Creating a Custom GPT Bot Advisor for SPE-PRMS Reserves Guidelines

By now, you've likely experimented a bit with ChatGPT. If you haven't done so yet, you can start by reading my 'ChatGPT for Oil & Gas' blog series, starting with part 1 available here.

Today, we're going to explore CustomGPTs, which are tailored versions of AI language models designed to perform specific tasks.

OpenAI, the creator of ChatGPT, now allows anyone to create their own custom GPT with the data they want. This makes them more effective because they can understand and respond to queries in ways that are directly relevant to their particular use case.

In simple terms, you now have the ability to build your own ChatGPT, without any coding knowledge. Amazing, isn't it?

The best use cases are those where the data you want to train the bot on is contained in simple PDF files, which you just pass on wholesale, without any file processing.

The idea is for users to interact with the bot interface and ask direct questions about the file, without having to dig into the file itself. This creates a much better user experience, especially when the files to be consulted are large, or when you don't know exactly where to look, and the search process becomes a painful time-waster.

To demonstrate how anyone can build a Custom GPT, let's build a PRMS Guidelines Advisor as an example.

As you may know, the PRMS Guidelines are outlined and consolidated in a few PDF documents, the key ones being the 2007 PRMS foundational PDF, the 2011 PDF with extended guidelines and examples, and the 2018 Update. They are all publicly available for download from the SPE website at https://www.spe.org/en/industry/reserves/.

If you have ever been involved in a Resource and Reserves classification process, you most likely needed to consult these guidelines, and often, tracking down those elusive PDFs can be a hassle.

So, I thought this would be a good use case to have this advisor readily available in your pocket, where one can ask specific questions and have the bot scan through the content of the guidelines and return the most relevant answers.

Here’s how the process goes:

Head to https://chatgpt.com/gpts and click on ‘My GPTs’ in the top right corner:

Once inside, click on ‘Create a GPT’:

Here, you’ll find a simple interface with the left half dedicated to your bot builder, and the right half serving as your preview, where you can see how your creation is shaping up as you build it.

From here on out, you interact with the builder by having a conversation in plain English, just as you would with any regular ChatGPT instance.

Let’s ask:

I want to build a 'PRMS Advisor' to offer guidance on the SPE-PRMS Guidelines for Oil & Gas Resources and Reserves Classification. The PRMS Advisor will respond to user queries about the guidelines, drawing information from four key PDF documents I will provide. The answers must adhere strictly to the terminology and principles outlined in these documents, specifying which version the advice is based on. For inquiries that fall outside the relevant domain, users should be directed to seek information elsewhere.

As seen above, I received a reply acknowledging my request, and the right-hand side panel already displays an icon placeholder along with a brief description of the bot.

I accept the suggested name. Then, the builder crafts an icon, which I also accept, and then it proceeds to inquire further about the bot’s role and abilities.

I want to avoid the risk of incorrect answers, changes in terminology, or misclassification, so I prompt:

Users will ask questions using specific technical terminology related to the oil and gas sector. The bot will search the provided PDF files and respond by quoting verbatim the relevant sentences or paragraphs that contain the answer to the query. I want to emphasize that the bot is not going to rephrase, paraphrase, or alter the terminology in any way. Answers must be provided exactly as they appear in the PDFs.

Then, I reiterate:

Any query unrelated to the specific topic of reserves classification covered by the PRMS Guidelines will receive a reply directing users to seek information elsewhere.

In response to the builder's query about unclear questions, I emphasize:

If unclear, ask the user for more details until you're certain you can provide an answer by quoting verbatim from the PDF documents, avoiding any misrepresentation or improper use of terminology. I emphasize that the use of precise terminology, exactly as it appears in the PDFs, is critical.

On the tone question that follows, I reply:

Use a formal tone, consistent with the tone in the PDFs themselves.

With that, the basic setup is complete. We now move to the 'Configure' tab at the top:

In the Configure window, we can see the Name and Description that will appear on the user interface (see preview in the right panel) and the Instructions we set up in the previous steps, as summarized by the bot builder.

We also have the opportunity here to add a few conversation starters. You can accept the ones offered by the bot or edit them with your own. I’m going to use the following:

- Outline the foundation principles of PRMS

- Is P90 1P if it's not economic, but P50 is?

- Should fuel gas be counted as reserves?

- What is TRR?

At the bottom, we see the ‘Upload files’ button.

We browse to the location where we have our PDF files to import them, and once they’re uploaded, we're ready to hit the 'Create' button.

When you click the 'Create' button, you’ll be directed to the 'Share GPT' screen, where you need to set your sharing options.

When I built the first version, I made it public and posted it on my LinkedIn feed to invite public scrutiny and feedback. However, the SPE contacted me and requested its removal. I won't delve into the details of those conversations, but I complied and removed it from the store.

In this case, or in any other case where you build a Custom GPT just for yourself, this is where you select the 'Only me' option, and you’re ready to go.

Voilà!

We are done creating our PRMS Advisor.

We click on the 'View GPT' button and can see how the live GPT appears, complete with the icon, name, author, description, and conversation starters:

Looks good! Now a thorough testing phase begins, where you need to ensure the GPT is responding as per the instructions given to it.

You can start by clicking on any of the conversation starters and see what responses you get.

Try to ‘break’ the bot by asking tough questions and ensuring it doesn't miss the mark on fundamental concepts.

You should also take the opportunity to test that it responds appropriately to questions not related to the domain.

Oops! Wrong answer.

When you encounter unexpected behavior, go back to your 'My GPTs' page (https://chatgpt.com/gpts/mine) and select 'Edit' on the Custom GPT you wish to tweak.

I go back to the ‘Create’ tab and add:

I just asked you 'when was World War 2' and you provided an answer. Please refrain from answering any questions that don't relate to the topic of oil and gas resources classification. Any question pertaining to a domain other than oil and gas should be answered with a message like 'Your question does not pertain to Oil and Gas. Please ask a relevant question,' or something along those lines.

I receive a satisfactory response and immediately put the bot to the test within the preview window itself, without having to leave the setup interface.

Here, I try again with a few more non-domain questions to ensure I get the correct responses.

Great!

You’re now equipped with your very own bot, which you can access at any point from anywhere, even on your phone. It can answer direct questions about the PRMS guidelines without you having to sift through those pesky PDF files. Amazing value for just a few minutes of your time!

While I used the PRMS Guidelines as an example here, but the exact same process can be applied to the COGEH Guidelines (as requested by a user), the VALMIN Code, a consultation manual, internal company operating procedures, environmental compliance documents, books—you name it!

There are undoubtedly hundreds of potential CustomGPTs across various domains in our industry. What do you think? Do you have any GPT ideas you’d like to share and collaborate on? Get in touch and let's collaborate!