💬 NLP

FT-RAG: A Fine-grained Retrieval-Augmented Generation Framework for Complex Table Reasoning

This paper introduces FT-RAG, a fine-grained retrieval-augmented generation framework that decomposes tables into entry-level semantic units and employs structural neighbor expansion to significantly outperform existing baselines in complex table reasoning, supported by the newly proposed Multi-Table-RAG-Lib benchmark.

Original authors: Zebin Guo, Weidong Geng, Ruichen Mao

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zebin Guo, Weidong Geng, Ruichen Mao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a complex case. Your evidence isn't just a stack of loose notes; it's a massive filing cabinet filled with thousands of spreadsheets (tables) mixed in with long reports (text).

Current AI detectives (Large Language Models) are great at reading the long reports, but when they hit the spreadsheets, they tend to get confused. They often treat a whole spreadsheet like a single, messy paragraph, or they try to flatten it into a list. This causes them to miss the tiny, crucial details hidden in specific cells, like a specific number in a specific row and column. They might find the right topic but grab the wrong number.

The paper introduces FT-RAG (Fine-grained Table-aware Retrieval-Augmented Generation), a new system designed to fix this. Here is how it works, broken down into simple concepts:

1. The Problem: The "Whole Pie" vs. The "Slice"

Think of a traditional AI trying to read a table like someone trying to eat a whole pizza in one bite. They swallow the whole thing (the whole table) without tasting the specific toppings (the specific data points). If you ask, "What was the profit in March?" and the AI just grabs the whole "Financial Report" page, it has to guess which number is the profit and which is the cost. It's messy and prone to mistakes.

2. The Solution: FT-RAG's "Lego" Approach

FT-RAG changes the game by breaking the table down into its smallest, most meaningful pieces before the AI even looks at it.

Step 1: The Micro-Scanning (Entry-level Decomposition)
Instead of seeing a table as one big block, FT-RAG breaks it down into individual "Cell Groups." Imagine taking every single square in a spreadsheet and wrapping it in a protective bubble. Inside that bubble, the AI doesn't just see the number "500"; it sees "500," plus the column header "Revenue," the row header "Q1," and the document title "2024 Annual Report." It knows exactly where that number lives.
Step 2: Building the "SAT" Map (Structured Graph)
Once the data is broken into these tiny bubbles, FT-RAG builds a giant, organized map called a SAT Graph.
- S (Subject): Who or what is this about? (e.g., "Company A")
- A (Attribute): What are we measuring? (e.g., "Revenue")
- T (Temporal): When did this happen? (e.g., "2024")
Think of this map like a subway system. Instead of wandering aimlessly through a city (the whole document), the AI hops on a train that goes directly from "Company A" to "Revenue" to "2024." It connects the dots logically, ensuring that the number it finds is actually linked to the right time and the right company.
Step 3: The "Neighbor" Check (Structural Neighbor Expansion)
Sometimes, the answer isn't just one number; it's a trend. If you ask, "How did revenue grow?", the AI needs to see the number for 2023 and 2024. FT-RAG has a special feature that automatically checks the "neighbors" on the map. If it finds the 2024 data, it instantly pulls the 2023 data next to it, just like a detective checking the alibi of a suspect's neighbor to get the full picture.
Step 4: Mixing the Ingredients (Multi-Modal Fusion)
Tables are often dry and lack context. A number might mean "Profit" in one report and "Loss" in another depending on the surrounding text. FT-RAG takes the precise number it found on the map and grabs the nearby sentences from the report to explain why that number matters. It fuses the hard data (the table) with the story (the text) so the AI understands the full context.

3. The New Training Ground: Multi-Table-RAG-Lib

The authors realized that to test this new detective, they needed a harder test than what existed before. Most tests only asked simple questions about one single table (like "What is the price of item X?").

They built a new, massive library called Multi-Table-RAG-Lib.

The Challenge: This library contains nearly 10,000 questions that require the AI to jump between multiple different tables and mix that data with text.
The Goal: It forces the AI to be a master of integration, not just a simple search engine.

4. The Results: A Clear Victory

When they put FT-RAG to the test against other top AI systems:

Accuracy: It found the right specific data points (the "cells") much more often than anyone else. It improved the ability to find the exact right cell by nearly 60%.
Truthfulness: When the AI generated an answer, it was much more likely to get the actual numbers right (a 62% improvement in exact value accuracy).
Consistency: It didn't just get lucky; it consistently outperformed the competition across all types of difficult, multi-table questions.

Summary

In short, FT-RAG is like upgrading a detective from someone who reads a whole file folder hoping to find a clue, to a forensic expert who has a high-tech map of every single piece of evidence, knows exactly how they connect, and can instantly pull in the surrounding story to make sense of the numbers. It stops the AI from hallucinating or guessing and forces it to ground its answers in the precise, structured reality of the data.

Technical Summary: FT-RAG for Complex Table Reasoning

Problem Statement

Retrieval-Augmented Generation (RAG) systems have proven effective for unstructured text but struggle significantly with structured tabular data. Conventional RAG approaches typically treat tables as monolithic units or flatten them into plain text, resulting in coarse retrieval granularity and a loss of the precise, cell-level semantic relationships required for high-accuracy reasoning. This simplification obscures the spatial and structural logic inherent in tables, making it difficult to perform rigorous cross-table information fusion or complex multi-hop reasoning. Furthermore, existing benchmarks primarily focus on single-table retrieval with simple factual answers, failing to evaluate the capability of systems to integrate fragmented data from heterogeneous documents (text and multiple tables) to generate complex responses.

Methodology: FT-RAG Framework

To address these limitations, the authors propose FT-RAG (Fine-grained Table-aware Retrieval-Augmented Generation), a framework designed to process tabular data at the entry level. The methodology consists of three core modules:

1. Fine-Grained Table Parsing

Instead of processing tables as blocks, FT-RAG decomposes raw tables into Cell Groups. Each cell group encapsulates the raw cell value along with comprehensive contextual metadata, including:

Global metadata: Document titles and captions.
Table-level metadata: Schema and headers.
Hierarchical paths: Explicit encoding of indices (absolute coordinates), types (row/column header roles), and hierarchy depth (nesting levels).

2. Structure-Aware Semantic Lifting (SAT Graph)

The framework constructs a unified Subject-Attribute-Temporal (SAT) Graph, a composite directed acyclic graph ( $G = (V, E)$ ) that organizes cell groups into interconnected knowledge units. This graph is built upon three dimensions:

Subject Sub-tree ( $T_S$ ): Generated via LLM inference to establish structural lineage and entity hierarchies (e.g., "Laptop" under "Hardware").
Temporal Sub-tree ( $T_T$ ): Organizes time intervals chronologically using pattern-based parsing, ensuring temporal validity (e.g., specific dates nested within years).
Attribute Node Set ( $V_A$ ): Acts as semantic convergence points, bridging disparate subject-temporal contexts.
Value Leaf Node Set ( $V_L$ ): Stores atomic data entries as key-value pairs, anchored to specific SAT combinations via a composite index ( $\langle s_i, t_j, a_k \rangle$ ).

3. Graph Retrieval and Multi-Modal Fusion

Retrieval occurs in two phases to ensure both structural precision and semantic completeness:

Phase I: Hierarchical Graph Navigation: The system employs a Dual-Path Traversal Strategy (Anchor-Driven Forward and Attribute-Driven Reverse) to align user queries with the graph's dimensions. It utilizes a Structural Neighbor Expansion mechanism to retrieve semantically connected entities (e.g., adjacent time nodes for growth rate calculations) without relying solely on vector similarity, which can lead to factual drift.
Phase II: Text-Bridged Augmentation: Retrieved structured facts are linearized into natural language templates and used to query the associated textual corpus. This fuses precise tabular metrics with surrounding narrative context (qualitative insights, caveats), providing the LLM with a holistic evidence package.

Key Contributions

FT-RAG Framework: A novel fine-grained approach that decomposes tables into semantic units and organizes them into a schema-aware SAT Graph, enabling precise cell-level retrieval and reasoning.
Multi-Table-RAG-Lib Benchmark: A new dataset comprising 9,870 QA pairs derived from 308 large tables across 81 documents. Unlike existing benchmarks, it specifically targets complex, multi-hop reasoning requiring the integration of multiple tables and text-table fusion. It includes both "Pure Tabular" and "Context-Augmented" retrieval tasks.
Structural Neighbor Expansion: A mechanism that expands retrieval scope along the graph topology (e.g., temporal neighbors) to capture necessary context for complex queries without sacrificing precision.

Experimental Results

The authors evaluated FT-RAG against baselines including Standard RAG, LangChain, Self-RAG, and MixRAG on the Multi-Table-RAG-Lib benchmark.

Retrieval Performance: FT-RAG achieved state-of-the-art results.
- Table-level Hit Rate: Improved by 23.5% over the best baseline.
- Cell-level Hit Rate: Improved by 59.2%, demonstrating superior ability to locate specific atomic values.
- In Context-Augmented tasks, FT-RAG maintained high precision where other models suffered from noise introduced by textual chunks.
Generation Performance:
- Exact Value Accuracy Recall: Increased by 62.2% compared to baselines.
- FT-RAG demonstrated a significant advantage in grounding numerical attributes, achieving a Value Accuracy of 54.5% compared to 22.0–33.9% for other methods.
Ablation Study: Removing the SAT-Graph caused performance to plummet below 40%, confirming the necessity of the graph structure. The Multi-modal Fusion module was found critical for value accuracy, while Structural Neighbor Expansion improved evidence comprehensiveness.

Significance and Claims

The paper claims that FT-RAG establishes a new state-of-the-art performance for complex reasoning over mixed-modality documents (tables and text). By shifting the paradigm from coarse-grained text retrieval to fine-grained, structure-aware graph navigation, the framework effectively mitigates hallucinations and factual errors common in table-based RAG. The authors assert that their method successfully bridges the gap between the rigorous logic of structured data and the descriptive power of unstructured text, providing a robust solution for high-value knowledge extraction in domains like financial analysis and scientific review. The introduction of the Multi-Table-RAG-Lib benchmark is also presented as a critical step toward more realistic and challenging evaluations of table-centric RAG systems.