From RAGs to Riches

Generative AI (GenAI) is evolving so fast that it feels like a moving target - but retrieval augmented generation (RAG) is here to stay. The seminal paper on RAG was first submitted only 4 1/2 years ago, and now RAG is the primary method for building applications that include large language models (LLMs). RAG solves the problems caused by large language models' reliance on dated, static training data, which often results in outdated information, hallucinated responses, and an inability to leverage newer external data and/or an organization’s proprietary data.    

RAG is a technique that leverages the functionality of large language models with relevant information from internal and external sources that was not used to pre-train models. Applications using RAG leverage information beyond what is in the LLM, such as a company’s internal databases or specialized knowledge repositories before generating responses. This process helps improve the accuracy and relevance of the output by retrieving and incorporating the most current and context-specific data available, and reduces hallucinations and inaccurate responses by grounding the model to use only data selected by the application developer to generate more contextually relevant responses.

If you're not familiar with retrieval augmented generation and would like to learn more, check out our RAG resource page. 
  

InterSystems released support for RAG about a year ago, and was one of the first data management software vendors to ship vector search capabilities embedded in an integrated data platform. With InterSystems IRIS, vector embeddings and vector search capabilities are built into the platform, which is different than, and often superior to a standalone vector database. This means that GenAI and vector search capabilities can augment traditional analytics like BI and machine learning to generate results that can be more accurate than a single analytical technique alone. For example, consider an analysis to identify customers at risk of churn. Traditional machine learning can provide good results by executing a predictive model created via logistic regression, decision trees, GBM, or other. Similarly, RAG can provide good results by feeding the LLM with all of the relevant information on customers, including not only structured data but also reams of unstructured data including free text notes in CRM and call center systems. The benefit of a combined approach that leverages both machine learning and GenAI is that it will typically produce more generated responses compared with either approach alone.    

We recently presented our GenAI and RAG capabilities at the North East Database Day (NEDB) conference (check out this link to my slides and poster from the event) and I had some insightful conversations with top researchers in the field regarding our approach. This was tons of fun and also personally gratifying to hear from these experts that our approach is innovative and leading edge. 

We are seeing strong and rapidly growing adoption of RAG from our customers that are using our InterSystems IRIS data platform with fully integrated vector search capabilities. At our 2024 user group conference, InterSystems Global Summit 2024, (less than 3 months after our first release of our vector capabilities) we already had 17 customers showing off their RAG and vector search implementations, and it's only grown since then.  

If you want to play with a cool application that uses our InterSystems IRIS RAG capabilities, you can visit the InterSystems Developer Community "Ask DC AI" page where you can ask natural language questions to get answers from the vast set of content in our developer community. I use this implementation of InterSystems IRIS and RAG often for programming tips to receive code snippets along with clear explanations – a huge timesaver for me! 

From the initial development of RAG (now called "naïve RAG") many variants have evolved, including RAFT, FLARE, REACT, GraphRAG, Agentic RAG, and others. RAG is a fairly straightforward concept, but making it successful can sometimes be complex, and new techniques are constantly evolving. For those not into reading academic papers, check out a recent Gartner® report (note that you must be a Gartner client to read the report, sorry!) that starts with "Creating highly accurate retrieval-augmented generation systems on business use cases is a challenge due to the many factors that contribute to overall performance." The highlighted boxes on the diagram below from the report are all key areas for optimization. 

  

The good news is that we already include all of these capabilities within our InterSystems IRIS data platform. Our philosophy of integrating capabilities in a single platform and executing the processing on the data (rather than copying the data to a separate analytical engine or vector database) turns out to be very well suited for GenAI, from orchestrating LLMs with production-grade controls to managing RAG with the sophistication needed for high accuracy in complex applications. Developers can start using basic RAG and quickly add the optimizations they need for advanced production use. And since InterSystems is staying on the forefront of this technology as it evolves, customers can rest assured that they'll have best-of-class scale, performance, and ease of use. The robustness and industrial grade reliability of InterSystems IRIS is critical for customers as their projects move to production.  

2025 will be a breakout year for our customers using GenAI. Many projects that started as pilots in 2024 are now moving to production, and many more of our customers are starting to use RAG. We're also using it heavily ourselves in our own products and operations, and in the process building real-world experience and best practices.  

Helping customers leverage their data to create new insights and value-added services is core to InterSystems culture, and GenAI is opening up a whole new, exciting world of opportunities and possibilities. 

That's how you get “From RAGs to Riches."