Using Neural Graphs (GNNs) for Intelligent List Generation and Organisation

What is the value of a well-organised and insightful list? It transforms overwhelming data into actionable intelligence. In this research we introduce AI models designed to identify, organise and evolve lists of information from proprietary data. These models dynamically discover meaningful organisational structures within unstructured or fragmented data. The system provides the backbone for both highly personalised search and living knowledge systems that actively reorganise information to surface unseen relationships and insights that drive better decisions.

Abstract

We propose a dual-agent architecture that fundamentally changes how AI understands and organizes lists. The first agent executes precise list operations using graph neural networks, while the second continuously builds and updates knowledge structures from your data. This separation allows each agent to excel at its task—one achieving millisecond response times for sorting and filtering, the other identifying deep patterns and relationships. The result is an AI-powered search and organization system that adapts to individual contexts, learns from usage patterns, and handles enterprise-scale data without sacrificing speed or accuracy.

1. Introduction

The exponential growth of unstructured data across fragmented sources presents an unprecedented challenge for both individuals and enterprises seeking to organise, search, and extract meaningful patterns from information. While recent advances in neural sequence modelling and graph neural networks have demonstrated remarkable capabilities in specific domains, current approaches suffer from critical limitations when applied to the dynamic, multi-faceted nature of real-world list organisation tasks.

1.1 Why This Matters: The Hidden Cost of Poor Information Organisation

Knowledge workers spend an average of 2.5 hours per day searching for information—nearly 30% of their working time. For a company with 1,000 knowledge workers, this translates to £12 million in annual productivity loss at average UK salaries. The problem isn’t just time—it’s missed opportunities, duplicated efforts, and decisions made without access to relevant information.

Current enterprise search solutions and knowledge management systems fail because they treat information organisation as a static problem. They create fixed taxonomies, unchanging search rankings, and rigid folder structures that become obsolete the moment they’re deployed. As projects evolve, teams change, and priorities shift, these static systems become barriers rather than enablers.

The financial impact extends beyond lost productivity. McKinsey research indicates that effective knowledge management can reduce project development time by 35% and increase innovation success rates by 28%. For a technology company with £100 million in R&D spending, this represents £35 million in accelerated value delivery.

Personal productivity faces similar challenges. Professionals manage information across dozens of tools—email, documents, chat, wikis, code repositories—each with its own search and organisation paradigm. The cognitive overhead of remembering where information lives and how to find it creates a constant drain on mental resources better spent on creative and strategic work.

1.2 The List Organisation Challenge

Lists, in their broadest sense, represent fundamental structures for human information processing. From simple ordered sequences to complex hierarchical taxonomies, lists serve as cognitive scaffolds that enable us to navigate, understand, and manipulate information. However, the diversity of list types—ranked results, categorised data, temporal sequences, hierarchical structures—combined with the variety of operations required—sorting, filtering, grouping, transforming—creates a combinatorial explosion of complexity that existing AI systems struggle to address comprehensively.

Current approaches typically fall into two categories: task-specific models that excel at narrow operations but lack generalisability, and large language models that demonstrate broad capabilities but suffer from computational inefficiency and lack of precise control. Neither approach adequately addresses the need for systems that can both understand the semantic intent behind list organisation tasks and execute precise structural operations with mathematical certainty.

1.3 Limitations of Static Embedding Approaches

The predominant paradigm in information retrieval and AI-powered search relies on static embeddings—fixed vector representations computed once and reused across contexts. Systems like SBERT have revolutionised semantic search, but their static nature fundamentally limits their ability to adapt to evolving information landscapes or context-specific organisational requirements in enterprise search solutions.

Consider an enterprise knowledge management scenario where the same document might need to be organised by technical relevance for engineers, business impact for executives, or regulatory compliance for legal teams. Static embeddings cannot capture these multiple, context-dependent organisational schemas. Similarly, as new information enters the system or relationships between entities evolve, static representations quickly become outdated, requiring expensive recomputation of entire embedding spaces.

1.4 The Promise of Graph Neural Networks for Intelligent Information Organisation

Graph Neural Networks offer a compelling alternative by representing information as interconnected nodes and edges that can be dynamically traversed and reorganised. Recent advances in hierarchical GNNs and differentiable graph algorithms demonstrate that neural networks can learn to perform complex structural operations while maintaining the flexibility to adapt to new patterns.

Critically, GNNs can embed atomic operations—fundamental list manipulations like sorting, filtering, and Boolean operations—directly into their architecture. This capability enables end-to-end learning of complex list organisation strategies while maintaining mathematical guarantees about operation correctness. When an operation is truly atomic (e.g., alphabetical sorting), the model can learn to execute it with perfect accuracy, unlike probabilistic approaches that approximate such operations. Furthermore, these atomic operations serve as building blocks that can be combined through the graph structure to create increasingly sophisticated organisational behaviours.

1.5 Bayesian Foundations for Dynamic List Sorting and Filtering

While GNNs provide the structural framework for embedding atomic operations, the true power of our approach lies in incorporating Bayesian statistical primitives that can be composed to create arbitrarily complex filtering and sorting capabilities. By embedding fundamental probabilistic operations—conditional probability (A given B), joint probability (A and B), and negation (not A)—directly into the neural architecture, we enable the system to learn sophisticated organisational strategies through composition of simple, mathematically rigorous building blocks.

This Bayesian approach addresses a critical gap in current list organisation systems: the ability to handle uncertainty and partial information while maintaining mathematical correctness. Traditional sorting algorithms assume complete information and deterministic outcomes, but real-world list organisation often involves probabilistic reasoning—ranking documents by likely relevance, filtering items by probable category membership, or grouping entities by uncertain relationships.

The compositional nature of these Bayesian operations enables exponential expressiveness. Simple primitives like P(A|B), P(A∩B), and P(¬A) can be combined through the graph structure to express complex queries such as “find all documents probably relevant to machine learning but not deep learning, given the user’s reading history.” This compositional approach mirrors how humans naturally combine simple logical operations to form complex organisational criteria, providing both intuitive interpretability and mathematical rigour.

1.6 Dual-Agent Architecture: Intelligent AI-Powered Search

Drawing inspiration from recent advances in multi-agent AI systems and Google DeepMind’s Talker-Reasoner architecture, we propose a dual-agent framework that separates list organisation into two complementary components:

Agent 1: List Organisation Agent (LOA)

• Specialises in executing precise list operations using GNN-embedded atomic operations
• Handles sorting, filtering, grouping, and transformation tasks
• Maintains operation correctness guarantees for atomic operations
• Optimised for speed and accuracy in list manipulation

Agent 2: Knowledge Structure Agent (KSA)

• Dynamically constructs and maintains graph-based knowledge representations
• Identifies patterns, relationships, and emerging structures in data streams
• Continuously updates knowledge graphs based on new information
• Provides contextual embeddings and structural priors to the LOA

This separation enables each agent to be optimised for its specific role while maintaining tight integration through well-defined interfaces. The KSA can employ more computationally intensive methods for knowledge extraction and pattern recognition, updating its representations asynchronously, while the LOA maintains real-time performance for user-facing operations.

1.7 Research Objectives and Contributions

This research aims to develop, validate, and deploy a production-ready dual-agent system for intelligent list organisation. Our primary contributions include:

1. Theoretical Framework: A formal model for representing list operations as graph transformations, enabling principled reasoning about operation composition and correctness guarantees.

2. Neural Architecture Design: Novel GNN architectures that embed atomic operations while maintaining differentiability, enabling end-to-end learning of complex organisational strategies.

3. Dynamic Knowledge Construction: Algorithms for real-time knowledge graph construction and update, including methods for conflict resolution and uncertainty quantification.

4. Dual-Agent Coordination: Protocols for efficient communication and coordination between agents, including attention mechanisms for dynamic context selection.

5. Empirical Validation: Comprehensive evaluation on both synthetic benchmarks and real-world applications in enterprise knowledge management and personal productivity.

6. Production Implementation: Open-source reference implementation demonstrating scalability to millions of items and thousands of concurrent users.

1.8 Applications and Impact

The proposed system addresses critical needs across multiple domains:

Enterprise Knowledge Management: Organisations struggle with information silos and inefficient knowledge sharing. Our system enables intelligent aggregation across disparate data sources, automatic organisation by multiple criteria, and real-time adaptation to changing business contexts. This AI-powered search capability transforms how enterprises access and utilise their collective knowledge.

Personal Productivity and Personalized Knowledge Management: Individuals face information overload from emails, documents, messages, and web content. The dual-agent framework provides personalised organisation that learns from usage patterns while maintaining user control over organisational principles. This personalised approach to information organisation reduces cognitive load and accelerates decision-making.

Complex Project Management: Modern projects—from software development to construction to product design—involve intricate hierarchies of interconnected tasks, requirements, and stakeholders. Software projects exemplify this complexity with architectural decisions cascading through design patterns, integration requirements, and testing protocols. Our system enables dynamic organisation of these multi-layered concerns, allowing seamless navigation from high-level strategic views to developer-specific daily tasks, while automatically maintaining consistency across all organisational levels.

Intelligent and Personalised Search: Traditional search engines return static rankings that fail to adapt to individual contexts and evolving needs. Our dual-agent system enables truly personalised search experiences where results are dynamically organised based on user patterns, current context, and learned preferences, while maintaining near real-time performance.

By replacing static, one-size-fits-all approaches with dynamic, context-aware organisation, this research promises to fundamentally transform intelligent information organisation in an increasingly complex digital landscape. The convergence of AI-powered search, personalised knowledge management, and dynamic list sorting creates unprecedented opportunities for both individual productivity and enterprise efficiency.