TEKLIA: Open-Source Arkindex for Archives and Libraries
On This Page
TEKLIA is a French company that builds and sells services around Arkindex, an open-source platform for handwriting recognition, layout analysis, text recognition and indexing of large collections of scanned documents. Its customers are mostly archives, libraries, museums and research projects. It is not a tool for invoices or forms in a business process.
How this profile came about: Clémence Fontaine requested the listing through the contribute form on 5 November 2026. IDP Software wrote the profile from TEKLIA's website, its documentation, package registries and Hugging Face. Figures marked as vendor claims have not been verified independently. About the author: Clémence Fontaine.
Company and ownership
The company page says TEKLIA was founded in 2014 by Christopher Kermorvant, who is CEO. The legal notice names TEKLIA SAS, registered in Paris, and says the website is hosted in the European Union. There are offices in Paris and in Grenoble. The team page lists 13 people, among them a CTO (Bastien Abadie), a COO (Yoann Schneider), several machine learning scientists and engineers, and one open role for a Python/Django developer. Funding and ownership are not published. The LinkedIn company page gives 2015 as founding year; the website says 2014. Contact is through the contact form.
The company says it puts a large share of its work into collaborative research projects in France, Europe and Canada. Its first project funded by the French National Research Agency (ANR) was HORAE in 2018.
Where TEKLIA fits / limits
TEKLIA fits projects where the documents are old, handwritten or poorly structured, the volume runs to thousands or millions of pages, and the output is an index, a catalogue or a research dataset rather than a posted invoice. Typical examples on its references page: handwritten registers, census returns, index cards, loan registers and auction catalogues. The open-source licence also fits institutions that must keep documents on their own servers and want to inspect or change the code.
It is a weaker fit in these cases:
- Business documents. There is no invoice, receipt or form extraction product. Teams with those needs should look at the IDP vendor directory, for example Planet AI for handwriting in business forms, and the guide to invoice processing.
- Proprietary products built on Arkindex. The Community edition is AGPL-v3. TEKLIA's licensing page says "All projects using Arkindex must also be open-source"; closed projects need the Enterprise edition. Have your legal team read the licence before you decide.
- Teams without operations staff. Self-hosting means Docker on Linux, an IIIF image server, S3-compatible storage and an SMTP account. TEKLIA also offers hosted and managed options, at a price.
- Fixed-price buying. The pricing for services and AI models depends on person-months and volume. Only the Arkindex licence has published prices.
- Published accuracy. The website states no character or word error rates. The research papers do, for specific datasets, but they do not predict results on your documents.
How TEKLIA processes documents
Arkindex is a pipeline manager. Documents (JPEG, TIFF, PNG, PDF or IIIF manifests) are imported into a corpus with a freely defined hierarchy. Processing steps run as "workers", which are Docker images with Python code, and each step stores its output and history. The Arkindex page lists the models it can run:
| Task | Models named by the vendor |
|---|---|
| Layout analysis and segmentation | Doc-UFCN, MaskRCNN, YOLO, Kraken, Grounding DINO, TableTransformer, LayoutParser |
| Text recognition (OCR and handwriting) | Tesseract, PyLaia, Kraken, DAN, Google Vision |
| Named-entity recognition | spaCy, Flair, Stanza |
| Other | Image classification (ResNet, YOLO), image description (Llava), translation (MarianMT), LLMs (Qwen, OpenAI, Claude, Gemini) |
The vendor says it does not build a one-size-fits-all model. Its project method starts with a benchmark on about 10 pages, then ground truth collection, evaluation, fine-tuning or training of custom models, and production with human review. The page names Doc-UFCN and PyLaia as its own libraries; the other models come from third parties. Data can be corrected in the annotation interface or in Callico, TEKLIA's annotation platform. Tasks can be monitored and re-run per element. Ocelus is a separate, older web service for handwriting recognition (French, English, Arabic in 2021); the current website does not describe it.
Open source: licence, code, models and papers
Licence. Arkindex went open source on 15 April 2024 with version 1.6.0, under the AGPL-v3. The Community edition is the open-source version. Per the licensing page, it lacks project access control, slurm support, network-segregated ML tasks, snapshot-based exports and secrets management. The Enterprise edition has them and is licensed commercially. The documentation is under GFDL. The Arkindex product page itself does not name the licence; you find it in the blog post and the documentation.
Code. The source is on TEKLIA's own GitLab at gitlab.teklia.com/arkindex (backend in Python, frontend in Vue.js). The GitLab server uses bot protection, so some pages need a real browser. Around it, TEKLIA lists these open tools on its open-source page:
| Tool | Purpose | Where |
|---|---|---|
| PyLaia | Handwriting recognition library; PyPI lists version 1.1.2 under MIT | PyPI |
| Doc-UFCN | Object and line detection in scanned documents | PyPI |
| Callico | Annotation and validation platform; documentation | GitLab |
| arkindex-cli, arkindex-client, arkindex-export, arkindex-base-worker | Command line, API client, SQLite export reader, base class for workers | PyPI |
| Nerval | Evaluation of named-entity extraction | GitLab |
The Arkindex packages show regular releases in 2026: arkindex-cli 0.5.0 in May and arkindex-base-worker 0.6.0 in July. The documentation lists Arkindex releases up to 1.13.1 but shows no dates. A public demo and a support forum are available. The open-source page links a guide for contributing code. The number of outside contributors was not counted.
Models and data. The Hugging Face organisation holds 27 models and 21 datasets (state on 9 November 2026). The models are mainly PyLaia handwriting models (for example for RIMES, NorHand, Belfort and Esposalles), Doc-UFCN and YOLO layout models, spaCy entity models and, for the DAI-CReTDHI project, fine-tuned Qwen2.5-VL models. Datasets include RIMES-2011, NorHand, KHATT and an ATR benchmark. The licence per model is not shown on the organisation page; check each model card before commercial use.
Papers. The publications page lists over 90 items, including ICDAR 2024 papers on Callico and on language models in PyLaia, and journal articles in IJDAR and Pattern Recognition Letters. Staff such as Mélodie Boillet and Solène Tarride appear as authors, and two PhD defences are announced on the blog. This research base is what separates TEKLIA from many providers that only reuse public models.
Community. Support for the Community edition is a public forum; private forum and tickets belong to the paid editions. The documentation describes importing a Transkribus collection, and the open-source page lists a virtual keyboard extension for eScriptorium. This makes Arkindex usable next to those tools.
Technical specifications
| Feature | Specification |
|---|---|
| Document Types | Scanned and digitised historical and administrative documents: registers, census returns, index cards, letters, newspapers, excavation notebooks, photographs |
| Input Formats | JPEG, TIFF, PNG, PDF, IIIF manifests; JP2 named on the technology page |
| Output Formats | PDF, CSV, PAGE XML, ALTO, Docx, SQLite; METS, TEI and JSON named on the technology page |
| Processing Pipeline | Import, layout analysis, text recognition, named-entity recognition, classification, indexing, export; human correction in the interface or in Callico |
| API/Integration | Documented REST API, command line client, Python API client, worker SDK; custom workers as Docker images; OpenID Connect |
| Deployment Options | Self-hosted with Docker on Linux; TEKLIA-managed on-premises; dedicated TEKLIA-hosted instance |
| Certifications | None named; site claims "strict GDPR compliance" |
| Claimed Accuracy | None published on the website; research papers report results per dataset |
Deployment, security and pricing
Prices for Arkindex editions are on the Arkindex page:
| Edition | Price | Hosting | Support |
|---|---|---|---|
| Community (open source) | Free | Self-hosted | Public forum |
| Academia | 1,500 EUR per month | Self-hosted or TEKLIA-hosted | Private forum |
| Enterprise | 4,000 EUR per month | Self-hosted or TEKLIA-hosted | Private tickets |
The page does not say whether the prices include VAT. Academia is a discounted licence for university research projects. The services page describes three tiers for services, infrastructure and AI models: free (open-source models, you host), service-based, and full-service. Prices there depend on person-months and volume and are not published. Enterprise servers contact a remote licence server a few times a day. Self-hosting suits projects whose documents cannot go to public instances for privacy or legal reasons. The website lists no security certificate and no hosting provider for the TEKLIA-hosted instances; ask for both.
Evidence
The open-source side is verifiable: the licence is named in the documentation, packages and models are public and recent, and research papers carry DOIs. The references page lists 22 projects with country, years and a one-line scope, including the Louvre, the Library of Congress, the National Library of Norway, the Service Historique de la Défense and INED. These are TEKLIA's own statements, not confirmed with the institutions, and the pages give no volumes or accuracy per project.
Other statements are claims without method: 20 million images, 33.7 billion characters, "6x faster than manual entry" and "90% savings compared to manual processing" on the homepage have no basis, date or customer. "Over 20 years of research and development" and "european leader" sit next to "since 2014" and "over 10 years" on other pages. Treat all of these as vendor claims and run a pilot on a sample of your own documents; TEKLIA itself starts projects with a benchmark on about 10 pages.
Resources
- Website
- Arkindex and documentation
- Deployment guide
- Source code on GitLab
- Open-source tools
- Hugging Face models and datasets
- Publications
- References
- YouTube
- Atom feed (blog)
- Related on this site: Open-source OCR tools, handwriting recognition tools, document archiving solutions
Company Information
TEKLIA SAS is registered in Paris, with offices in Paris (30 rue Raymond Losserand, 75014) and Grenoble (La Turbine, 5 esplanade Andry Farcy, 38000). It was founded in 2014 according to its website. The team page lists 13 people. Funding, ownership and revenue are not published.