Ability to customize initial prompt

[sadiogo]

Greetings!

I want to use the program to rename scientific papers instead of invoices. Where can I customize the initial prompt given to the AI? The only thing I found was the ability to extend the initial prompt, but not change it.

Can this be done?

[ptmrio]

You could edit the prompt directly in _ai_processing.py and recompile or run it through python directly.

Fork > Edit Prompt > pip install > python autorename-pdf.py file.pdf

HUGE however: still won't work, as the software expects some hard coded fields DocumentMetadata (document date, company name, ...).

I was thinking about generalizing, but this is a free open source side project, so it might take a while.
If you are curious, you could fork it and tell claude code to modify it accordingly, which probably will work fine.

[sadiogo]

Fork > Edit Prompt > pip install > python autorename-pdf.py file.pdf

Sorry for the dumb question, but what should I be installing in the pip install step?

HUGE however: still won't work, as the software expects some hard coded fields DocumentMetadata (document date, company name, ...).

Are these fields harded coded throughout all scripts? I can probably just try to change what is being returned for each field, e.g., the paper's title instead of company name.

How can I recompile the program?

[sadiogo]

Nevermind! I managed to successfully run the python script and recompile the program (at least the CLI... I'm still having trouble with the GUI). I will attempt to change the prompt and see how that goes.

[sadiogo]

I managed to extract the correct fields. Now I will work on building a harmonized list of journal names and abbreviations.

What function is responsible for building the final string of the pdf filename?
I want to change the order in which fields appear and add underscores in between them: date_author_journal_name

[ptmrio]

Hi again, I really appreciate your interest, but I'll have to remind you that this is a free, open source software project. You are free to modify it yourself, but this is not a commercial support. In your whole request, there is no please or thank you at all.

It might not be your intention, but your words/demands/requests sound rather rude, requesting assistance as if you paid for a premium SLA. I have a life aside of open source coding and need to make a living. This is also true for almost all other open source devs, many of them much more intelligent, capeable and giving than myself.

Nonetheless, we are living in great times nowadays, so I shall summarize the options you have:

1. Either you wait and give me some time, with proper please and thank you, so I might help you.
2. Or you request paid support, I'll happily send you an invoice. In this case, you can omit any friendly words, but it would still improve the experience.
3. And third, go pay $20 for a Claude code and instruct it to change it for you.

The feature request is noted, and may be implemented once enough requests/demand is noticed or appropriate sponsorship is contributed.

[sadiogo]

Sorry, I had no intention of being rude, nor had I any idea that you might interpret my comments as such. I was just genuinely interested in making the program work for scientific journals, but I can see now that I might have been a bit too pragmatic on my approach. My comments were not meant as requests which I expected you to respond promptly, or even respond at all (but yes, I was rather thankful that you did reply promptly; my bad for not demonstrating this explicitly).

I will continue working on this feature the best I can. Feel free to close this discussion.

[ptmrio]

I really appreciate your answer, and I'm glad we got this out of the world. Thank you very much. It takes courage to write an answer like that. Thank you.

I'll try to look into the matter deeper as soon as I got some time, as I was thinking about optimizing the prompt - even before you asked - to do different tasks. Maybe next week.

[sadiogo]

It's all good, mate! I totally understood your perspective. I've been looking for a program like this for quite a while, so I got a bit carried away in my enthusiasm; I acknowledge this now.

I also believe that we should strive to be kinder to one another, especially in online environments, where we have become increasingly accustomed to conversing with chatbots.

It's great to hear that you'll be generalizing the program! I think this will make all the difference going forward. In any case, the initial directions you provided were very helpful. The program is well-documented and I was able to find sufficient information in the README file on how to tinker with the program. It also encouraged me to learn more about "AI-ready programs" and "configuration and skills files". I have never used Claude Code, but yesterday I watched about 2 hours of videos on how to get the most out of it. So thank you for documenting the program and putting me on that direction!

Even without employing Claude Code I was able to achieve outstanding results. I basically kept document_date and replaced company_name and document_type for journal_name and first_author_surname, respectively. I did this throughout all the files. I should also point out that naming the variables according to the fields I wanted to extract was VERY important for achieving good results (at least for the low-resource models I was working with). (This probably means you'll need change the variable names dynamically to generalize the program; this is annoying to do with python, but possible. Of course, it is better to try something like field_1, field_2 first, but I just wanted to share my experience.)

I also did some prompt engineering, but the real "game changer" was using the gemma3:4b model with Ollama (I wanted to run the program locally). I tested several low-resource models (about 10-12 of them) and gemma3:4b simply nailed it! It was much faster than the other models and was able to resolve even the toughest cases while only needing to read 1 pdf page. There was even one case where it named the file using one of the middle authors, but it turned out that he was the correponding author; this information was written in minuscule letters on the footnote. I also noticed that chain-of-thought/reasoning models are not very efficient for this task, at least not for running the program locally; they take forever to rename even a single file.

To sum it up: I am pretty satisfied with how the program is running right now (for the purpose of renaming scientific articles), but there are still some improvements I want to implement, such as harmonizing the journal names and changing the order of fields in the filename.

Nonetheless, I look forward to testing your generalized version of the program, as I believe you will do a much better job than I did!

One suggestion: once you generalize the program, it might be a good idea to create a Discussion forum where users can post their best prompts and models for achieving a certain renaming task. Once enough of these are available, you can create a database and maybe incorporate this into a future version of the program to automate the most used renaming tasks.

All the best to you,
have a good one!

[qdm12]

(Hopefully not hijacking the issue too much)

First of all, I love the project and finally giving up on manual renaming is ❤️

I have my little convention to name all my files as yyyy-mm-dd <provider> <some-useful-short-title>. This sounds like it would require a custom initial prompt;

On top of this, in my case, some documents have dates as dd/mm/yyyy and some have it as mm/dd/yyyy (god damn it 🇺🇸 ). It would be pretty amazing if the program can understand the date format according to the country the document is from, or just if it's not a possible date i.e. 13/01/2026 means 13 january 2026, since 13 > 12. That would likely require some LLM to figure this out. Any suggestion on how this should be approached? More than happy to contribute with a PR to add such feature, but I just thought I would poke you first.

Thanks!

[ptmrio]

Hi @qdm12

The LLM should automatically recognize the date format. More intelligent/modern/expensive models typically should recognize it better. If you have some typical documents, you could hint it even more, like with the following prompt extension, which is easily configurable in config.yaml.

prompt_extension: "Make sure you get the document date correct whenever the document is sent from a U.S. company, it's likely to be U.S. format. Make sure to always sanity check if the document might use U.S. date formatting and extract accordingly."

As for the slightly different date format document naming, this would indeed need some rewriting. I'm in the midst of planning, but it is a heavy code rewrite to make this more universally configurable. The hard part especially is the newly introduced GUI, but I'm already in the works.

[qdm12]

thanks! indeed I did sort of miss the prompt_extension my apologies. Date-wise larger models do well most of the time, llama3.2-vision is decent and qwen2.5vl does really well but that's 22GB of memory used.
on the other side, they both tend to hallucinate dates when there is an incomplete date, like flight tickets have just DEC10 without the year so they put some random recent year 😄 but that might be just more "prompt engineering". Ideally we could do some old programming to figure this out without AI but I gotta admit it's also tempting to just throw more vram at it 😄!

edit: tldr this prompt field is plenty awesome, no need to add more really from my point of view, sorry for the half useless comment

[ptmrio]

All good, thanks for the input. Just a hint: gemma4 should work great, way better then qwen etc.

[austriaka]

Hi Gerhard,
I came across your program and managed to install it with ollama as local KI. I use it to rename scanned bills (Kassenbons) to show date and company names which works like a charm and does a better job than a payed program I tried -- thanks for that!
But when I tried to add the company location (city) within the file name via the prompt extension, it doesn't work. Whatever I use as a prompt, it seems not to read it at all, since I tried the prompt outside of your program and it worked. Can you tell me please what I am missing?
Thanks and greetings to Austria!

[ptmrio]

Could you please paste the exact prompt extension?

[austriaka]

prompt_extension: "The name of the city must be included if present in the text. The name of the city usually follows after the 5 digit code of the german ZIP code like '29378 Wittingen' where the city is Wittingen. If more than one city is present, take the first one. It must appear at the end of the filename, separated by an underscore (_). Filename format: 'Date_Company_City.pdf' (if city name is available). 'Date_Company.pdf' (if no city). Don't add any additional words like 'receipt', 'AR' or 'ER'."

The result of my last renames were "2025_02_09 BurgerKing ER.pdf" and "2025_06_01 McDonalds Receipt.pdf". Beside the harmonizing worked great.
(BTW thanks for the tip with gemma4, it is much faster. Still 2 minutes per file, but faster than qwen3 which took up to 5 minutes per file)