How to Effectively Search with Regulatory Agents?

During a recent webinar, an analyst raised a question regarding hierarchical indices for regulatory documents, specifically the structure of 'document → article → section.' This model allows for a descent from the root, akin to how one navigates a table of contents. However, the discussion quickly turned to the costs associated with indexing, which involves traversing the entire document corpus. I likened this indexing method to a B-tree structure. After some calculations, the comparison proved valid. In a B-tree, nodes are optimized for disk blocks to minimize head movement. In this context, a block represents an efficient window into a smaller model, while head movement refers to accessing that model. For instance, with six acts spanning approximately 200 pages, the depth is two, and with 2 million tokens, the depth increases to three. The navigator can read between 8,000 to 12,000 tokens instead of the entire corpus. The hierarchical structure is already present in legal texts, including chapters, articles, parts, and sections, constructed using regular expressions based on numbering without needing a model. The navigator's task is to select one of fifty branches based on a brief description, leaning more towards classification than reasoning. There was a question in the room about whether a model with 1.5 to 3 billion parameters could handle this task. The intrigue continues as we explore methods for obtaining precise answers from legally significant documentation.