Working alongside Keepl, the Hartree Centre North East Hub created an end-to-end encrypted chatbot that allows users to converse with their personal documents in a secure way.

Challenge

Keepl, an innovative startup building a digital platform for consumers for secure storage and organisation of their own data, plans to introduce a feature allowing users to converse with their stored documents. The document storage process involves secure storage via encryption, allowing users to control their own personal data, how it is stored and if, when and how it is shared. Keepl wants to see if it is possible to securely store the underlying meaning of documents, which are represented as vector embeddings (lists of numbers). The project is to determine if it is technically possible to store these embeddings in an encrypted format within a vector database whilst still being useful for a Large Language Model (LLM) to query it, at an acceptable speed for a commercial consumer-focused product.

‍

Approach

Data Scientists from the Hartree Centre North East Hub worked with Keepl to explore the production of a proof-of-concept encryption, decryption, and end-to-end Retrieval Augmented Generation (RAG) model for their personal data. The initial stages involved understanding and exploring the literature and methods involved in RAG, Question-Answering chatbot best practices, LLM’s and Encryption schemes. The team undertook feasibility experiments of the minimal code for the processes involved in the steps of RAG QA and Encryption. Then, with standard encryption approaches applied, benchmarking code was produced, allowing measurement of the time taken for steps of the process when scaled, which was a crucial goal of the project. With a minimal viable output produced, the team then explored useful advanced encryption techniques such as Homomorphic Encryption applied to the operations involved in RAG. Once this reached a natural conclusion, the RAG model was padded out, with modular code enabling systematic configuration of the parameters, evaluation and assessment of the output. The project ended with the production of a personal document chatbot. Keepl will now explore advancing the Encryption methods worked on during this project.

‍

‍

Benefits

The initial literature exploration and code produced by the Hartree Centre North East Hub will enable Keepl to demonstrate the ability to encrypt, decrypt and ask questions of your personal documents. The Keepl team have a deeper understanding of the advanced encryption methods’ capability and potential for further development. They have also received a valuable framework to build up a production RAG QA system as a feature for their users, with models to be hosted on-premises or on Azure cloud platform. As a result of the project, Keepl is exploring further work through the Arrow Programme and potentially creating a funded PhD project on applied encryption.

‍

"The Hartree Centre North East Hub has provided a useful framework and knowledge base to enable us to move towards our vision of a fully secure and useful personal document chatbot."

- Chris Campbell, Keepl

‍

"“Through the sessions we have seen the direction in which encryption is headed, which can be used to complement and increase user retention, engagement and attract new users to our platform. This will allow us to develop our platform offering with an exciting and novel feature.”

- Karl Hewitson, Keepl

‍

Further Information

This work was completed as part of one of our collaborative data projects. The projects are up to 12 weeks in duration and give you access to a wide range of expertise across our team of data scientists and data engineers. We will work alongside your team to scope your data science or engineering project, build a prototype solution, and explore options to deploy it within your organisation. You can learn more about them on our webpage here.

If you would like to learn more about the Hartree Centre North East Hub or our collaborative data projects, please get in touch with us at: hello@hartreenortheast.uk

‍