AI enables telecoms research toolkit
The UK Telecoms Innovation Network is developing a toolkit that uses a combination of AI techniques to help researchers, investor and companies explore telecom research projects.
The AI discovery toolkit is being developed by the Smart Internet Lab at the University of Bristol for UKTIN with a new way to search telecom research data.
The toolkit, currently in prototype, uses a combination of a generic large language model, in this case Google Gemini, and retrieval augmented generation (RAG) to explore a ‘data lake’ of projects. The data comes from open sources such as the UK and EU research agencies, and is hosted by the Lab in a ‘data lake’.
The interface of the tool is an interactive chatbot but also shows the underlying data with references to provide transparency.
“Back in 2020 the whole issue of supply chain diversification with anxiety about security in the telecom infrastructure coming from high risk vendors in our supply chain, particularly Huawei,” said Prof Dimitira Simeonidou, UKTIN Lead for UK Research Capability and head of the Lab.
“Looking across the telecoms sector it is hugely fragmented, with universities, large operators such as BT and startups but not large vendors. One of the large things OpenRAN and Open Networking initiatives to create an active market in the UK to commercialise innovation assets that we have and the idea was that to motivate the market we need to mitigate the fragmentation. So we need a front door for users for companies to find out what is going in the UK and make it easy to navigate,” she said.
“We still believe in this, and this tool is the front door.”
“The Discovery Toolkit unlocks the potential of our national telecoms R&D capabilities in a way that has never been seen before – the technology is fast and sophisticated. AI-driven semantic search and large language models enhance data discoverability, allowing users to easily define their interests and access the latest information. The service it provides is slick – its interactive data interpretation feature, enables users to “chat” with the data and transform raw information into actionable insights,” she said.
The design of the tool has evolved to include large language models (LLM) such as GPT, Gemini and Llama, but without training them with data. Instead theses are used for the natural language processing, coupled with semantic structures called embeddings that provide more natural communication. The data lake approach also means that the data is always up to date and can be used to avoid the risk of the LLM hallucinating and inventing data. The results can also be exported as a CSV file.
“The access to the live sources puts us ahead,” said Simon Bond, director of Bristol Innovations at the University.
UKTIN sets up 5G R&D special interest groups, looks for chairs
“We started with people, but this doesn’t scale so the plan was to create a database service, a data lake, for people to discover things,” said Simeonidou. “But that’s only as good as the data you import and when you do it is immediately out of date. We knew about large language models and early on we realised the potential of LLMs to access databases in real time and a natural language interface.”
“The technology at that point was extremely early and the proposal was met with scepticism, so we started with the data lake and explored LLMs as search tools and it became clear that this was a valid way to go forwards.”
The cloud-based Google Gemini LLM was chosen as this gives the highest number of tokens to be used, so the 2000 token inputs cloud can allow the details of 250 projects to be accessed. This is hosted by Google.
The architecture allows for other LLMs to be used, for example for a less detailed query, or to use another LLM to check the results of the first one.
The group is now looking to extend the tool to search the written elements of patents from the UK, EU and possibly the US. The tools could also be extended to other areas such as wider semiconductor research.
The prototype AI discovery toolkit is at discover.uktin.net/ but requires a UKTIN registration.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
