DFS-RAG

Since OpenAI O1 came out I have been seeing a lot of discussion around Chain of Thought + Reinforcement Learning fine-tune, actor-critic models, monte-carlo tree search and so on.

But for people who are working with structured content like code generation or legal circulars, you can hotfix a lot of this with DFS RAG.

DFS RAG follows the same process as you see in your OpenAI O1 model's chain of thought but because all of these are distinct model calls, the flow is much more parallelizable in production. And is also more constrainted when being using temperature > 0.2 and top-p < 0.95, which may be required when the model must use some freedom.

Here is a flow below for code generation style of usecases:

As you can see there are 2 databases, here you can prepare them by extracting, annotating, and linking type definitions of custom structs. In case you have nested custom types in structs you need to flatten them out when generating context.

We have adapted this a similar depth-first flow to work with compliance documents at OnFinance AI at 1/20th the cost of OpenAI O1 using open source models. This makes sure you are able to track version control from regulator documents without needing any manual intervention, no matter how complex the gazette/regulation.

Email:[email protected] to get a demo of our compliance copilot.