Skip to main content

Deep Research Mode in a Nutshell

Deep Research Mode enables even more precise searches. Complex problems are broken down into a logical search plan. Each sub-question is researched individually. A built-in quality control system immediately validates the relevance of the documents found and automatically improves the search if information is missing. This gives you more informed, verified, and highly accurate results directly from your company data.

Regular Questions - The Baseline

Some questions are simple. Questions like: “What spare part number corresponds to this item?”. These questions are concise and require only one precise search.

Complex Questions - Where Deep Research excels!

Other questions are more complex. Questions like: “What are the key differences between proposals in the automotive industry and proposals in the public sector in regards to intended project team size and duration?” These questions cannot be answered with a single look-up. They need to be divided into sub-queries and sub-tasks. In this example, finding automotive sector proposals, finding public sector proposals, extracting properties about team size and duration of each, and comparing those properties. Our new Deep Research Mode imitates exactly this human behavior of planning ahead, dissecting into sub-tasks and merging the information back together. To provide the most insightful answer possible.

How to activate Deep Research Mode

You will find the Button to select the Deep Research Mode in the Use Case Hub right below your input field, between the web search and the data upload functionality. Note that the availability of the web search and data upload can vary depending on your Use Case settings. Upon selection, Deep Research is ready to use!
Deep Research Activation
For now, the Deep Research Mode is in its Beta phase. It can be activated on a Use Case level by the Genow team.

How to use Deep Research Mode

Once selected like shown in the section above, simply type in your complex, long and twisty question. Deep Research Mode will figure out a plan to divide, search and consolidate, stopping only when the answer fits your question.

How the Deep Research Mode works

Conventional AI-supported information searches often reach their limits when it comes to complex questions. If inaccurate information is retrieved in an early stage of the search, this error can carry over into the final answer. Our new Deep Research Mode solves this problem by integrating a new architecture. This works like a methodical human analyst and goes through a structured, five-step process:
  1. Search planning: Instead of starting the search immediately, the original question is broken down into a plan of logical sub-questions. The system explicitly records which questions build on each other and in which order they need to be clarified.
  2. Targeted information gathering: The system performs a separate, targeted search for each sub-question. Sub-questions that are independent are researched in parallel. Each sub-search can be performed on company data, i.e. on a Knowledge Assets, or - if enabled in your Use Case settings - as a Web Search.
  3. Integrated quality control (self-reflection): Deep Research mode does not blindly trust search results. After each retrieval, the system independently evaluates whether the documents found are sufficient to answer the respective sub-question factually correctly.
  4. Automatic course correction (refine): If the system determines that the information is insufficient, the search queries are automatically adjusted and an improved search is started. This controlled review loop prevents incorrect or incomplete data from creeping into the further process at an early stage.
  5. Final synthesis: Once all facts have been thoroughly verified, a well-founded, coherent final answer is generated from the collected individual results.
In summary: Through this “plan-first” approach combined with continuous verification, Deep Research Mode achieves a high degree of accuracy in fact-finding for complex questions.

Exemplary Questions

These are examples of questions where Deep Research shows superior results compared to a regular single search. Note how each question combines two or more pieces of information.
  • What year did NVIDIA’s data center revenue (e.g. hyperscalers) surpass that of consumer products?
  • According to the 2000 United States census, what was the year 2000 population of the birth city of the only 21st-century mayor of Austin, Texas who also served as mayor in the 1990s? Round your answer to the nearest thousand.
  • How does Germany’s GDP compare now to when Google was founded?