LLM Processor

Two modes. Pick one per node.

Mode 1 — Extract a record from each page

{
  "extraction_strategy": "llm",
  "extraction_input_format": "markdown",
  "extraction_apply_chunking": false,
  "extraction_instruction": "Return the company on this page. If this is a list page, return nothing. Copy values exactly as written; use null for anything the page does not state. Never guess.",
  "extraction_output_schema": {
    "type": "object",
    "properties": {
      "name": { "type": "string" },
      "phone": { "type": ["string", "null"] },
      "email": { "type": ["string", "null"] }
    },
    "required": ["name"]
  }
}

In the app: Extract a record from each page, with a Fields list (name and type) instead of a JSON schema.

Do not set prompt in this mode

A prompt makes the fetcher judge every page with the model as well (“semantic crawling”), which can fail the run. Put your wording in extraction_instruction.

Write a good instruction

1. Say what to return when a page is not one record (“if this is a list page, return nothing”).
2. Say to copy values as written.
3. Say to use null for what the page doesn’t state, and never guess.

No model needed

extraction_strategy can also be css, xpath or regex, which need no model call. They are exact, free and fast for stable markup. The model can even generate the selector schema from a sample page (extraction_generate_schema).

Mode 2 — Answer a question about the pages

{ "prompt": "What does this company sell?", "output_format": "json", "schema": { "type": "object" } }

Answers are grounded in the crawled pages, with source references, and can report supported, partial, conflicting or not-found outcomes. Use output_format: "text" for prose.

Choosing a model

Leave model out for the platform default. To use your own endpoint, pick it in the node’s Model (see Resources).

Useful limits

max_context_pages, max_input_chars, max_tokens, temperature, extraction_parallelism, strict_json.

Inputs, outputs, and failures

The node processes upstream pages according to the selected mode. Extraction emits structured records and metadata; answer mode emits an answer with grounding information. Inspect the output before choosing Table (records) or Text (prose). An instruction asking for null values does not guarantee that the model follows it.

Missing or unavailable models, context/token limits, invalid output schemas, and malformed model responses can fail processing. Validate the graph first, inspect the failed-node message, and test a few source pages. Defaults and constraints for every option are in the generated configuration reference.