A Human-in-the-Loop, LLM-Centered Architecture for Knowledge-Graph Question Answering

Technical Summary: A Human-in-the-Loop, LLM-Centered Architecture for Knowledge-Graph Question Answering

Problem Statement
While Large Language Models (LLMs) excel at natural language understanding, they face significant limitations in knowledge-intensive domains, including hallucinations, outdated information, and a lack of explainability. Existing Retrieval-Augmented Generation (RAG) approaches often rely on text-based retrieval, which struggles with multi-hop reasoning required to connect indirectly related entities. Conversely, Knowledge Graphs (KGs) offer precise, explainable querying but require expertise in query languages like Cypher or SPARQL. Current KG-based RAG pipelines often suffer from three specific design flaws: (1) linearizing graph triples erases structural context; (2) black-box retrieval hides the reasoning path; and (3) the lack of user-in-the-loop control prevents iterative refinement of mistaken entity matches or scope. Previous work by the authors demonstrated that while LLMs can translate questions to queries, the inability to debug or explain these queries necessitated complete regeneration upon failure, creating a disconnect between user intent and execution.

Methodology
The authors propose an interactive, modular framework where LLMs generate, explain, and iteratively refine Cypher queries based on natural language user input. The architecture consists of four core components:

1. Query Generation: Translates natural language questions into executable Cypher queries using a schema-aware prompt that constrains vocabulary to valid node and edge types.
2. Query Execution: Connects to a Neo4j database to execute the generated query and return structured results.
3. Query Explanation: Generates human-readable, step-by-step explanations of the query logic, including purpose summaries and plausibility checks (e.g., flagging inverted relationships or invalid node types).
4. Query Amendment: Allows users to provide natural language feedback to iteratively update an existing query without regenerating it from scratch, enabling targeted corrections to filters, entity types, or traversal logic.

The framework was evaluated through three distinct experiments:

• Synthetic Benchmark (Movie KG): A 90-query benchmark designed to test the LLM's ability to explain Cypher queries and detect injected errors (e.g., flipped relationships, nonsense labels, contradictory constraints). Five LLMs were tested on explanation quality and fault detection.
• Real-World Case Study 1 (MaRDI KG): A mathematical research data graph containing software packages, authors, and publications. Fourteen LLMs were tested on nine natural language questions to assess query generation accuracy with up to two amendment attempts.
• Real-World Case Study 2 (Hyena KG): A biological graph derived from long-term ecological data on spotted hyenas. The same 14 LLMs were tested on five expert-formulated questions regarding kinship and dispersal, again allowing for iterative refinement.

Key Contributions

1. Interactive Framework: A system that combines the expressiveness of Cypher with LLM accessibility, explicitly targeting the transparency and control limitations of existing KG-RAG systems. It enables users to inspect, explain, and refine queries through conversational feedback.
2. Empirical Evaluation: A comprehensive assessment across synthetic and real-world domains (mathematics and biology), measuring explanation quality, fault detection, and query generation accuracy across multiple LLMs.
3. Insights into Failure Modes: Identification of common LLM failure patterns, such as the omission of temporal constraints in summaries and significant performance variability across models when handling complex, expert-level questions.

Results

• Explanation Quality: On the synthetic movie benchmark, top-performing models (o1-preview, deepseek-reasoner-api, o3-mini) achieved over 70% accuracy in generating correct one-sentence summaries. However, a recurring failure mode across models was the omission of specific years mentioned in queries. In fault detection, models varied in their balance of sensitivity and specificity; o1-preview and deepseek-reasoner-api achieved high defect detection (>85%) with low false positive rates (<10%), whereas others struggled with specific perturbation types like flipped relationships.
• Query Generation (MaRDI): On the mathematical KG, o3-mini and GPT-5.2 solved all nine tasks, while most other models solved 7–8. Simple queries were handled reliably, but complex tasks like identifying shared authors across packages proved difficult for many models.
• Query Generation (Hyena): Performance gaps widened significantly on the biological KG. Only o3 and deepseek-reasoner-api achieved 100% accuracy (5/5 questions). Other models, including those that performed well on MaRDI, struggled, with several Ollama-hosted models achieving 0/5. Efficiency varied, with some models requiring multiple amendment attempts to reach a correct answer.

Significance and Claims
The paper claims that its primary contribution is bridging the gap between natural language and formal graph querying by making the reasoning path explicit and auditable. The authors argue that while LLMs can generate workable queries, their utility in real-world, expert-driven scenarios is highly variable and dependent on the domain and question complexity. The study demonstrates that synthetic benchmarks may not fully capture the difficulties of expert-sourced questions, as evidenced by the performance drop in the Hyena case study.

The authors modestly conclude that their framework successfully enables non-experts to interact with complex KGs without learning query languages, provided that the system supports iterative refinement. They emphasize that future work should focus on refining explanation prompts to retain temporal constraints and developing interactive tooling (e.g., diff views) to improve the efficiency of the human-in-the-loop process. The work does not claim to solve all KG-RAG challenges but provides a structured approach to transparency and control that reduces the risk of hallucination and improves trust in graph-based answers.