I have spent a lot of time this week working on my latest MyGPT creation, which is a customised historical query-bot that has been fed a decade’s worth of Bradford-specific information from the GPO circulars. Some of this work may soon be superseded by Nayomi’s RAG work, which looks very promising, but I think it is still useful to see what can be done with GPT given it is already so user friendly.

The information I have used to create the MyGPT knowledge base is all directly from the circulars themselves. I used ABBYY to turn the PDFs into searchable PDFs, and then ran a search for Bradford. I then copy and pasted entire pages where Bradford was mentioned, editing the OCR reading as rarely as possible. Despite aiming for minimal intervention, I did have to edit a lot of the tables in the circulars as their awkward formatting was not read well by ABBYY.

I then began probing the MyGPT for answers to questions around Bradford’s communication links in the period (1880 – 1889). Rather geographically challenged maps courtesy of the DALL-E extension are below.

The GPT reported that “the mention of telegram rates to places like Egypt, Tasmania, and New Zealand in the 1886 circular shows Bradford's participation in a global communication network, essential for international commerce and correspondence.” To fact-check this, I went back to the full circular myself to search for references to communication links between Bradford and Egypt.

While there are no explicit links made between Bradford and Egypt, the circulars do show the cost to send a telegram between London and Egypt, as well as list several telegraph offices in Egypt. We can confidently assume London and Bradford were connected, and so that Bradford could get a message to Egypt via London – however, is this level of ‘assumption’ something we are comfortable with a LLM doing?

The MyGPT itself admitted it wasn’t totally sure if Bradford could reach Egypt, but only when prompted to reveal how certain it was: “The information in the General Post Office Circulars indeed indicates that Bradford was part of a telegraphic network that allowed the sending of telegrams to various international destinations, including Egypt… However, it's important to note that the circulars primarily provide guidelines, rates, and instructions related to telegraphic services and may not explicitly confirm the sending of a telegram from Bradford to Egypt… The presence of a tariff for Egypt and other countries suggests the capability and the established route for such communication during the period the circulars cover.”

So, on the one hand, the MyGPT is disguising itself as a historian quite well and making decent historical hypothesis. On the other, it is disguising its guesses as fact, and I am sure we would need to feed it more knowledge to be comfortable with the historical conclusions it is drawing – this knowledge would include secondary literature (eg. Connecting the Nineteenth-Century World, Wenzlhuemer) and further primary data such as submarine cable maps (BT Archives).

Following a conversation with Tim, and moving away from training GPT to take over my job for me, I started thinking about how the circulars could be linked to physical objects. An easy first step was to ask GPT if it understood its knowledge base contained technologies, so I prompted it with “Could you write a list of every object or technology which is referenced in the dataset you have?”

The output is below:

In the spirit of linkage, then, and given the MyGPT had provided 17 options of technologies (in the loosest sense of the word) within its very specialist slice of the dataset, I fed a catalogue to the MyGPT and asked it “can you pull out objects from this catalogue that link to the technologies you have noted are relevant, bearing in mind the dates and location of the original circulars information, so that you are only pulling out objects that directly link?”

The catalogue used for this experiment was a GPT generated one, which is available here for anyone who wishes to see what GPT thinks a museum holds. MyGPT’s answer to the question posed to it is below:

Although the catalogue is completely imaginary, I think this approach shows some promise. If I can train a GPT using my own dataset, and potentially also feed it a full catalogue in the training stage, it appears that a customised MyGPT would be able to identify links between historical sources and contemporary museum catalogues. All of this can be done by anyone with the required source material, a $20 ChatGPT subscription, and fairly minimal technical knowledge.

Beyond the circulars, I have been continuing to work on the personal researcher notes project which I hope to share reflections on next week. Tim, Daniel and myself are also currently calling out for papers to present at the BSHS conference in the summer, so do get in touch if you are interested!