Node reference

For Each

Execute an inline subworkflow once for each upstream record.

Use forEach when every record needs its own fetch-and-process graph, for example visiting each company’s website and extracting a contact record.

Input and output

Connect a node producing records. Each item supplies values to the entry node’s configuration through bind, whose keys are config fields and values are dotted record paths. Missing, empty-string, or empty-list bindings skip that item.

The node emits combined records in llm_result.data and extraction_result.data, with _foreach_index and _foreach_key provenance. foreach_result reports item counts, failures, skipped items, record count, and truncation when max_items cuts the input.

Configuration

Field	Default / requirement
`flow`	Required inline engine graph: `nodes` with `config`, and `edges`
`entry`	Required id of a node with no incoming edge in that graph
`bind`	Required nonempty map, for example `{"url":"website"}`
`concurrency`	`4`; from `1` to `8`
`max_items`	`500`; from `1` to `5000`
`item_timeout_seconds`	`300`; from `10` to `3600`
`on_error`	`continue` or `stop`; default `continue`
`key`	Optional record path for item identification

Example

This config visits each upstream website and returns an answer. The inner graph uses engine config, even when the outer For Each node is saved as canvas data. Its placeholder URL makes the inner graph valid before execution; bind replaces it for each item.

{
  "entry": "fetch",
  "bind": {"url": "website"},
  "max_items": 8,
  "concurrency": 2,
  "flow": {
    "nodes": [
      {"id": "fetch", "type": "urlFetcher", "config": {"url": "https://example.com", "depth": 0, "max_pages": 1}},
      {"id": "markdown", "type": "markdownConverter", "config": {"content_filter": "none"}},
      {"id": "answer", "type": "llmProcessor", "config": {"prompt": "What does this company do?", "output_format": "json"}}
    ],
    "edges": [
      {"source": "fetch", "target": "markdown"},
      {"source": "markdown", "target": "answer"}
    ]
  }
}

Current limitations

Execution and failure behavior

Item timeouts count as failures. With continue, failed items are reported and successful output is retained; if every taken item fails, the node raises an error. Results are assembled in input order, regardless of completion order.

Subflows use the run’s shared model/page/byte limits. Inspect foreach_result before trusting a partial result.

For cross-branch enrichment, see Combine. For reusing stored URLs, see Query.