Best Document Splitting APIs for Insurance, Legal, and Healthcare Claims (October 2026)

Insurance claims, legal case files, and medical records share one problem: dozens of documents arrive as a single PDF, and every downstream step depends on knowing where each one starts and ends. Extend's Split API is the strongest fit when those boundaries feed extraction and review, because it splits on identifiers you choose, such as a claim number, Bates prefix, or exhibit number, and returns each sub-document as its own file with a written reason for the grouping. Reducto Split and LlamaSplit both classify pages into categories you define, and a frontier LLM can split short packets directly but misses many boundaries on long ones.

TLDR:

• Extend splits on an identifierKey such as a claim number, exhibit number, or Bates prefix, and returns startPage, endPage, type, identifier, observation, and a standalone fileId per sub-document.
• Reducto Split returns page lists per category with high or low confidence, and can partition a category by an identifier with partition_key.
• LlamaSplit, in beta, classifies pages into 1 to 50 categories, returns high, medium, or low confidence per segment, and accepts PDF, DOCX, and PPTX.
• On PoliTax-Split, frontier models splitting long packets directly scored 37.6%-64.1% F1. The same models inside Extend's splitter scored 64.1%-72.5%.
• None of the three split APIs redacts content. Legal teams that need redaction add it as a separate step.

Why claim packets and case files need splitting first

Extraction schemas are written per document type. A repair estimate, a police report, and a medical bill each need different fields. If a 90-page claim packet goes to extraction as one file, fields from one document bleed into another, and line items from the estimate can land in the bill. Splitting first turns one unknown packet into a list of typed documents, each routed to the right schema and reviewed against the right pages.

The three industries differ in how boundaries are marked:

• Insurance: claim numbers and policy numbers repeat across documents in the same packet, so the identifier alone doesn't mark a boundary. Document type changes do.
• Legal: exhibit numbers and Bates ranges mark boundaries explicitly, but exhibits can contain several document types that should stay together.
• Healthcare: patient identifiers repeat throughout, while encounters, claims, and remittances change. Forms such as CMS-1500 and UB-04 have fixed layouts, and clinical notes don't.

How we compared these APIs

We compared documented API behavior, published benchmark results, and published compliance terms. We did not run a new head-to-head test. Where a vendor's documentation does not describe a capability, the table says so rather than guessing.

Extend Split

Extend's Split API takes your document types as splitClassifications, each with a description, and an optional identifierKey that extracts an identifier from each sub-document. With strictIdentifierModeEnabled, adjacent pages that share an identifier merge into one split, which keeps a multi-document exhibit together. Plain-language splitRules handle cases such as keeping a multi-page contract intact.

Each sub-document returns a page range, type, identifier, an observation explaining the grouping, and a fileId you can send straight to Extract or Classify. Workflows route each sub-document to the right processor and can pause for human review. For volume, async /split_runs supports webhooks and batch requests take up to 1,000 files. In production, Extend split a 1,044-page loan document into 315 sub-documents in 4 minutes and 43 seconds, parsing included.

Extend supports SOC 2 Type II, HIPAA with BAAs, GDPR, and EU processing, with zero data retention available on request.

Reducto Split

Reducto Split takes a split_description list of categories, each with a name, description, and optional partition_key. It returns the pages for each category with a high or low confidence label. Partitions name each sub-section by the identifier value found, such as an account number. Pages can belong to more than one section by default. deep_split enables a higher-quality split agent, and /split_async adds webhooks and a priority queue.

Reducto covers SOC 2 Type II and HIPAA. Its published BAA, zero data retention, and regional processing terms require a custom-priced Growth or Enterprise plan.

LlamaSplit (LlamaParse)

LlamaSplit is a beta API that classifies each page of a concatenated document into 1 to 50 categories you describe, with optional custom instructions. It returns segments with page ranges and a high, medium, or low confidence. It accepts PDF, DOCX, and PPTX files up to 512 MB and runs as a job you poll. Because it is in beta, LlamaIndex notes the API is subject to breaking changes.

LlamaParse documents SOC 2 Type II and Enterprise HIPAA BAA options. It fits teams already on LlamaParse that want splitting in the same platform as parsing and indexing.

Frontier LLMs without a splitter

Sending a packet straight to a frontier model works for short files. Long ones break down: on PoliTax-Split, raw models merged supplements into parent forms and combined back-to-back forms of the same type, so they missed many boundaries. Context limits force chunking, and chunk edges are exactly where boundaries fall.

Comparison table

Insurance: claim packets and submissions

A claim packet can hold a first notice of loss, ACORD forms, a police report, photos, repair estimates, medical bills, and correspondence, all under one claim number. Because that number repeats on every document, split on document type and use the claim number as the identifierKey only to confirm which packet each page belongs to. Write splitRules that keep multi-page estimates and invoices together, and add an "other" type so unrecognized pages are flagged rather than absorbed into a neighboring document.

Legal: multi-exhibit filings and discovery productions

Court filings and discovery productions mark their own boundaries. Use the exhibit number or Bates prefix as the identifierKey and turn on strict identifier mode so every page of Exhibit 12 stays in one split, even when the exhibit holds an email, an attachment, and a contract. Classify each split afterward when you need document-level types inside an exhibit.

Redaction is a separate step. Extend's Edit API fills and flattens PDF form fields; it does not black out content, and neither Reducto Split nor LlamaSplit redacts. Run splitting first so each document reaches your redaction tool on its own.

Healthcare: medical records and claims

Medical record bundles mix encounters, lab results, imaging reports, and claim forms under one patient. Split by document type, use an encounter or claim number as the identifierKey where it exists, and route CMS-1500 claims, EOBs, and prior authorization packets to their own schemas. Confirm a BAA is in place before sending protected health information to any vendor; see HIPAA and SOC 2 requirements for parsing.

How to choose

Choose Extend when splits feed extraction and review, when boundaries follow identifiers such as claim, exhibit, or Bates numbers, and when each sub-document needs its own file and recorded reason. Choose Reducto when you already parse with it and want account-style partitions. Choose LlamaSplit when the rest of your stack runs on LlamaParse and a beta API is acceptable. For traceable output and throughput, see document splitting with source citations: Extend vs Reducto.

FAQ

What is the best API for document splitting in legal?

Extend, for multi-exhibit filings and discovery productions. It splits on an identifier you choose, such as an exhibit number or Bates prefix, and strict identifier mode keeps every page sharing that identifier in one split. Each sub-document returns a page range, type, written observation, and standalone file ID for extraction or classification. LlamaSplit fits teams already on LlamaParse that accept a beta API, and none of these split APIs redacts content.

What is the best document splitting API for insurance claims?

Extend, when claim packets feed extraction and adjuster review. Split by document type with an "other" catch-all, use the claim number as the identifier to confirm packet membership, and keep multi-page estimates together with split rules. Each sub-document becomes its own file for extraction with citations, and workflows can pause uncertain values for human review.

Can a frontier LLM split medical record bundles without a splitter?

For short files, often yes. For long bundles, accuracy drops. On the PoliTax-Split benchmark of long compound documents, frontier models splitting directly scored 37.6%-64.1% F1, and the same models inside Extend's splitter scored 64.1%-72.5%. Raw models mostly missed boundaries by merging adjacent documents.

Which document splitting APIs offer HIPAA BAAs?

Extend supports HIPAA with BAAs. LlamaParse documents Enterprise HIPAA BAA options. Reducto covers HIPAA, with its published BAA terms tied to custom-priced Growth or Enterprise plans. Confirm the BAA before sending protected health information to any vendor.