OpenAI’s deep research enabled users to go beyond a single question and into complex multi-step tasks using chain-of-thought reasoning. This allowed an LLM to “reason” and utilize more compute, iteratively reusing the output of the last step as the context for the next. With execution broken down into steps, it was also able to fetch additional knowledge through web search as new ideas arose between steps, no longer constrained by having to know everything at the start.
HyperArc’s deep exploration is this jump for analytics — from single text-to-sql natural language queries to enabling complex analytical tasks that require iterative queries to build a foundational knowledge to complete.
Planning
Given a task, deep exploration first finds relevant memories to construct a plan with. This plan become our <thinking> tokens and is provided to later execution steps as the reasoning behind the task.
Unlike reasoning models, HyperArc allows for human in the loop before the final answer. A user is able to modify, remove, and add new steps to augment the plan with their unique business knowledge — such as changing the metric for best assignee from resolution time to CSAT.
Steps
Much like what makes deep research so effective, deep exploration will begin each step by figuring out what additional information is needed and retrieve it from memories. The first handful of memories were helpful to build the plan, but with each new step and query we gain additional knowledge and a better understanding of what other memories would be helpful. These memories inform the execution of each step with the construction of a new query.
Using chain-of-thought reasoning, deep exploration will build out queries step-by-step and is able to automatically fix mistakes, branch when necessary, and most importantly — critique itself with new information with each new query. This self evaluation is what enables it to decide if the plan is sufficient or if changes are needed.
Stories
Rarely is a single visualization enough to tell the whole story and although deep exploration will attempt to build a final visualization, the steps taken to reach that visualization are often just as important. With Narrate, we’ll help build that story for you.
Just shift to multi-select multiple memories to view as small multiples before prompting for a common thread to be extracted and narrated across them. From the deep exploration of “what does the best assignee do differently?”, were able to tell of story that takes the analysis from individual assignees to that of the best vs the rest.
Share this post