We’re a curious bunch here at Transform, so naturally we wanted internal use cases of our own to show that the AI craze had legs. But we also love learning; so, we wanted to disseminate the knowledge we'd gain into the organisation so we could build similar solutions for clients going forward. All while making some of our internal tasks quicker in the meantime.
We wanted to play with AI, but to what end? We challenged ourselves to see if we could build a bot that would help us quickly source the right data to design solutions for government customers, amplifying efficiency.
We got to work, data processing all our previous client challenge solutions to ensure they were in the right shape for vectorisation (the process of turning words into numbers).
Once the responses were stored in a vector database, we considered next steps. We were cautious of AI hallucinations (where it would make up false responses) in a process where accuracy would be key.
Because augmented retrieval required far less examples than the thousands of data points you’d need to train a model completely, whilst still allowing us to be really complex with our prompting, this was the method we went with.
Augmented retrieval unlocked another powerful component; we could feed it our corporate style guide and teach it to ‘talk’ in the same way.
In addition to this, we also enabled semantic searching in the back end that would allow the model to 1) answer the prompt by finding 10 - 25 examples based on your search but also 2) create a new example based on those 10 – 25 snippets in your corporate tone of voice.
The Knowledge Bot has since served as a knowledge transfer hub, an indexing tool and has helped our staff craft solutions – all in far less time than it has taken us before. One of our proposed solutions, using the tool, was also shortlisted by the client, showing the value of having AI do the heavy lifting, especially when it comes to outlining solutions that a Transformer can then polish.