Tried with Ollama and Deepseek-r1

[nabbl]

Hi,

So I was successful in setting everything up and I can chat with my ollama deepseek-r1 model via the paperless-ai gui.

However it has trouble analyzing the provided file (it is a pdf and the ocr from paperless was pretty spot on). It was able to read parts of it but couldn't manage to answer in more detail:

Alright, so I need to figure out what the document is talking about. The user greeted me in German and asked what's going on with it. I responded by saying that without seeing the content, I can't provide a specific answer.
Okay, since the user is asking about the topic of a document they provided earlier, but now it seems like there might be some confusion because the initial query didn't include an attachment or any specific details. Maybe the user sent the question after sending the document but didn't attach it properly, or perhaps they only included text.

I should make sure to inform them that I can't access external content and ask for more information. It's important to prompt them to provide specifics like headings, keywords, or key points from the document so I can assist better. Also, if there were images or charts, mentioning those would be helpful.

Maybe I'll phrase it in a friendly and supportive way to encourage them to share whatever details they have. That way, I can help them understand how to proceed or what information is needed for an accurate response.

Ihr Dokument seems to be about an unspecified topic since no specific content was provided. To assist you effectively, please provide more details such as the main subject, key points, headings, keywords, or any relevant context from your document. This will enable me to offer a more precise and helpful response.

After inquiring more I get this response:

Okay, so the user sent me this pretty detailed query about a Kia electric car and then added that they provided a document. Hmm, but when I look at their message, it's all in one block without any separation between their initial question and the additional info.
Wait, maybe they didn't realize how to format their messages properly here. They probably intended to ask something specific about the car details they provided. So my response should acknowledge that they included a document but also make sure I understand what exactly they need help with.

I should check if there's any specific information they want me to focus on, like the electric range, CO2 emissions, or maybe something else from the document. It seems they might be looking for assistance interpreting or analyzing the provided data about their Kia EV9.

Also, the user mentioned that the CO2 emissions are 0 g/km, which is interesting because traditional combustion engines have high CO2 output. So perhaps they're curious about why this is happening with an electric vehicle and what implications that has on fuel costs over time.

I should also consider if there's any regulatory information or tax benefits since the steuerbefreiung (tax holiday) was mentioned. The user might be looking into whether this car could save them money in taxes or running costs over a few years.

Another thing to think about is the upfront cost versus the long-term savings with an electric vehicle compared to a traditional one. Maybe they're evaluating if it's worth switching to electric based on their driving habits and budget.

I need to make sure my response is clear, friendly, and open-ended so they know I'm here to help them explore whatever aspects of the document they're interested in.

It seems you've provided a detailed document about a Kia EV9. However, your message appears to be a bit disjointed and may not have been formatted properly. Could you clarify or provide more specific details about what you'd like me to focus on? For example, are you asking for an interpretation of the data, an analysis, or something else related to this document? Let me know how I can assist!

With which models was paperless ai tested? Maybe deepseek just doesn't do a good job with it?
The response doesn't seem to be very good:

No tags or correspondent found in response from Ollama for Document.
Please review your prompt or switch to OpenAI for better results.

Realizing that Deepseek-r1 might be a bad choice for this I tested llama 3.1 with similar results.
It seems that the AI doesn't get the whole ocr'd document as a prompt

[clusterzx]

I tested all of them and all did pretty well.

[gima84]

Deepseek-r1 is not ready to respond with a structured JSON. So automation won't work with paperless-ai I think (I tried and nothing happened, as expected).

[clusterzx]

I used deepseek chat, not r1.

[iyoschi]

Have testet both. But GPT 4o gave significantly better results. For around 100 normal documents without excess length, i paid 5 cents in tokens, absolutely fine! With Ollama: I was wondering, did it include anything at all? :D