"
    },
    {
      "id": "map-3",
      "column": "Deadline",
      "value": "{{ upstream.review-1.output.response.deadline }}",
      "valueMode": "expression"
    },
    {
      "id": "map-4",
      "column": "Owner",
      "value": "{{ trigger.body.owner }}",
      "valueMode": "expression"
    },
    { "id": "map-5", "column": "Status", "value": "open", "valueMode": "fixed" }
  ]
}

Run it twice with the same contract reference and the register holds one row: the node's output says "action": "inserted" the first time and "action": "updated" the second.

Tick Dry Run while you are still wiring this up. The node reports what it would write — and for an upsert that would insert, the exact cells, as result.insertValues — without touching the register.

Workflow 2 — watch the register

Recurrence — daily

A Recurrence trigger set to run once a day. A scheduled run carries no payload, so everything this workflow needs comes out of the table.

The cutoff date

{{ }} is dot-path substitution, not an expression language: there is no arithmetic in it, so today + 30 days has to be produced by a node. An AI Prompt node asked for exactly that is the simplest version:

Answer with today's date plus 30 days, as YYYY-MM-DD, and nothing else.

Its output is then {{ upstream.horizon-1.output.text }}.

Data Table — what is coming due

A Row · Get row(s) node: everything still open whose Deadline is at or before the cutoff, soonest first.

node data — Row · Get what is coming due

{
  "resource": "row",
  "operation": "get",
  "tableId": "8a9f0e42-1d7c-4a3a-9c2e-6b4d1f77e001",
  "tableName": "Contract register",
  "conditions": [
    { "id": "cond-1", "column": "Status", "operator": "eq", "value": "open", "valueMode": "fixed" },
    {
      "id": "cond-2",
      "column": "Deadline",
      "operator": "lte",
      "value": "{{ upstream.horizon-1.output.text }}",
      "valueMode": "expression"
    }
  ],
  "conditionMatch": "all",
  "orderBy": [{ "id": "sort-1", "field": "Deadline", "direction": "asc" }],
  "returnAll": true
}

conditionMatch is all: both have to hold. Switch it to any and you would email about every open contract and every past deadline, which is not the same question.

Send Email — the reminder

A Send Email node. The rows are {{ upstream.due-1.output.result.rows }}, and how many there were is {{ upstream.due-1.output.affected }}.

To send nothing on a quiet day, put an If / Else node in front of it and branch on {{ upstream.due-1.output.affected }} — or use a Row · Row exists node, whose {{ upstream.<id>.output.result.exists }} is a plain boolean.

What this pattern gets you

The register outlives the run. A contract read in October is still found by the schedule in November.
Corrections stick. A deadline fixed by hand in Studio → Data Tables is what the next day's Get row(s) reads. Nothing re-extracts over the top of it.
Re-uploading is safe. The upsert's key makes the second upload an update.
The two workflows stay independent. Replace the extraction with a better one, or the email with a Slack message, and the other half does not change.

Worked example: extract once, check daily

On this page