Currently, I am build and open-source pipeline for archival documents processing and I want to include Arkindex into the workflow.
Can you please help me with the below questions?
I will be dealing mostly with letters in Russian. I already have some good results with Tesseract OCR processing of typescripts. But most letter are handwritten and I don’t see a suitable model for that in Arkindex. What should I do in this case?
Currently I mange all the letters in my Zotero library. Would it be possible to automatically take the PDFs from Zotero, do the transcription in Arkindex and and then save those transcriptions back to Zotero in the form of an Item attachment or note?
The current dataset is pretty small (approx. 400 letters / 700 pages). Can I do the processing on demo.arkindex.org in this case? Or should I go for the self-hosted option?
I would very much appreciate the access to Russian HTR model if possible – If it will be activated for me only, then please use the account racyn@phil.muni.cz on demo.arkindex.org.
Regarding the files – is there any other solution than manual upload and download of the PDFs? Would it be possible to pull the files from Zotero via API for example?
we can give you the access on Monday. Regarding the import, we don’t have a zotero connexion in Arkindex, so the PDF export is still the best solution.
I don’t have a permission to use metadata in my project:
“There are no allowed metadata for this project. Please contact your administrator.”
What is required to enable and use this feature?
I decided to test API connection further in my pipeline bypassing Zotero and linking Arkindex directly to Nodegoat app, where I do the main data analysis. I was able to pull data from one of the the public projects (Demo | IAM) during testing. However, when I tried the same thing with the element from my own project I got the below error:
{"detail":"Element afa63104-3561-4f2a-a0e2-4218a49df3a9 not found."}
Just to give you some context: the idea is that via API I will get the processed data from Arkindex and items from Zotero directly in the Nodegoat and then match Arkindex elements with Zotero Items in Nodegoat based on their unique ID pairs. In Zotero that would be the Item ID. In Arkindex I would like to store that same Zotero Item ID in metadata section.
I already have Zotero connected to Nodegoat. Now need to do the same thing for Arkindex and include Zotero Item ID in the Arkindex element metadata.
I am wondering whether you were able to provide me with access to the Russian HTR model already? If so, where should I look for it? I don’t see any new items in the list of workers.
You have now access to Microsoft OCR in Arkindex, and I credited your account with 5 euro. You can test it, you can specify the language to Russian with a configuration.
I tried to run the Microsoft OCR worker three times on two different files but always ended-up with task error (see warning in the line 3) below:
2025-09-24 09:16:53,190 INFO/arkindex_worker: Worker will use /data/current as working directory
2025-09-24 09:16:53,951 INFO/arkindex_worker: Loaded Worker Microsoft OCR @ main (123a5a) using configuration ‘Russian’ from API
2025-09-24 09:16:54,086 WARNING/arkindex_worker: Secret tmocr/api.yaml not available: {‘detail’: ‘No Secret matches the given query.’}
Traceback (most recent call last):
File “/usr/local/bin/worker-microsoft-ocr”, line 8, in
sys.exit(main())
^^^^^^
File “/usr/local/lib/python3.12/site-packages/worker_microsoft_ocr/ocr/worker.py”, line 169, in main
MicrosoftOCR(description=“Arkindex worker to use TMOCR transcriptions”).run()
File “/usr/local/lib/python3.12/site-packages/arkindex_worker/worker/init.py”, line 159, in run
self.configure()
File “/usr/local/lib/python3.12/site-packages/worker_microsoft_ocr/ocr/worker.py”, line 28, in configure
super().configure()
File “/usr/local/lib/python3.12/site-packages/arkindex_worker/worker/base.py”, line 341, in configure
self.configure_worker_run()
File “/usr/local/lib/python3.12/site-packages/arkindex_worker/worker/base.py”, line 280, in configure_worker_run
self.secrets = {name: self.load_secret(Path(name)) for name in required_secrets}
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File “/usr/local/lib/python3.12/site-packages/arkindex_worker/worker/base.py”, line 390, in load_secret
raise Exception(f"Secret {name} is not available on the API nor locally")
Exception: Secret tmocr/api.yaml is not available on the API nor locally
I did a brief testing of Microsoft OCR worker on handwritten letters in Russian. Unfortunately, the results are quite disappointing. Indeed, I was able to get better (but not good enough) detection with native Tesseract OCR worker when doing back-to-back testing on the sample document.
So I still do not have a suitable solution for the Russian HTR at this moment.
Meanwhile I got in touch with Achim Rabus from Freiburg University. He pointed me to a couple of promising models below. Could you please tell me whether it would be technically possible to use those as workers in Arkindex?
I am following-up this thread with a couple of updates.
First of all I was able to test trOCR – unfortunately, the result are far from perfect especially in comparison to commonly used LLMs. In this regard I had an excellent results with Google Gemini Pro 2.5 subscription and pretty good results with Qwen3-VL-30B-A3B (even without any additional tweaking).
Thus, I wanted to ask, where is it possible to use Gemini and/or Qwen models directly in Arkindex for HTR tasks? If so, how can I setup and access this models in Arkindex?
Is there someone who would be willing to help with the above question? I would like to know, where is it possible to use Gemini and/or Qwen models directly in Arkindex for HTR tasks? If so, how can I setup and access this models in Arkindex? Google just released Gemini 3 Pro and transcription results are very impressive in this model (I already have a subscription for Gemini).
we are updating the LMM worker to support Gemini3, we 'll let you know when and how you can test it. Regarding Qwen, we run it only internally because it requires a larger GPU. How did you test it ?
Regarding Qwen3 – I am testing Qwen3 Coder 480B Instruct model that runs via Open WebUI manitained by the CERIT-SC and e-INFRA CZ (I have access via Masaryk Unviersity).
