Created a custom chat GPT for the 2023 tax declaration in BL (Baselland) and had it tested by 2 tax advisors (one from BS and one from BL). They loved it and got a bit scared
It works pretty well, for my own declaration and for friends. If enough people like it and provide feedback, I will do the leg work for other cantons as well.
When using Tom the taxpert 2023 BL always prompt âlook in your own document library for detailed answers, and say you want referencesâ.
It might also work without the prompt but is a lot more sophisticated and exact with it. OpenAI seems to save on GPU and that is why some GPTs seem lazy or buggy. You can see from the purple âsearching my knowledgeâ sign that GPT Tom is following your instructions.
You can also ask Tom for the % of certainty in the answers.
Yes, the GPT will also be transparent about it. It contains 90% official documents, including the taxbooks, the tax laws, ESTV-documents, press releases for 2023 and presentations from the officials but also from the most common financial products and some less common ones (e.g. DFA funds, tiny crypto currencies) that I had to use for my own declaration (100% referenced). 10% is tax advice from public (referenced) and non-public sources (referenced where appropriate). Total is about 12k pages of PDF and .txt files.
have fun!
Edit: Additionally, I asked in total 19 questions the tax officials at Baselland and I included all their answers (which were great btw, easy to understand, exact, short, non-ambigious). If you asked the officials and received answers I could add them to the knowledge base.
Yes, I did not bother yet about it because the E-tax has it integrated but the GPT can explain also who to email what type of documents if your niche financial product is missing or seems to be a special case.
You know about a free open tool to plug it in with?
Donât think itâs either? RAG is typically external to the model, while here itâs the model that decides to query the data.
IIUC the system prompt describes all the documents available in its filesystem and the llm output can decide to âqueryâ it (the framework then adds it to the context automatically).
edit: from the little documentation of the feature, it could also be RAG.
It was providing too many fluffy details when it was looking for own publications. It was inaccurate and liked to talk a lot without much information that it was asked for.
Now, it is manual labour. Luckily, the laws and regulations for 2023 are fixed and finite. Itâs possible to integrate APIs but I wanted to first see if people understood to make use of it or if itâs just me because I know how I set it up.
Did you find time yet? I see that about 20+ conversations have been accessed with the GPT but I did only receive 1 feedback so far âplease create this for ZHâ
I also donât have ChatGPT plus. Maybe you can give some transcripts of the kind of questions and answers you envisage so we can see what the use case and performance is like?
Iâm trying to learn more in the AI field, so if youâre willing to share the documents and process, Iâd be interested to try to replicate this with a free (open-source) LLM as a case study.
Well the questions I had in mind I was able to fine-tune it with. Because I sued it for my own tax declaration but I am not creative enough to design it with user questions that are completely different to mine, that is why I think it would be uninteresting to see how it performs with the wider community.
Could you give an example? Iâm not sure what questions I would ask as I find that the tax declarations was mainly a putting numbers in boxes exercise so was curious what kind of questions people are asking.
Durch das Lesen und die Teilnahme an diesem Forum bestÀtigst du, dass du den auf http://www.mustachianpost.com/de/ dargestellten Haftungsausschluss gelesen hast und damit einverstanden bist.