Hype aside – the usefulness of AI integrations for natural language assistance and Retrieval-Augmented Generation (RAG) knowledge retrieval is impressive, and improving fast.
We are pleased to have produced our own limited proof of concept chat bot application. We chose the UK Tax sector as our subject due to the availability of authoritative information, and the challenges of application (e.g. calculations / exceptions / applicability). We deployed a test chat bot initially trialing ‘GPT 4o’, LLAMA3.3, and the recently announced DeepSeek (R1), (and others) as the underlying LLM. We seeded it with carefully selected sources (both long form tax PDF info guides and HMRC site page URLs to crawl). Results were impressive and we can see its potential, and its application in other sectors too.
The caveat here as ever, is careful attention to content safeguards, good prompt design and reference pre-selection; through test, training and filtering we endeavour to ensure that results are pertinent, accurate, legal and truthful – and importantly avoiding any ‘hallucination’.
The sheer speed with which these tools can now be seeded, trained and deployed in useful applications is impressive (safeguards notwithstanding).
This now presents fascinating and productive back ground R&D work for us, and already we can see its usefulness for immediate application in all sorts of businesses.
It also felt good to produce something potentially commercially useful, rather than just playing with the production of mad images (although that’s also fun).
