Topo-RAG: Topology-aware retrieval for hybrid text-table documents

TL;DR

Topo-RAG challenges the default assumption that everything in enterprise documents can be treated as text by introducing topology-aware routing that splits narrative and structured data into two specialized retrieval paths. Route A uses a dense narrative Bi-Encoder for text, while Route B employs Cell-Aware Late Interaction (CALI) to preserve table topology via per-cell vectors and a MaxSim scoring scheme. On the SEC-25 synthetic corpus, the dual-path approach yields an $18.4\%$ improvement in $nDCG@10$ for hybrid queries and robust performance as table width grows, thanks to per-cell interactions and efficiency techniques like WARP and CRISP. The work has practical industry impact, advocating composite indexes and graph-inspired integrations to bridge the gap between unstructured text and structured data for accurate, scalable enterprise search.

Abstract

In enterprise datasets, documents are rarely pure. They are not just text, nor just numbers; they are a complex amalgam of narrative and structure. Current Retrieval-Augmented Generation (RAG) systems have attempted to address this complexity with a blunt tool: linearization. We convert rich, multidimensional tables into simple Markdown-style text strings, hoping that an embedding model will capture the geometry of a spreadsheet in a single vector. But it has already been shown that this is mathematically insufficient. This work presents Topo-RAG, a framework that challenges the assumption that "everything is text". We propose a dual architecture that respects the topology of the data: we route fluid narrative through traditional dense retrievers, while tabular structures are processed by a Cell-Aware Late Interaction mechanism, preserving their spatial relationships. Evaluated on SEC-25, a synthetic enterprise corpus that mimics real-world complexity, Topo-RAG demonstrates an 18.4% improvement in nDCG@10 on hybrid queries compared to standard linearization approaches. It's not just about searching better; it's about understanding the shape of information.

Figures (5)

• Figure 1: The linearization bottleneck versus Topo-RAG. (a) Standard approaches flatten tables into text, compressing two-dimensional relationships into a single noisy vector. (b) Topo-RAG preserves the topological grid: each cell becomes an independent embedding, allowing the query to interact precisely with the relevant values (e.g., matching “B” and “20”) via Late Interaction.
• Figure 2: The Topo-RAG architecture. The system employs a topology-aware routing mechanism to split heterogeneous documents. Narrative text follows a standard dense retrieval route (top, blue), while tabular data is processed via a cell-aware Late Interaction path (bottom, orange), using the WARP engine for greater efficiency. Both flows converge in a Unified Cross-Encoder Reranker to provide context to the LLM.
• Figure 3: Cell-Aware Late Interaction (CALI). Unlike dense retrieval, which compares one vector against another, Topo-RAG compares each token vector of the query ($q_i$) with all cell vectors ($d_j$) of the table. The MaxSim operator (orange arrows) independently identifies the best-matching cell for each term (for example, “Price” matches the header, “Verna” matches the row identifier), regardless of their distance in the linearized text. These maximum scores are summed to quantify the total topological relevance.
• Figure 4: Retrieval performance by query type. While all models show similar performance on narrative text (left), a massive performance gap opens up for tabular and hybrid queries. Topo-RAG (orange) maintains high accuracy in complex scenarios where linear approaches (gray/blue) collapse due to loss of structure.
• Figure 5: Robustness to information density. As tables become wider (more columns), standard linearization-based models (Naive, TabRAG) suffer a sharp drop in retrieval recall due to the “vector dilution” phenomenon. Topo-RAG maintains an almost constant performance, demonstrating that Cell-Aware Late Interaction effectively decouples the information capacity from the fixed dimensions of the vector.