Comparison at a glance
As documented by each vendor, checked 14 Nov 2026. “Not documented” means we found no statement on the vendor’s official pages.
| Type | Local / self-hosted | EU processing | Output | |
|---|---|---|---|---|
| LlamaParse | Managed parsing service | PartlyBYOC and on-premises options | YesEU region, AWS eu-central-1 | Markdown (default), HTML tables, XLSX and more |
| Docling | Open source (MIT) | YesYes, incl. air-gapped | YesRuns locally | Markdown, HTML, JSON |
| Unstructured | Open source (Apache-2.0) and platform | YesLibrary and Docker | YesDedicated instances in EU regions | md, json, html, txt |
| Azure AI Document Intelligence | Cloud API | PartlyContainers; disconnected on request | YesProcessed in the resource’s region | JSON; Markdown for layout |
| Google Cloud Document AI | Cloud API | Not documented | Yeseu multi-region | Document object in JSON |
| AWS Textract | Cloud API | Not documented | YesFrankfurt and Ireland endpoints | JSON Block objects |
| Mistral OCR | API | PartlySelf-hosted container for enterprise customers | YesEU by default | Markdown per page |
| PaperOffice AI IDP | Cloud API and MCP server | PartlyStorage can stay on your own NAS/S3 (BYOS) | YesOwn data center in the EU | Extracted fields with source location, via API |
The options in detail
LlamaParse
LlamaIndex’s parsing service. Besides the managed service there are BYOC and on-premises options [2], and an EU region runs on AWS eu-central-1 [1]. Markdown is the default output, with HTML tables, XLSX and other formats available [3].
Best for: teams already using LlamaIndex that want to stay with it and need the EU region.
Docling
An open-source toolkit (MIT) started by IBM Research and hosted by the LF AI & Data Foundation. It runs locally, including air-gapped environments, and exports Markdown, HTML and JSON [4].
Best for: developers who want parsing fully inside their own infrastructure.
Unstructured
An open-source library (Apache-2.0) that runs via pip or Docker [5], plus a platform with dedicated instances in AWS and Azure EU regions [6]. Output formats include md, json, html and txt [7].
Best for: pipelines that need many file types and want a choice between self-hosting and a dedicated EU instance.
Azure AI Document Intelligence
Microsoft’s document API processes incoming data in the region where the resource was created [9]. It can also run in containers, including disconnected containers on request [8], and the layout model outputs JSON or Markdown [10].
Best for: Azure customers who want a managed API in a chosen EU region.
Google Cloud Document AI
Google’s document processing offers an eu multi-region and several European single regions [11]; results are returned as a document object in JSON [12].
Best for: Google Cloud users who need EU processing for document extraction.
AWS Textract
Amazon’s OCR and extraction service has endpoints in Europe (Frankfurt) and Europe (Ireland) [13] and returns a list of JSON Block objects [14].
Best for: AWS users who want text, tables and forms as structured JSON.
Mistral OCR
Mistral hosts data in the European Union by default [16]; OCR results are returned as Markdown per page [17], and OCR 4 is available as a self-hosted container for enterprise customers [15].
Best for: teams that want an EU-based OCR model with Markdown output.
PaperOffice AI IDP
IDP agents extract fields such as supplier, VAT ID, IBAN, invoice number and line items, each with its source location in the document. Processing runs in PaperOffice’s own data center in the EU, storage can stay on your own NAS or S3 (BYOS), and the same functions are available via API and the MCP server.
Best for: business documents where you need fields and checks rather than raw Markdown.
How to choose
- Decide between raw text and business fields. Parsers such as Docling or LlamaParse produce Markdown for RAG; IDP services return fields such as invoice numbers or IBANs.
- Decide where processing may happen. Local tools keep everything in your infrastructure; cloud APIs name the regions in their documentation.
- Check the region in the documentation. Look for the exact region name, for example eu-central-1 or the eu multi-region, before you send documents.
Sources
Vendor information from official vendor pages, checked 14 Nov 2026. Product capabilities change; check the linked pages for the current state.
- LlamaParse: regions
- LlamaParse: deployment options
- LlamaParse: configuring parse
- Docling on GitHub
- Unstructured on GitHub
- Unstructured: dedicated instances
- Unstructured: output formats
- Azure AI Document Intelligence: containers
- Azure AI Document Intelligence: data, privacy and security
- Azure AI Document Intelligence: layout model
- Google Cloud Document AI: regions
- Google Cloud Document AI: layout parser quickstart
- AWS: Amazon Textract endpoints
- AWS: AnalyzeDocument API
- Mistral: OCR 4
- Mistral Help Center: where data is stored
- Mistral docs: OCR endpoint