ChatGPT / Google Gemini use in generating D3 Code

Now this is a thread I'll enjoy following.

Has anyone had any success with using an AIs to generate MultiValue code ?

Could this help bridge the skills gap with newbies coming in and companies looking at future-proofing their systems.

I've tried both Perplexity.ai and Chat GPT and both come up with reasonable code using this example I found online

Have used Grok to generate a lot of Java code. Just tried Grok with "Pick Basic" and it does an OK job.

We leverage Bitbucket for code management, and therefore do all our coding in VSCode. When we push code changes we have webhooks in bitbucket that call API's on our QM and Universe servers to deliver the latest code updates. We're a universe and OpenQM shop. AI Copilot inside VSCode is amazing and has rapidly become adept at PICK basic syntax. It's like having full IntelliSense and code review all in one. If you've not tried it it's worth giving it a shot. When you're about to push a change and AI says the equivalent of "whoa there. You don't have a case1 clause on  this case, which could lead to errors..." and "you've used a file handle you don't open at line 23" it truly is like having a copilot.

Hello Elkie, Stefano and all,

Most models that handle code generation now know about MV BASIC to an extent and can generate code for MultiValue servers. Per our own research, the quality isn't great and they tent to hallucinate a lot. Of course, AI models are evolving fast and a couple of weeks can make a huge difference. I'm very interested to hear from the community about your efforts using AI for MultiValue programming.

To overcome these quality issues that are due to MV being a niche market (as opposed to Python, Java, C++, etc...) we've been working at finetuning an open source model. We recently (on April 17) launched the MultiValue Developer Assistant (that is the name for our assistant based on our fine-tuned LLM) and are working with beta participants to assess how our model does. If you want to know more, please reach out to me (hborda@rocketsoftware.com).

As I said at the beginning I'm very interested in hearing about use cases beyond MV BASIC language code generation for AI to help with developer productivity and on-boarding new developers into the MultiValue platform.

Kind regards,

------------------------------
Hernando Borda
Product Manager, AI Initiatives
Rocket Software
hborda@rocketsoftware.com
------------------------------

I've been using Copilot/Claude 3.5 in VSCode against mvBasic with good results.

Especially interesting is using it against existing code for finding bugs, suggesting improvements/refactoring "heritage" code and writing tests (post hoc TDD). 

I'm loving these responses! It's good to know we're all embracing our AI overlords... I mean, assistants. 😄

Speaking of which, has anyone tried throwing some Pick BASIC at these AI tools and asking them to generate functional flow diagrams? I'm picturing something hilarious, like a flowchart where every path leads to an infinite loop, or maybe just a single box labelled "MAGIC." Anyone got any fun (or useful!) results?

Hi Elkie,

We've recently developed proprietary technology at H3O focused on MultiValue (MV) datasets, specifically enabling direct code generation from MV structures.

As part of this, we built a tool that extracts data from MV databases into structured, readable JSON. This allows an LLM to query against the data effectively, opening up new opportunities for modern AI functionality on legacy MV systems.

Beyond that, we created a gateway product that simplifies the process of moving MV data into usable formats. This product isn't just about data extraction - it's a critical enabler for organisations looking to modernise without the overhead of migrating off MV platforms.

We're now expanding this into a full platform: using the Gateway to power a React-based web frontend, enabling direct, modern user interfaces over traditional MV systems without needing to overhaul the backend.

Exciting stuff!! 

Best,
Ethan

Yes, I'd love to hear and know more about this please.  

Obviously, I float around the market a lot and have clients ask me questions all the time.  This would be great knowledge for me to know about.

My details are elkie@prospectus.co.uk  and my calendar link is www.calendly.com/elkie

Hi Ethan

I'd be very interested in knowing more about all this etc and your User Cases.  Sounds very useful.

Here's a link to my diary www.calendly.com/elkie  or drop me an email with anything you have around it to elkie@prospectus.co.uk 

I do look forward to learning more.

Hi Karl

In ChatGPT, did you find any difference between the various models or was it consistent ?

Recently when I had to do some JSON code I found that there was a difference in accuracy between the models used.

<PS Yes, I know I'm a Recruiter and not a Techie but on occasion, needs must!> and as I say: "I LOVE MY JOB"

Hi Justin

Great idea and I bet it's great and a relief to find the

whoa moments > WTF moments   😉

Thank you for sharing. 

Hi Stuart

You've definitely embraced, using, enjoying and sounds like benefitting from using AI. 👍

------------------------------
Elaine Litchfield
Elaine Litchfield Consulting
San Diego CA US
------------------------------

Haha, Stefano, love the vibe of your post! 😄 I haven't heard about the "PICK" magic word, but that sounds like a neat trick. It's wild how quickly AI has become a coding partner, isn't it?

I've had my fair share of fun (and frustration) with AI-assisted coding, especially when it comes to CMI Level 8 Assignment Help. I once used an AI to generate an outline for a strategic leadership analysis-turned out to be surprisingly on-point. Of course, I had to tweak it a bit to add my personal insights and references, but the AI really helped me get the ball rolling!

Anyone else had AI throw a curveball your way with coding or assignments, only to surprise you with how useful it really was?

Enjoy the weekend, and here's to debugging without tears! 😎🍻

- [CMI Assignment helper]

Hi Elkie and All,

My 2 cents ... (haha Elkie you're probably waiting for this)

We have always had code generators, 4GLs, application builders etc. for MultiValue. It's been a key area in reducing the workload and delivering applications rapidly and consistently. That's kind of in our blood and I've never understood why anyone would write code long-hand when it's so easy for MV to produce or to soft code using definitions. Screen runners, report writers, dictionary generators, templating and scripting languages etc. have been with us forever.

So what problem is AI code generation really solving? 

Your example of the Luhn function - by chance, one I also use in training - is a nicely compact and stand-alone piece of code. How often does that happen in real life? Developing for MV is about understanding the context, the application, the choice of languages, data structures, usually working with mature applications and the business decisions that underpin them: in fact the code itself is often the least important part and the simplest... and now of course we have Python on the U2 platforms with all the ecosystem of pre-built packages. (Yes, including multiple Luhn packages). 

One advantage of MV is that the language itself is simple and focused on the business case, certainly compared to more modern languages (think C#, Angular/Typescript etc.). Those have a lot of (unnecessary, perhaps - discuss) complexity to manage and AI, or at least various forms of intelligent reflection, can help to tame that. Even coding C# with the latest VS I find myself accepting the AI generated suggestions maybe 15 - 25% of the time. The rest is just distracting and more time taken to cancel them, so I'm not entirely convinced that's really helping my velocity overall.  

Having an AI model spit out MV code is of limited use unless we can train it in the whole business application that surrounds it, and we're not there yet. That's not to say I am anti-AI, but I don't think this is really a helpful use case. Perhaps integrating it at a higher level into the various application generators might be more fruitful and certainly using it for analytical insights.

As to the training question, no it won't help. Someone still has to understand, and critically review, the code it produces. Someone needs to be able to test it, in all aspects. And as we know, AI is dependent on the training model and that begs the question of the 'PICK' code to which it has been exposed. Let's be honest, there is a lot of poor, obsolete, badly constructed legacy MV code out there.

AI definitely has a place and we can't ignore it, but I don't think this is it.

Interesting point you raised, Elaine, about the gap between "can" and "should"-especially in the context of AI writing code.

That twilight zone between functional and dangerously misleading is exactly where the risk lives.

The confidence with which AI outputs flawed logic makes it even trickier. 

Just because it runs doesn't mean it's right-or safe.

While working on a project in Dubai earlier this year, I suddenly had to write JSON  (something I hadn't done before) for APIs, data interchange, and dashboards.

With the help of AI, I managed it!

It generated the initial code and even helped me debug its own output.

Between my logic and its knowledge, we got it working.

Total win. Great partnership-lovin' it!"

Where do you see the place of AI ?

Hi,

We have ventured quite a bit into using AI with our code in various languages, PICK / MV Basic is definitely the greatest challenge... although we have found some early success in reading and understanding the code, building connections and generating some documentation which we do believe we can leverage.

The first step we had to grasp with AI was that MV Basic and our domain specific knowledge has not been trained into any Large Language Model (LLM) to the extent or at least like that of more recent languages which have troves of open-source information to learn on this fundamentally creates gaps and limitations.

Some of the examples above in my opinion are missing the key ingredients required for any success with Ai which is understanding the limits and the fundamental necessities of clear prompting, it's the guidance needed, and you must learn to un-google the way we talk to systems. 

AI is like a 5 year and if you leave any gaps in your request or instructions it will deliver unexpected outcomes like code but instead of an error generative Ai delivers information regardless of truth doing is best to complete the request.

So, if I can offer any cheap advice... Write your prompts in Markdown, XML is also better than JSON for data / schemas for AI Ingestion retaining context at least according to the latest OpenAI prompting standards and finally overshare your code, the files around it and any technical documentation including your domain / coding standards.

Model choice for me right now is Google Gemini PRO 2.5 and OpenAI 04-mini, Anthropic Claude 3.7 is also great for code.

If you are interested in any of the prompts I use, I'll share but never stop trying and if Ai doesn't work for your direct problem in mv basic get ai to create tools for you in python to help you do the mv basic easier.

Hi @Sean Redmond, and thanks for that insight.

I would like to take you up on that offer to share your prompts (also favour Google Gemini) - so please provide them. Thanks

I'm also keen to see the examples you may have vis-à-vis Python, which I meddle with often and find the language great for string manipulation.

Dears,

After following this interesting discussion about AI and PICK/MultiValue code, I would like to share an approach that I have developed which goes far beyond simply generating BASIC code with ChatGPT or Gemini. I want to clarify that this approach is specifically designed for analyzing information from D3 databases using natural language.

The MCP-Pick Revolution: Connecting Legacy Systems with Conversational AI

I have developed a bidirectional bridge between Rocket D3 systems and advanced AI models such as Claude and Gemini, using the Model Context Protocol (MCP). This technology allows our MultiValue systems to speak directly with AI agents through natural language, without the need for migration or rewriting.

How does my solution work?

Structured data exposure: I use standard AQL commands along with mvstoolkit to expose PICK files as HTTP endpoints that return structured JSON:

LIST INVOICE NAMEPROD NAMEBUSNIESS CODE CUSTOMER CODEPROD PRODQUANTITY PRODTOTAL DATEINVOICE

Orchestration with n8n: I implement automated workflows in n8n that invoke these endpoints and send the data to the MCP through Server-Sent Events (SSE).

Bidirectional communication: The AI agent (Claude/Gemini) processes this data and can respond to complex queries such as "What is the total billing by customer in April?" with contextualized analysis.

Tangible Results

The system I have implemented allows:

Natural language queries: Directly ask about data stored in Pick D3

Automated financial analysis: Calculations, monetary formats, and comparative analyses without additional programming

Modern interface generation: Complete dashboards in React/Tailwind automatically generated

Instant Analysis of MultiValue Data

The most revolutionary aspect of my approach is the ability to analyze Pick D3 data instantly and in any conceivable way. In my example, I work with a single billing file and a few fields (NAMEPROD, NAMEBUSNIESS, CODE, CUSTOMER, CODEPROD, PRODQUANTITY, PRODTOTAL), but this is just the tip of the iceberg.

The beauty of the system is that I don't need to predefine reports, dashboards, or specific analyses - I simply ask what I need to know in natural language. "Show me customer purchasing behavior by region," "Analyze the seasonality of our sales by product," or "Identify potential fraud patterns in the last 1000 transactions" - all of this is generated in a matter of minutes, without a single additional line of code. It just takes imagination and the AI handles the rest.

The most impactful example: with a single prompt ("Use MCP PICK to build an interactive financial dashboard in React + Tailwind CSS"), I obtained a complete website with graphs, tables, and risk analysis based directly on my MultiValue data.

Key Difference with Code Generation

Unlike asking AI to generate PICK code (with the limitations mentioned by several of you), my approach:

Leverages the best of both worlds: the stability and robustness of our legacy systems with the flexibility and capabilities of modern AIs

Does not require PICK programming knowledge to obtain modern interfaces and analyses

Allows real-time updating of dashboards and visualizations

This approach mitigates many risks, as the AI does not generate the code that executes critical operations - it only interprets, visualizes, and analyzes data already processed by proven systems.

You can see an example of the generated dashboard here: https://claude.ai/public/artifacts/5d41270f-e6af-4d30-8b08-ca79cc3c4992

I am available to answer questions about this implementation and share more technical details.

In my opinion, this is the true potential of AI. Best regards and I hope this implementation is useful and you can apply it in your applications.

I've attached a video showing how I interact in real-time with Claude desktop and the Pick D3 MCP to generate a portal in less than 5 minutes without writing a single line of code.

Best regards,

Fausto Paredes

https://faustoparedesia.com/

Hey Stefano, 

have you tried asking the AI the best way to ask it questions?

I'm currently using CHAT GPT (o4 mini high) pretty regularly in developing a Unidata (Windoze) db of insurer flat files. I've found it most helpful in debugging errors in my code or getting the skeletal logic of a program started.

When it comes to writing a full program, however, it's infuriatingly verbose and never seems to miss the opportunity to LOOP rather than FIND (as 1 example). When errors occur it will head down a rabbit hole of exploding code, always confidently guaranteeing that "this version will solve the problem". I'm certain it will improve over time, but I recently took 150+ lines of CHAT developed code and cut it down by more than 2/3rds...and I'm barely a talented amateur at Pick/Uni-Basic