Skip to main content
Version: v1.7.0

Extractor settings

When you create or edit an extractor, you can configure settings that control vision capabilities, prompt customization, chunk limits, and guardrails. To access these settings, open the extractor creation or edit form. For details, see Create an extractor.

Vision​

note

Vision settings appear only when your deployment includes vision-capable models.

Enable vision​

Pass document context as images to a vision-capable LLM, in addition to text.

Select one of the following:

  • Off: Sends document context as text only.
  • Automatic (default): Automatically determines whether to use vision based on the document content and selected LLM.
  • On: Always passes document context as images to the vision LLM.

Requires vision-capable models in your deployment. Enabling vision mode can increase latency and cost. For details, see Enable vision.

Vision LLM​

Select a specific LLM for vision and image processing tasks. When set to Automatic, the auto-selected model name appears below the dropdown. If no vision model is available for the selected LLM, a warning appears instead.

Vision Settings

Chunks​

Max number of chunks​

The Max Number of Chunks slider sets the maximum number of document chunks the extractor processes. The range is 10 to 1,000, with a default of 100. More chunks increase coverage of large documents but require more processing time and LLM calls.

Chunk Settings

Prompts​

Prompt template​

Populates the System prompt, Pre-prompt, Prompt, and image batch prompt fields with values from an existing template. For details, see Prompts.

System prompt​

Set instructions to send to the LLM before any document context or user prompts. Use this to define the LLM's behavior and role for the extraction task.

Prompts before/after document context​

Add custom text before and after the document context. This setting has three fields:

  • Pre-prompt (top field): Text that appears before the document context.
  • Document context (middle, read-only): The document chunks being processed.
  • Prompt (bottom field): Text that appears after the document context.

When you define a JSON schema, the output is always a JSON object. The prompt can customize extraction behavior but cannot override the JSON output format.

caution

When a JSON schema is defined, the Pre-prompt field is ignored. Only the Prompt (after document context) is applied.

note

If the document exceeds the LLM's context window, the system applies these prompts multiple times — first to the original document chunks, then to intermediate results. To apply the LLM only once to the original chunks, enable Keep Intermediate Results.

Prompt per image batch for vision models​

Set the prompt that processes each batch of images when vision mode is enabled. This prompt appears after the batch of images and before the user prompt.

Prompt for final image batch reduction for vision models​

Set how the system combines per-batch answers into a final answer when vision mode is enabled. This prompt appears after the collected per-batch answers and before the user prompt.

note

These image batch prompt fields are always visible in the form, but only take effect at runtime when Enable Vision is set to Automatic or On.

Prompts Settings

Schema​

Input JSON schema​

Switch between two modes for defining the extractor's output schema:

  • Off (builder mode): Use the visual JSON schema builder to add fields and tables interactively.
  • On (code mode): Enter a valid JSON schema directly in a text editor.

For details, see Create an extractor.

Schema Settings

Guardrails and PII​

The extractor form includes Enable guardrails and Enable PII controls toggles. When enabled, these expose the same sub-settings available when creating a collection. For details, see Guardrails and PII Detection.

note

The Parse Action setting (available in Create a Collection) is not available in extractors, as extractors process documents that are already ingested.

Advanced​

Keep intermediate results​

The Keep Intermediate Results toggle stores intermediate processing results during extraction.

  • Enabled: The final result is a list of each chunk's individual extraction output. Useful for debugging or when documents exceed the LLM's context window.
  • Disabled (default): Intermediate results are combined through recursive summarization into a single final output.

Guardrails, PII, and Advanced Settings

Feedback