September 30, 2025

The latest release of our RheinInsights Retrieval Suite offers algorithms for deep research. Here, we use the combination of large language models and vector search to iteratively generate the best results for the user's request. But why does this approach work so well?

Consistent Data in the Search Index

Our connectors use a uniform index schema. This means that the requirements of the index schema are taken into account directly when implementing the content source connectors. In other words, the fields we develop the connectors directly to provide consistent data - across all content sources.

For example, the title field for Jira issues is filled with the issue’s subject, for SharePoint documents it is the extracted title of the document, for web pages it is the title tag, or as a fallback the file name. The same applies to the author field, the modification date, and much more. Furthermore, the file types are homogenized and well-defined.

Deep Research

On this basis, the AI ​​can then easily make the right decisions when searching for answers.

For example, the AI ​​can decide in advance which file types are suitable for the search query and which are not. For example, contracts are usually stored as PDFs, as Word documents, or Word-like documents. But (legal binding) contractual information usually cannot be found in source code files or on web pages. Documentation, however, lives in wiki pages, web pages, or similar.

This allows that the AI can determine which filters to apply and to increase the vector search’s precision. At the same time, we use mechanisms to expand the query to increase the search’s recall.

In combination, the AI ​​can furthermore decide whether the first wave of documents already answers the search query or whether the search needs to be refined to gather further context and answer the query in a second attempt.

Either way, the consistent data by our connectors is the foundation of enabling the AI to produce the best answers.