# InsurGPT nodes InsurGPT nodes are the document-intelligence steps in a flow. They apply Bevaya's insurance-trained models to the documents a flow has ingested — segmenting a combined page stream into separate documents, extracting structured fields and classifying document types, and generating short business insights from the content. Three InsurGPT nodes are available today: - **Split Documents** — segment a multi-document page stream into separate documents. - **InsurGPT: Custom** — extract the fields you define and classify the document. - **Insights** — generate a short, written insight from the content and extracted data. These nodes produce AI output, so they are designed to work with the controls around them. Pair extraction with [Field Validation](./utility-nodes.md) and route low-confidence results to [Flag for Human Review](./utility-nodes.md) so a person confirms results before they're written back. **Permissions.** Building a flow requires the **Admin** or **Partner** role. The AI analysis and extraction these nodes perform is an *AI intelligence* capability available to the **Admin** role; the **Partner** role can build flows but is restricted from AI intelligence features. See [Users and access](../onboarding/users-access.md). **Inputs from upstream.** These nodes act on the documents prepared by [Read Files](./utility-nodes.md). The **Documents**, **Pages**, and **Item ID** fields default from the flow, so in most flows you don't set them by hand — you connect the InsurGPT node after Read Files (or after Split Documents) and configure the prompt and fields. **Model.** Where a node offers a **Model** selector, the default is **InsurGPT**, Bevaya's insurance-trained model. You can choose another model from the options in the selector. Leave the default unless you have a specific reason to change it. --- ## Split Documents In the Nodes library this node is labelled **Split Documents** (Page Stream Segmentation). **What it does.** Takes a single file made up of several documents — a scanned batch or a combined PDF — and segments it into separate documents by detecting where one document ends and the next begins. **Status:** Available. **When to use it.** Near the start of a flow, after [Read Files](./utility-nodes.md), when a single uploaded file actually contains multiple documents that should each be handled on their own. **Inputs.** The documents and pages prepared upstream (defaulted from the flow). **Configuration fields** - **Model** — the model used to detect document boundaries. Defaults to **InsurGPT**. - **PSS User Prompt** *(optional)* — free-text guidance to steer how the page stream is segmented, for cases where the default behavior needs a hint. **Outputs.** A list of **documents**, each with its own pages — one per detected document in the source file. **Downstream usage.** Feed the segmented documents into a [For Loop](./control-nodes.md) so each one is extracted and validated individually, or into [InsurGPT: Custom](#insurgpt-custom) directly. **Boundaries.** Split Documents decides where documents begin and end; it does not extract field values or classify the content — that's the job of InsurGPT: Custom. **Limitations.** Segmentation quality depends on the source file. A single-document file simply returns one document. Review results when batches are unusually formatted. **Example.** In [Claim Indexing](../use-cases/claims-indexing.md), a combined claim file is split into its constituent documents so each is indexed and routed separately. **Common setup problems.** Nothing to split because the node isn't placed after Read Files; or expecting multiple outputs from a file that is genuinely a single document. **Related.** [For Loop](./control-nodes.md), [InsurGPT: Custom](#insurgpt-custom). --- ## InsurGPT: Custom In the Nodes library this node is labelled **InsurGPT: Custom** (LLM Extraction). **What it does.** Extracts the specific fields you define from an insurance document and classifies the document. This is the core extraction step in most document flows. **Status:** Available. **When to use it.** After the documents are prepared (and split, if needed), to pull structured data — such as FNOL fields, policy details, or amounts — out of the document content. **Inputs.** The documents and pages from upstream (defaulted from the flow). **Configuration fields** - **Classification Prompt** *(required)* — instructions for the extraction and classification. A starting prompt frames the model as an insurance professional; tailor it to your document type. - **Fields to Extract** *(required)* — the specific fields you want returned. These define the extraction boundary: the node returns the fields you list. - **Grounding** *(optional, off by default)* — ties extracted values back to the source document to improve accuracy and reliability. - **Output as table** *(optional)* — return results in tabular form, with **Table name** to label the output. - **Pages** — which pages to read; defaults to all pages of the document. **Outputs.** The extracted fields, which are added to the flow's data so later nodes can use them. **Downstream usage.** Send the extracted fields into [Field Validation](./utility-nodes.md) to check them, branch on the results with [Control nodes](./control-nodes.md), route exceptions to [Flag for Human Review](./utility-nodes.md), and write confirmed values back with [App integration nodes](./app-integration-nodes.md). **Prompt and extraction boundaries.** The node extracts the fields you define from the documents and pages provided — it does not invent fields you didn't ask for, and it works from the supplied content rather than outside knowledge. Define clear field names and a prompt specific to your document type for the most consistent results. **Evidence behavior.** With **Grounding** enabled, extractions are tied back to the source document, which the option describes as improving accuracy and reliability. Grounding is off by default; enable it where traceability to the source matters. **Limitations.** AI extraction is not guaranteed to be correct on every field or document. Treat outputs as needing validation: use Field Validation thresholds and human review rather than writing extracted values straight to a system of record. **Example.** In [Submission Intake](../use-cases/submission-intake.md), InsurGPT: Custom extracts the submission fields from a broker's documents and classifies the document type, then the flow validates and routes them. **Common setup problems.** Nothing is returned because **Fields to Extract** is empty; results are inconsistent because the prompt is too generic for the document type; or the relevant pages are excluded by the **Pages** setting. **Related.** [Field Validation](./utility-nodes.md), [Flag for Human Review](./utility-nodes.md), [App integration nodes](./app-integration-nodes.md). --- ## Insights In the Nodes library this node is labelled **Insights**. **What it does.** Generates a short written insight from the document content and the data already extracted in the flow. **Status:** Available. **When to use it.** After extraction, when you want a brief narrative summary or highlight to accompany the structured data — for example, a one- or two-sentence summary of a claim file. **Inputs.** The document pages and the extracted fields from upstream (defaulted from the flow). **Configuration fields** - **Insights Prompt** *(required)* — instructs the node what insight to produce. A starting prompt asks for one to two sentences of key takeaways; tailor it to your need. - **Model** *(required)* — the model used to generate the insight. Defaults to **InsurGPT**. **Outputs.** The generated insight, added to the flow's data for downstream use. **Downstream usage.** Include the insight in a write-back, attach it to the work item, or send it in a [Send Outlook Email](./app-integration-nodes.md) summary. **Prompt boundaries.** The insight is generated from the document content and extracted data in the flow. Keep the prompt specific and bounded (for example, "summarize the loss in one or two sentences") rather than open-ended. **Limitations.** Insights are generated text and should be treated as a summary, not an authoritative record. Where the wording matters, route it through [Flag for Human Review](./utility-nodes.md). **Example.** In [Claim File Summarization](../use-cases/claims-file-summarization.md), Insights produces a short summary of each claim file for adjusters to read before opening the full documents. **Common setup problems.** The insight is too long or off-topic because the prompt is open-ended; or there's little to summarize because extraction didn't run first. **Related.** [InsurGPT: Custom](#insurgpt-custom), [Send Outlook Email](./app-integration-nodes.md). --- ## Where to go next - [Node catalog](./node-catalog.md) — every node and category at a glance. - [Utility nodes](./utility-nodes.md) — Read Files, Field Validation, and human review that surround extraction. - [Control nodes](./control-nodes.md) — loop over split documents and branch on extracted values. - [App integration nodes](./app-integration-nodes.md) — write confirmed results back to your systems. - [Test and debug a draft](./test-debug-draft.md) — run a draft on sample documents to check extraction.