I have a worrying observation about the accuracy of Notebook Assistant (NA). I was doing some work in Fourier theory and there is a reasonably well known result involving the equality of negative exponential and the integral on the half plane of a weighted negative exponential and a Gaussian. Without going into all the details you get a transformed integral which has a nice known solution which Mathematica correctly outputs - it is in Gradsheytn and Ryzhik and Watson’s tome on Bessel functions and involves Bessel’s function of the second kind K_(1/2)(b) where b>0. It is not amenable to naive approaches. I tested ChatGPT 4o , DeepSeek R1 and Notebook Assistant (NA) by asking how to perform the integration. I got the usual pattern recognition from ChatGPT with arithmetic errors and nonsense about how to integrate it which was just wrong. After some Socratic prompting it finally got the correct answer. DeepSeek R1 delivered similar stuff but did not make the arithmetic errors in the detail and I took it through a similar Socratic process and it got the right answer. I then tried out NA which at one stage invoked WolframAlpha but nevertheless produced a howler of an arithmetic error. It is as through the computational aspect went out to lunch. NA got the wrong answer so I tried again. It correctly made reference to the BesselK function but got the wrong answer and when I pointed out that Mathematica gave the correct answer it manfully stood its ground with an answer which was out by a factor of $ \frac{ \sqrt{\pi}}{2} $ . Both ChatGPT and DeepSeek R1 actually got the correct answer finally but NA did not change its “opinion” when contested. This is not commercially acceptable for a system embedded in Mathematica for a separate non-trivial charge. So to step back - the LLMs without actual computational functionality eventually got the integral right but NA didn’t despite being told that Mathematica delivers the correct answer AND along the way it made arithmetic errors. There is a LOT more work needed to make NA actually useful. Here is the howler

Bradley - When I use ChatGPT through my browser, I can upload a PDF of a technical paper and the bot reads it instantly and communicates at near perfection to convert equations in the paper to Wolfram code. The problem is that after a bit the OpenAI link wants me to upgrade at a cost of $? or it shuts me off until tomorrow. Since I have a ChatGPT API key I presume that is what they want. I bought the Wolfram notebook assistant and assumed it - or a Chat Notebook - would provide the same functionality. Apparently not. In the Chat notebook I have selected OpenAI as the model and input my API Key. Even with that I am told to Import the pdf as plain text into the notebook. I can do that but the LaTex functionality is lost in the Plaintext conversion. Division signs are omitted and subscripts are ignored so differentials and exponents are not right unless I do a HUGE amount of tedious editing one character at a time. It seems like there is a LOOOOOONG step between what OpenAI offers directly and the version of it that comes through the notebook.

I would much prefer to go through Wolfram for these things. If I am doing something wrong, please advise. But the Wolfram Chat bot seems to be well below the OpenAI equivalent. I am happy to pay $20 a month to one or the other of you but it does not seem to make sense to do both since OpenAI will supply everything I need without the Notebook Assistant.

It is not associated with the plan. It is a limit of the Model. Whenever you are carrying on a conversation with the LLM, for example, in a Chatbook, it sends the full history of the conversation to the model. At some point, the number of tokens that are being sent to the model exceeds the limit of the model. You can see the various context limits of the OpenAI models here OpenAI Models

From your messages I am guessing that you are using one of the older gpt-4 models which have a total context limit of 8192 tokens

No. Grok is XAi (from Elon Musk). Groq is from a distinctly different vendor. I'm surprised this is not well known by now.

I'm not sure how I can directly make suggestions to the developers (other than ones that may read this forum on an uncertain frequency). I would think that a platform on the level of XAi would be on their radar without users needing to make suggestions. Even the just announced DeepSeek from the CCP is in the list you posted. You and Bradley have Wolfram Research in your profiles, maybe you could pass it along?

Hi David. I can confirm that the Wolfram Notebook Assistant + LLM Kit will work on both of your computers where you have Wolfram desktop products installed. It's access is tied to your Wolfram ID, not your activation codes, so it will work on any computer where you sign in.

Bradley -- I just found out about this and, late on a Saturday before Christmas, I excitedly purchased a year's supply. I have a license number for the product but do not find anywhere to register it. Chat enabled notebooks seem to work as usual. Should I imaging there is enough AI in it to connect my subscription to my license number. That brings up the question: Can I use it on my laptop with my Home License I get from Premier Support?

Bradley, question here please:

regarding the Essentials and Pro Suscription to NoteBook Assistant+LLM Kit; I assume both functionalities are included on both type of suscriptions, but a comment on FAQs on https://www.wolfram.com/notebook-assistant-llm-kit/, makes me doubt it:

"Wolfram Notebook Assistant is intended for human users; LLM Kit is intended for programmatic use. Pro should be sufficient for light programmatic use".

Can you confirm both NoteBook Assistant and LLM Kit functionalities are included on both Essentials and Pro Suscriptions? If so, what is the difference between Essentials and Pro suscription? Are there different plan limits between suscriptions? Could you explain them?

Thank you for clarification.

Thanks for the clarification! Regards

Very excited. How will this work in the cloud offering?

Well done with this! Within 10 minutes of using Notebook Assistant I was discovering improvements in coding a commercial solution that we're developing. Looking forward to all of the possibilities!

Thank you. Now it works . I still cannot get my DateListPlot with the sunrise time but it is another issue :-). What is fun is that the assistant produced the code and then corrected it without result. How human this trial, error ... Thanks again for your very fast answer.

Thanks for bringing this up. We've found the issue and deployed and update that should fix this for you. You can use ResourceFunction["InstallNotebookAssistant"][] to update all of the relevant paclets and then you'll need to restart the app afterward.

Can it be used in Country like Hong Kong ?