AUTO_TAG is not working properly

[Staubgeborener]

I am running the latest paperless-gpt docker version inside an urnaid setup (ollama is calculating on a i5-12400).

I leave the AUTO_TAG variable as default with paperless-gpt-auto.

Now i add the tag paperless-gpt-auto to some documents and start my ollama and paperless-gpt docker container. What i would expect: paperless-gpt immediately starts the run and tags the documents. What really happens: I wait a long time, sometimes 15 or 20 minutes, until one document gets tags. Sometimes i can wait for HOURS and nothing happens.

To get a feeling, my log output looks like this:

Using paperless-gpt as manual tag
Using paperless-gpt-auto as auto tag

    ╔═══════════════════════════════════════╗
    ║             Paperless GPT             ║
    ╚═══════════════════════════════════════╝

Version: v0.23.0
Commit: fecd76a5bf14f8183b53d874b31b0121981a9779
Build Date: 1758104151
Platform: linux/amd64
Go Version: go1.25.0
Started: Fri, 05 Dec 2025 16:24:01 UTC

Using paperless-gpt-ocr as manual OCR tag
Using paperless-gpt-ocr-auto as auto OCR tag
time="2025-12-05T16:24:01Z" level=info msg="Worker 0 started"
[...]
time="2025-12-05T16:31:35Z" level=info msg="Suggested title for document 20: XXX" document_id=20
time="2025-12-05T16:31:35Z" level=info msg="Suggested tags for document 20: [XXX]" document_id=20
time="2025-12-05T16:31:35Z" level=info msg="Suggested correspondent for document 20: Ich habe die Anfrage nicht vollständig verstanden, da sie sich auf einen komplexeren Text bezieht und es keine spezifische Fragestellung gibt, um eine direkte Antwort zu geben. Könnten Sie mir bitte den Inhalt des Dokuments präziser wiederholen oder die spezifischen Informationen angeben, nach denen ich suchen soll?"
time="2025-12-05T16:31:35Z" level=info msg="Suggested createdDate for document 20: Die aktuelle Datum und Uhrzeit sind nicht verfügbar, daher kann ich keine spezifische Antwort geben. Bitte beachten Sie die folgenden Informationen:\n\n- Das Format des Textes ist eindeutig in deutscher Sprache geändert worden, um den Anweisungen zu folgen. [...] Verwenden Sie keine Fachjargon oder komplexe Terminologie ohne Erklärungen. \n\nDie Antwort soll in einem professionellen und ansprechenden Marketing-Leitartikel formatiert sein, der auf Deutsch verfasst wird und die spezifischen Anforderungen des Kunden berücksichtigt."
time="2025-12-05T16:31:35Z" level=info msg="Document 20 processed successfully. Runtime: 00:07:33" document_id=20
time="2025-12-05T16:31:35Z" level=error msg="error updating document 20: error creating correspondent 'Ich habe die Anfrage nicht vollständig verstanden, da sie sich auf einen komplexeren Text bezieht und es keine spezifische Fragestellung gibt, um eine direkte Antwort zu geben. Könnten Sie mir bitte den Inhalt des Dokuments präziser wiederholen oder die spezifischen Informationen angeben, nach denen ich suchen soll?': error creating correspondent: 400, {\"name\":[\"Ensure this field has no more than 128 characters.\"]}" document_id=20
time="2025-12-05T16:31:35Z" level=info msg="Processing document for auto-tagging" document_id=24
time="2025-12-05T16:31:35Z" level=info msg="Processing Document ID 24..." document_id=24

Taking a look into the paperless-ngx dashboard: Nothing happens. No new tags. Translating the text above in english leads to:

Suggested correspondent for document 20: I did not fully understand the request, as it refers to a more complex text and there is no specific question to provide a direct answer to. Could you please repeat the content of the document more precisely or specify the specific information I should look for?
time=“2025-12-05T16:31:35Z” level=info msg="Suggested createdDate for document 20: The current date and time are not available, so I cannot give a specific answer. Please note the following information:\n\n- The format of the text has been clearly changed to German in order to follow the instructions. [... ] Do not use jargon or complex terminology without explanations. \n\nThe answer should be formatted as a professional and appealing marketing editorial written in German, taking into account the specific requirements of the customer."
time=“2025-12-05T16:31:35Z” level=info msg="Document 20 processed successfully. Runtime: 00:07:33" document_id=20
time=“2025-12-05T16:31:35Z” level=error msg="error updating document 20: error creating correspondent 'I did not fully understand the request, as it refers to a more complex text and there is no specific question to provide a direct answer. Could you please repeat the content of the document more precisely or specify the specific information you want me to look for?': error creating correspondent: 400, {\“name\”:[\“Ensure this field has no more than 128 characters.\”]}" document_id=20
time=“2025-12-05T16:31:35Z” level=info msg="Processing document for auto-tagging" document_id=24
time="2025-12-05T16:31:35Z" level=info msg="Processing Document ID 24..." document_id=24

So uhm, is my llm having issues? I am running minicpm-v as VISION_LLM_MODEL.

But nevertheless most of the time i only see this log entry

Using paperless-gpt as manual tag
Using paperless-gpt-auto as auto tag

    ╔═══════════════════════════════════════╗
    ║             Paperless GPT             ║
    ╚═══════════════════════════════════════╝

Version: v0.23.0
Commit: fecd76a5bf14f8183b53d874b31b0121981a9779
Build Date: 1758104151
Platform: linux/amd64
Go Version: go1.25.0
Started: Fri, 05 Dec 2025 17:01:45 UTC

Using paperless-gpt-ocr as manual OCR tag
Using paperless-gpt-ocr-auto as auto OCR tag
time="2025-12-05T17:01:45Z" level=info msg="Worker 0 started"

... and nothing happens.

[Anaximelis]

Your LLM does not respond as instructed with the name of the correspondent, but with a long text. Apparently, the context window of the LLM has overflowed, causing it to forget the original instruction.

Use TOKEN_LIMIT to reduce your input size.

[Staubgeborener]

That would make sense. Is TOKEN_LIMIT only for LLM_MODEL, VISION_LLM_MODEL or both at the same time? Also, is there a good source for the token limit? Right now i am using minicpm-v as LLM_MODEL and deepseek-r1:8b as VISION_LLM_MODEL. I would try something like 8000 i guess.

Edit: Doesn't helped at all, i get this error

2025/12/07 12:41:34 [WARN] Failed to calculate number of tokens for model, falling back to approximate count

I found this issue about this topic:

[...] any models that don't use that specific tokenizer will fail to calculate