CQs as Executable Plans for Controlled RAG

💡Novel CQ plans make RAG hallucination-proof for KG storytelling apps
⚡ 30-Second TL;DR
What Changed
Repurposes design-time CQs into runtime narrative plans for evidence-closed generation
Why It Matters
Provides blueprint for auditable, controllable storytelling systems reducing LLM hallucinations. Enables personalized heritage narratives with quantifiable RAG trade-offs. Actionable for KG-based AI applications beyond culture.
What To Do Next
Download Live Aid KG from arXiv and prototype CQ plans in your RAG pipeline.
Key Points
- •Repurposes design-time CQs into runtime narrative plans for evidence-closed generation
- •Introduces Live Aid KG dataset linking concert data to Music Meta Ontology and multimedia
- •Compares three RAG strategies: symbolic KG-RAG, text-enriched Hybrid-RAG, structure-aware Graph-RAG
🧠 Deep Insight
AI-generated analysis for this event — not the original article.
🔑 Enhanced Key Takeaways
- •The approach addresses the 'hallucination-by-omission' problem in RAG by enforcing a strict evidence-closed constraint, where the LLM is restricted to generating content only from the subgraph retrieved via the CQ-derived plan.
- •The Live Aid KG utilizes the Music Meta Ontology (MMO) to bridge structured concert metadata with unstructured multimedia assets, enabling cross-modal retrieval that standard vector-based RAG often fails to capture.
- •The study identifies that while Graph-RAG excels in structural connectivity, it often suffers from 'context dilution' in long-form narrative tasks, a limitation this CQ-based planning architecture specifically mitigates by prioritizing path-based relevance over global graph density.
🛠️ Technical Deep Dive
- •Architecture: Neuro-symbolic pipeline where a 'Planner' LLM decomposes user queries into a sequence of SPARQL-like executable operations based on predefined Competency Questions (CQs).
- •Execution Engine: A runtime interpreter that maps CQ-derived plans to specific graph traversal patterns, ensuring the retrieved evidence is strictly aligned with the narrative requirements.
- •Dataset Composition: The Live Aid KG integrates 1985 concert performance logs, artist discographies, and event-specific multimedia metadata, structured to support multi-hop reasoning across temporal and spatial dimensions.
- •Evaluation Metrics: Uses a combination of ROUGE-L for coherence, FactScore for factual consistency, and a custom 'Plan-Adherence' metric to measure how strictly the generation follows the CQ-derived execution path.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.