Hierarchical Summaries

Hierarchical Summaries provide multi-level abstraction over the CueCrux corpus using the RAPTOR approach, enabling retrieval at any granularity from specific passages to broad thematic overviews. This capability addresses the fundamental tension between precision and breadth in information retrieval, allowing the platform to serve both narrow fact-finding queries and wide exploratory questions from the same knowledge base.

The hierarchical structure is built as a tree. At the leaf level are the original semantic chunks produced by the chunking pipeline. These leaves are then clustered by topical similarity, and a language model generates a summary for each cluster. These cluster summaries become intermediate nodes in the tree, representing broader topics than any individual chunk. The clustering and summarisation process repeats at progressively higher levels, creating additional layers of abstraction until the top level captures the broadest themes across entire domains.

When Engine processes a query, it can retrieve evidence from any level of the hierarchy. A specific factual query like the maximum permissible exposure limit for a particular chemical will be best served by leaf-level chunks that contain the precise data point. A broader query like the key themes in recent food safety regulation will be better served by intermediate summaries that synthesise across multiple documents. The retrieval system dynamically selects the appropriate abstraction level based on the query's specificity and scope.

This multi-level retrieval has significant benefits for answer quality. Broad queries that would retrieve dozens of marginally relevant individual chunks instead retrieve a handful of well-crafted summaries that cover the same ground more coherently. Specific queries bypass the summaries entirely and retrieve the original chunks with their precise factual content. The result is answers that are appropriately detailed for the question being asked.

Hierarchical Summaries also improve the efficiency of the retrieval pipeline. Rather than processing hundreds of candidate chunks for a broad query, the system can work with a much smaller set of summary nodes that collectively represent the same knowledge. This reduces computational cost and latency while maintaining or improving answer quality for broad queries.

Each summary node in the hierarchy maintains provenance links to the chunks it summarises, and these links are recorded in CROWN receipts. When an answer cites a hierarchical summary, the receipt includes the full chain from the summary down to the original source chunks, enabling verification at any level of detail.

For the CueCrux platform, Hierarchical Summaries make the corpus equally effective for detailed professional research and high-level strategic overviews, serving the full spectrum of user needs from a single knowledge infrastructure.