Changelog

Document Extraction and Document to Markdown v2 upgrade the extraction pipeline while keeping the same request and response shape. That means higher-quality OCR and extraction results without a migration on your side.

At the same time, v2 replaces the remaining OCR and extraction dependencies on non-EU providers, so the full pipeline runs on infrastructure we operate inside EU datacenters from end to end.

What will change

• Fully EU-hosted — No Vertex AI, no Gemini, no transatlantic round-trip. The DPA covers every step end-to-end.
• Higher-quality results — Upgraded extraction and OCR pipeline. Same confidence scores, more often correct.
• Same API shape — Request and response formats are unchanged. No code changes needed on your side when v2 ships.
• Same pricing — Still 1 credit per page.

We’ll post the launch date here as soon as we have one. If you’d like early access, contact us.

Pulling structured data out of calls, interviews, meetings, or voice notes usually means stitching together transcription, speaker labeling, and custom parsing before you can use the result anywhere else in your workflow.

Audio Extraction is our roadmap answer to that problem. The goal is simple: send an audio file, define a schema, and get structured JSON back through the same extraction model you already use for documents.

What it will include

• Schema-based extraction from audio — define the fields you care about and get structured JSON back instead of a transcript you still need to parse manually.
• Same extraction conventions as documents — confidence scores, citations, and structured outputs fit the same downstream workflows you already use elsewhere in Iteration Layer.
• Speaker-aware results — speaker detection labels who said what, which matters for interviews, support calls, meetings, and multi-person recordings.
• Timestamped citations — extracted values point back to the moment in the recording where they came from, which makes results easier to verify.
• EU-hosted processing — transcription runs on self-hosted open-weight models on EU infrastructure, so recordings stay on EU servers.

Instead of transcribing a conversation in one tool and then trying to parse the transcript in another, you will be able to extract structured values directly from audio and feed them into the same downstream reporting, document-generation, or agent workflows you already use elsewhere in Iteration Layer.

We published OCR benchmark results for the document pipeline behind our extraction and markdown APIs, so teams can evaluate the first document-conversion step before wiring it into a larger workflow.

The benchmark covers 41 real workflow files across forms, invoices, scans, tables, charts, photos, statements, reports, and receipts. Each file was converted to markdown, then judged against the source image for completeness, accuracy, structure, and extra text.

Iteration Layer ranked second in our benchmark with a 0.93 average score, behind Chandra-OCR-2 at 0.94 and ahead of Qwen3-VL-Instruct at 0.91. The current API passed all 41 files in the suite.

The page includes the model comparison table, tested document categories, methodology notes, the judge setup, and a hand-written markdown version for agents and LLM tools: OCR benchmark.

Scraping public websites usually means writing selectors, patching parsers when layouts shift, and maintaining a second extraction path that behaves differently from the one you already use for documents.

Website Extraction applies the same schema-based extraction model as Document Extraction to public web pages. Send a URL and a schema and get structured, typed JSON back from the page content.

What changed

• Schema-based extraction for live web pages — use the same field definitions you already use for documents instead of maintaining a separate scraping model.
• Typed, explainable outputs — fields come back with value, type, confidence, citations, and source, so website extraction behaves like the rest of the platform.
• Advanced extraction features included — nested arrays, validated field types, calculated fields, and async webhooks work here too.
• Fetch controls for difficult pages — use fetch_options for JavaScript rendering, proxy routing, region targeting, locale, or user-agent control.
• Available everywhere you need it — REST API, SDKs, and the extract_website MCP tool.

This is useful when a workflow starts on the web and ends in the same downstream systems as your document workflows. Extract structured data from a public page, then pass it into reporting, generation, or agent flows without building a separate scraping stack.

For pages that need JavaScript rendering, proxy routing, region targeting, locale, or user-agent controls, pass fetch_options to customize how the page is fetched. Available via the REST API, all SDKs, and as the extract_website MCP tool.

The Iteration Layer n8n node is now available as a verified community node in n8n. That makes it easier to discover from the n8n canvas and gives teams a reviewed integration for document, image, and generated-output workflows.

What changed

• Verified by n8n — The integration is listed in n8n’s official integrations directory and marked as verified by n8n.
• Available in n8n Cloud — Instance owners can add the verified community node from n8n, then everyone on the instance can use Iteration Layer resources in workflows.
• Same API coverage — The node keeps the same Iteration Layer capabilities: extract data from documents, convert files to markdown, transform images, generate images, create documents, and generate spreadsheets.

See the n8n page for setup details or install it from the verified n8n listing.

Losing credits on a request we never executed is the kind of billing friction that makes testing slower and production workflows harder to trust.

Failed requests are now refunded automatically, regardless of whether the failure came from our side or yours. Previously only server errors were refunded. Validation failures, malformed payloads, and other client-side errors could still consume credits even when no real work was performed.

What changed

• 4xx errors are refunded — Validation failures, missing parameters, invalid base64 buffers, unknown content block types, and any other client-side error returns the credits that were reserved for the request.
• Trial credits are refunded too — Previously the refund path skipped trial grants. Trial users evaluating the platform are no longer penalized for an off-by-one in their first curl.

Credits should reflect work performed, not requests attempted. A request we never executed shouldn’t draw down your balance — even if your payload was the reason we never executed it.

You do not need to change anything. Refunds happen automatically when the failed response is returned.

Document Extraction and Document to Markdown now accept a much broader set of document, email, text, and image formats, so more real-world inputs can stay inside the same pipeline.

What changed

• Fewer special cases — more uploads can go through the same extraction or markdown flow without format-specific glue code.
• Broader ingestion coverage — email, office, ebook, notebook, markup, and archival formats can now enter the same downstream workflow.
• Same platform conventions — the new formats work through the same API shape, SDKs, and MCP tools.

New document formats:

• PPTX — PowerPoint slides extracted as markdown with formatting, bullet lists, numbered lists, and tables
• ODT — OpenDocument Text with full rich text support: bold, italic, strikethrough, links, footnotes, nested lists, and tables
• ODS — OpenDocument Spreadsheet with multi-sheet support
• EPUB — Ebook chapters converted to markdown via the HTML pipeline
• RTF — Rich Text Format with bold, italic, strikethrough, Unicode, and special characters

Email formats with attachment extraction:

• EML — Standard email format with header extraction, body parsing (plain text, HTML, multipart MIME), and automatic attachment extraction
• MSG — Outlook email format (OLE2/CFB binary) with the same header, body, and attachment extraction

Email attachments are automatically ingested through the same pipeline. A PDF attached to an email gets the same OCR and layout analysis as a standalone PDF upload. Attachments are returned separately in the response and included in the extraction context for Document Extraction. Each attachment counts as its own document for billing.

New structured/text formats:

• TSV — Tab-separated values, parsed into markdown tables like CSV
• Jupyter Notebook (.ipynb) — Code and markdown cells extracted with outputs
• LaTeX (.tex, .latex) — Converted to markdown: headings, formatting, lists, tables, math equations, links, and code blocks
• XML, YAML, TOML, RST, Org, Djot, MDX, BibTeX, Typst — Text and markup formats passed through for LLM consumption

New image formats:

• BMP, TIFF — Common raster formats
• JPEG 2000 (.jp2, .jpx, .jpm, .mj2) — Used in archival and medical imaging
• Netpbm (.pnm, .pbm, .pgm, .ppm) — Unix-origin portable image formats

All new formats work with the REST API, SDKs, and the MCP server tools convert_document_to_markdown and extract_document.

Agencies feel the pain of fragmented tooling faster than almost anyone else. One client needs extraction, another needs generated deliverables, a third needs image workflows, and each extra vendor adds more setup, more billing overhead, and more delivery risk.

The Iteration Layer partner program is built for agencies that want one reusable setup for multi-client document and image-processing workflows.

What partners get

• Dedicated support channel for technical questions and feature requests
• Early access to new APIs and features before general availability
• Unified billing, project-level scoping, and active-project discount tiers that scale with real client usage
• Manual first-project guidance after approval
• Co-marketing and case study co-creation when projects become strong proof points

Apply now.

You can now share a referral link with other developers and teams. When someone signs up through your link and makes their first API call, you both receive 500 credits.

How it works

1. Visit the new Referrals page in your dashboard to find your unique referral link.
2. Share it with anyone who could benefit from document and image processing APIs.
3. When they sign up and make their first API call, both accounts receive 500 credits automatically.

Details

• No cap on referrals — refer as many people as you want.
• Referral credits never expire.
• Credits are granted on first API call, not just signup, to prevent abuse.
• Track your referrals and earned credits from the Referrals page.

See the Referral Program docs for full details.

Monthly commitments are a real source of friction when a developer or small team wants to move from “trying the API” to “putting a real workflow into production” without guessing usage in advance.

Pay As You Go removes that upfront commitment. You can use any API and get billed at the end of the month based on actual usage.

What changed

• No monthly commitment — use what you need, cancel anytime
• Volume discounts — per-credit rate decreases automatically as usage grows, from $0.033 to $0.022 per credit
• Trial credits included — same free trial credits as subscription plans
• Project budget caps — enforced per project, same as subscriptions
• Interactive pricing builder — estimate your cost by API on the pricing page
• Auto overage for subscribers — enable auto overage to seamlessly bill excess usage via pay-as-you-go rates when your plan credits run out
• Document Extraction and Document to Markdown now charge 1 credit per page — no more 5-credit minimum. A single-page document costs 1 credit instead of 5.

How it works

1. Sign up and add a card from the billing dashboard
2. Use any API — usage is tracked automatically
3. At the end of each billing period, you’re invoiced based on graduated pricing tiers
4. The more you use, the lower your per-credit rate

Subscriptions remain available for teams that want predictable monthly costs. At lower volumes, subscriptions often remain the better fit. Subscribers can also enable auto overage for overflow billing.

We updated our Terms of Service, Data Processing Agreement, and Privacy Policy so the legal layer matches how the platform works today.

What changed

• Pay-as-you-go pricing is now covered alongside subscription plans.
• Price increase protection — if we raise prices by more than 10%, you have 30 days to cancel without further charges.
• Budget caps and overage controls are documented for subscription plans.
• Force majeure clause — obligations are suspended during events beyond either party’s control. Either side can terminate after 30 consecutive days.
• Faster data deletion — account data is now deleted within 30 days of termination (previously 90), except billing records retained for tax compliance.
• Refund policy is now cross-referenced from the Terms of Service.

What has not changed: infrastructure remains EU-hosted, subprocessors are the same, and your data handling model stays the same.

Usage gets hard to reason about once one account serves multiple environments, teams, or client projects. Per-key tracking helps at the edge, but it does not match how people actually manage work.

You can now organize API keys and usage by project.

What changed

• Project-level organization — create projects in the dashboard and assign API keys to them.
• Project-level budget caps — limit credit consumption across all API keys in a project instead of managing caps key by key.
• Better usage visibility — filter the Usage page by both project and product to see where credits are going.
• Direct project usage views — projects link straight to their filtered usage pages.
• Safer transition — existing API keys without a project continue to work without budget enforcement.

This is especially useful when one account serves multiple clients, environments, or teams and you need the control boundary to match the way the work is actually organized.

The n8n-nodes-iterationlayer community node brings Iteration Layer directly into n8n, so those content-processing steps can live inside the same automation workflow with dedicated resources instead of custom glue.

What changed

• Extract data from documents inside n8n workflows
• Transform images without leaving the workflow builder
• Generate graphics, documents, and spreadsheets as downstream steps
• Use the same node whether you self-host n8n or run on n8n Cloud

See the n8n integration guide for setup instructions.

Document to Markdown turns PDFs, images, office documents, and websites into clean markdown, so messy inputs can move straight into search, extraction, agent, or generation workflows without custom conversion steps.

What changed

• Clean markdown from messy files — convert PDFs, images, and office documents into one text format before feeding them into search, prompting, extraction, or generation workflows.
• Preserved document structure — headings stay headings, tables come back as markdown tables, and layout-aware parsing keeps the output usable instead of flattening everything into raw text.
• OCR built in — scanned pages are OCR’d automatically, so image-heavy documents can enter the same text-first workflow as digital files.
• Image descriptions included — image responses include a description field that can be used for alt text, indexing, and LLM context.
• Agent-ready distribution — the API is available through REST, SDKs, and the convert_document_to_markdown MCP tool.

This is especially useful when the first step in your pipeline is normalization rather than extraction. Convert uploaded files to markdown first, then feed that result into a schema-based extraction flow, an agent workflow, or a generated report.

Available via the REST API, all SDKs, and as the convert_document_to_markdown MCP tool.

Document Generation now supports 9 font weights (thin, extralight, light, regular, medium, semibold, bold, extrabold, black) via the new font_weight field, replacing the previous is_bold boolean. PDF and EPUB render all weights as CSS numeric values (100–900). DOCX and PPTX map semibold and above to bold.

Styles and page settings are now optional. When omitted, sensible defaults are applied: PlusJakartaSans body text, Outfit headings, A4 page with 72pt margins. Only the fonts your document references are loaded, keeping output size minimal.

What changed

• More precise branding control — choose from thin through black instead of treating text as just regular or bold.

• Less request boilerplate — sensible defaults kick in when styles and page settings are omitted.

• Cleaner generated output — only referenced fonts are loaded, keeping document size under control.

• font_weight replaces is_bold

• font_weight replaces is_bold

• PDF and EPUB render weights as CSS numeric values

• DOCX and PPTX map semibold and above to bold

• Default styles use PlusJakartaSans for body text, Outfit for headings, and A4 with standard margins

This is most useful when you generate client-facing reports, certificates, or other branded deliverables and want better typography without turning every request into a styling exercise.

Layout layers in Image Generation now support background_layers, so you can place solid colors, gradients, or images behind the main content without polluting the foreground layer stack.

Content-processing workflows usually break down in the handoff between tools. You extract data in one service, transform assets in another, generate the final deliverable somewhere else, and the glue code between them becomes its own maintenance problem.

Iteration Layer is live for teams that want those workflows in one place instead of spread across separate vendors and custom integration code.

What changed

• Pipeline over point tools — extraction, transformation, and generation can live in one workflow with one API style.
• One credit pool — usage stays in one billing model instead of being split across vendors.
• EU-hosted processing — files are processed on EU infrastructure with zero data retention by default.
• Agent-ready access — the same capabilities are exposed through an MCP server for tool-using agents.

APIs available at launch

• Document Extraction — turn PDFs, images, and office documents into structured JSON, with multi-file merge and field-level confidence scores.
• Document to Markdown — convert documents and mixed file inputs into clean markdown for search, extraction, agent, or generation workflows.
• Image Transformation — resize, crop, convert, upscale, remove backgrounds, and prepare images for downstream use without building your own media pipeline.
• Image Generation — generate deterministic, pixel-perfect images from JSON templates with text, images, shapes, QR codes, and reusable layouts.
• Document Generation — generate PDF, DOCX, EPUB, and PPTX files from structured data for reports, certificates, invoices, and other deliverables.
• Sheet Generation — generate CSV, Markdown, and XLSX spreadsheets with formatting, formulas, merged cells, and multi-sheet output.

Example workflows

• Extract structured data from a document, then generate a report from it
• Transform uploaded images, then generate branded assets around them
• Normalize mixed content inputs before passing them into an agent or automation workflow