[Feature Request] Auto generate document title with AI support

[beber-msft]

Description

Loop in OpenAI API or other components to automatically create and fill title field.

Other

No response

[grostim]

Should be pretty easy to do, but i am not fully conviced by this proposal as this require to rely on external service and an openAI API will be needed.
I would rather see this done with a post-consume script.

[github-Jan]

Don't like to have a "expose my private parts documents to OpenAI"

[thiswillbeyourgithub]

there are interesting projects like litellm that provides a unique library handling dozens of API including "local" API if you're yourself running Ollama or localai.

Given the pace of developpment and compressing of LLMs by the end of the year we can expect current consumer hardware to be completely up to the task to many of the things done in paperless-ngx : better tagging, better search via embeddings, more importantly better date matching (no more thinking the date is your birthdate :) )

[jms92100]

I have done this as a preconsume script .
i use lm studio software wich launch the mistral ai in server mode. (lm studio does work on your machine and you need a good graphic card to make it work correctly. I have a 3080)
a python script then asks for the name of any scan in an input folder and transfert it to the output folder with the name given by the ai.
Then a process cron or whatever as to transfert it to the consume folder of paperless.
and voila
pro , it is automatic
cons it is not always perfect due to bad document scan
cons date of the document is not easy to ensure if multiples dates are identified

[jms92100]

i use mistral with LM Studio wich can be used for chatting or run as a server . So far i am very happy with the function as it changes the default scan value to something relevant for me and it is only working locally !
tell me if you can go further .

[sjafferali]

I also did this with a post-consume script: https://github.com/sjafferali/paperless-titles-from-ai

[grostim]

@sjafferali : i tried the prompt you use in chatgpt3.5 with the ocr content of some of my documents: impressive results.
I will try your script this week end.

[grostim]

@sjafferali Tested your post-consume script . Works quite well for me, even if i need to fine tune your prompt for my needs.

Do you know if there is a way to run the post-consume script on existing documents ?

[sjafferali]

@grostim I just pushed a commit that includes a script to allow running this on existing paperless documents to in order to backfill titles.

The documentation has details on how to run it. Feel free to give it a try.

I also added support for additional environment variables to allow overriding the default prompt, which you may find useful.