HIRO: Hierarchical Information Retrieval Optimization

TL;DR

HIRO is introduced, a novel querying approach that employs a Depth-First Search (DFS)-based recursive similarity score calculation and branch pruning that uniquely minimizes the context delivered to the LLM without informational loss, effectively managing the challenge of excessive data.

Abstract

Retrieval-Augmented Generation (RAG) has revolutionized natural language processing by dynamically integrating external knowledge into Large Language Models (LLMs), addressing their limitation of static training datasets. Recent implementations of RAG leverage hierarchical data structures, which organize documents at various levels of summarization and information density. This complexity, however, can cause LLMs to "choke" on information overload, necessitating more sophisticated querying mechanisms. In this context, we introduce Hierarchical Information Retrieval Optimization (HIRO), a novel querying approach that employs a Depth-First Search (DFS)-based recursive similarity score calculation and branch pruning. This method uniquely minimizes the context delivered to the LLM without informational loss, effectively managing the challenge of excessive data. HIRO's refined approach is validated by a 10.85% improvement in performance on the NarrativeQA dataset.

Figures (6)

• Figure 1: Architectural Overview of a Retrieval-Augmented Language Model (RALM). A query is inputted, processed through an embedding model to generate a vector representation, which is then matched in a vector database to retrieve relevant contexts. These contexts are fed, along with the original query, into the LLM, resulting in an informed response.
• Figure 2: HIRO Querying Mechanism. HIRO's recursive process employs Selection and Delta Thresholds to filter and prune document graphs for optimized information retrieval. It highlights how relevant contexts are dynamically identified and retrieved based on a balance of specificity and information breadth, tailored to each unique query.
• Figure 3: Confusion Matrices for RAPTOR with (a) HIRO Querying and (b) Collapsed Tree Querying on the QuALITY dataset
• Figure 4: Comparison of Time Complexities for the Three Hierarchical Retrieval Algorithms.
• Figure 5: Tree Traversal Querying Mechanism. The top-$k$ root nodes are selected based on their similarity to the query. In the next layer, the children of these nodes are considered, and the top-$k$ nodes are again selected based on similarity. This process continues until reaching the leaf nodes.