Explore Agentic KG construction

[jexp]

• Explore KG construcition as a tool based ReACT loop where processing of larger documents happens in multiple phases

Example tools:

• fetch existing entities from graph
• get graph schema
• merge entities
• extract entities
• extract relationships
• enrich entities or relationships
• ...
• consolidation

[dhiaaeddine16]

@jexp,

Your proposal for Agentic KG construction via a tool-based ReACT loop aligns well with modern AI agent architectures. Here are two key paradigms:

Single Agent with Dynamic Tool Use (The Orchestrator Pattern)

This approach uses one intelligent agent to dynamically select and sequence tools. It acts as a central orchestrator, choosing the right tool (e.g., extract entities, merge entities, get graph schema) at each step, managing the entire workflow, and handling final consolidation.

Pros:

Simplicity: Easier to design and implement initially due to centralized control.
Cohesion: Tighter integration of steps, potentially reducing coordination overhead.

Cons:

Scalability Limit: Can become a bottleneck for very large or diverse tasks.
Complexity Risk: A single agent can become overly complex if too many capabilities are packed into it.
Single Point of Failure: If the orchestrator fails, the entire process stops.

Multi-Agent Systems

For more complex scenarios, multi-agent systems offer specialization and scalability. Imagine dedicated agents: an "Extraction Agent," a "Merger/Consolidation Agent," or an "Enrichment Agent," each handling specific tasks. This allows for parallel processing, enhanced robustness, and easier module management.

Pros:

Scalability: Distributes workload across multiple specialized agents, enabling parallel processing.
Modularity: Easier to develop, maintain, and upgrade individual components.
Robustness: Failure of one agent may not halt the entire system.

Cons:

Coordination Overhead: Requires robust communication and conflict resolution mechanisms between agents.
Complexity: Can be more challenging to design and debug due to distributed nature.
Resource Intensive: May require more computational resources to manage multiple agents.

Both paradigms offer distinct advantages for building a robust and scalable KG construction system. I'm keen to discuss specific implementation details further.

[dhiaaeddine16]

@jexp

I’d like to highlight KARMA: Multi-Agent LLMs for Automated Knowledge Graph Enrichment (arXiv Feb 10, 2025) by Lu & Wang. KARMA uses a hierarchical, nine‑agent architecture spanning ingestion, summarization, entity/relation extraction, schema alignment, conflict resolution, and final evaluation coordinated by a Central Controller Agent.
This aligns strongly with our proposed tool-based ReACT loop for agentic KG construction. KARMA exemplifies the multi-agent paradigm, with specialized agents working in an orchestrated workflow. We could draw inspiration from their conflict-resolution debate phase and multi-stage verification when designing our merging, enrichment, and consolidation tools.

KARMA repo

[jexp]

Also checking papers and deeplearning AI course from Andreas Kollegger.
Collect ideas in a notebook and in this issue.
Need to find a good compromise between speed and quality.
Probably need to process larger document chunks or whole documents at once so that the agent has more full context.