News Article Allies work experience student trains ChatGPT on Postcoder
Norwich sixth form student, Joshua Cooper, has developed and demonstrated a proof of concept that enables ChatGPT to accurately answer detailed questions about Postcoder, during a work experience placement at Allies.
ChatGPT is a chatbot developed by OpenAI and built upon its GPT-3.5 and GPT-4 large language models. Since its launch in November 2022, ChatGPT has made it possible for users to ask questions and generate all sorts of written content – everything from detailed essays and complex code to LinkedIn posts and poetry.
It does, however, have limitations. ChatGPT does not always reflect the very latest information, there are gaps in its training data and it can sometimes provide inaccurate responses.
Allies challenged the Year 12 student to find a way to overcome these limitations and to build a ChatGPT-style chatbot that allows users to ask questions about Postcoder.
Joshua said “GPTs provide a text output in response to an input or prompt like ‘Tell me about Postcoder’. By carefully designing a prompt that includes instructions or background information in addition to the user input, you can essentially ‘program’ a GPT model.”
“This approach is called prompt engineering. My code calls the GPT-3.5 model with prompts that contain not just the user’s question but also relevant information about Postcoder, such as its features and their benefits, which I gathered from the website.”
During the work experience, Joshua found that it’s sometimes important for the model to reason about a problem before answering a specific question to avoid rushing to an incorrect conclusion.
“If the user asks, ‘Can Postcoder cuddle cats?’, the desired output is along the lines of ‘No, Postcoder is an API and not capable of cuddling cats’. To increase the likelihood of this response, I added reasoning steps into each prompt, for example, ‘Step 1: Is the user talking about Postcoder? Step 2: If so, is this a Postcoder feature? Step 3: If so, respond with relevant information about Postcoder’.”
Joshua created an OpenAI account to gain access to the GPT-3.5 model and built the chatbot with Python and Flask, and used a Jupyter notebook early on to experiment with structuring the prompts.
Tim Stephenson, Head of DevOps, said “This proof of concept showcases the potential of leveraging AI models to tackle specific business challenges. Joshua’s smart use of prompt engineering shows how large language models can be used to output relevant and up-to-date information which is vital in a business context. He’s done a great job and we are grateful to him for undertaking this research for us.”
Joshua starts his second year of A-Levels in Computer Science, Maths and Further Maths in September, and hopes to embark upon a Computer Science degree in 2024. We hope he enjoys the rest of his summer and wish him well with his future studies.