This week, we’ve invited Product Manager Seb May and Head of Software Engineering Toby Stephens to discuss Delfy, a BC start-up tech company developing a cutting edge AI-tool of the same name to empower businesses.
Tell us what Delfy is and why was it created?
Seb: It seems that generative AI has been the world's focus over the last year or so. However, Delfy became a thing when we realised that there were some key problems and blockers associated with the large language models (LLMs) produced by the tech's big players. Most prominently, LLMs hallucinate and have restrictive token limits, which means they can't be used for high value and intricately detailed tasks due reliability/ accuracy faults. This is where Delfy comes in. It’s an API-based tool which allow users to harness the power of GenAI to automate high value, specific tasks, without sacrificing the accuracy and usability of their content. We do this by allowing users to create their own corpuses of validated and/ or approved information that the user can then interact with and prompt as they would with an LLM.
What is your vision of Delfy? How do you expect for it to be used?
Seb: We have seen people use it in a number of ways, however typically, Delfy is used in two high level avenues: 1) Content marketing, and 2) Interactive knowledge base.
Our best validation has come in our work with Contentive, a specialist media provider. Their business model relies on producing vertical specific content in finance, accounting and HR - three major thought leadership channels. Delfy is now the foundational tool used to produce content for Contentive, allowing one editor to generate 3 newsletters, 8 articles and 8 article-specific social media posts across all three brands… per day. Their engagement benchmarks have increased by between 1-2%. Previously, a team of editors would have been expected to generate this volume of final version content over the span of several days. We hope, and expect, that Delfy can be used as a foundational piece of software at enterprise level in a manner that replicates what we have done with Contentive.
How does Delfy work? Could you tell us a little bit more about the underlying AI technology?
Toby: Delfy brings Retrieval-Augmented Generation (RAG) to organisations without the need for any software development at their end. As Seb mentioned, organisations can then create corpuses of trusted content that is used to answer questions and generate new content. When a new document is added to a corpus, it is processed by Delfy in a number of ways including by keyword, semantic embedding and topic classification. These characteristics are stored together with the document in a powerful search engine called Elasticsearch.
Documents can be added to a corpus via API, sent via Email, or entered directly into the Delfy web app. There is also Zapier and Power Automate integrations that makes use of the API to receive documents from these sources.
Organisations can then create workflows which select the most relevant information from the organisation's corpuses and generate content from that information. The LLM never uses its own knowledge, only the knowledge provided by Delfy that has come from the organisation's own corpuses.
Selecting the most relevant information is done using a complex Elasticsearch query that scores relevance for each characteristic of each document in the corpuses the workflow uses. The score can also be adjusted with things like date decay, which means that the most recent documents are considered more relevant than older documents.
Workflows are scheduled to run periodically and the output of each workflow can be either sent to a number of email addresses or can be sent directly as a JSON payload to an organisation's own backend software. This way, the organisation can take the workflow output and put it straight into a Wordpress Blog or a Newsletter Template without any human involvement.
Where do you see Delfy going next?
Seb: We're still figuring out exactly where to expand Delfy's product offering as the platform is a bit of a Swiss army knife - it isn't designed to serve only one distinct customer profile. Our product roadmap is aimed at developing our existing capabilities, enhancing the accuracy and quality of our output, in conjuncture with adding a few more niche offerings. For example, we intend to build some meta-tagging functionalities as well as operations to enable users to compare new document versions with previous versions of that document. This is so we can work towards our key objective this quarter which to start serving external clients, at both SME and enterprise level.
