HalluGuard: A Reverse RAG Hallucination Detection Framework with Zero LLM Inference

HalluGuard is an open-source hallucination detection framework developed by the nakata-app team. Its core feature is zero LLM dependency during the inference phase. By adopting a reverse RAG architecture and leveraging NLI validators, intelligent voting strategies, and stream processing mechanisms, it achieves efficient, low-cost real-time hallucination detection. It solves the problems of high cost and large latency caused by traditional methods' reliance on additional LLM calls, providing reliable content security guarantees for LLM applications.

With the widespread application of Large Language Models (LLMs) across various industries, the hallucination problem has become increasingly prominent—models generate content that seems reasonable but is inconsistent with facts or unverifiable, posing serious risks to critical applications. Traditional hallucination detection methods require additional LLM calls for verification, increasing inference costs, latency, and resource consumption.

1. Reverse RAG Architecture

Unlike traditional RAG, which retrieves from knowledge bases to enhance generation, HalluGuard uses generated content as a query to retrieve evidence from trusted knowledge sources and judges logical relationships via NLI models. Its advantages include zero LLM inference cost, low latency, and high scalability.

2. NLI Validator

A core component that judges the entailment, contradiction, or neutral relationship between generated content and evidence. It uses an optimized lightweight model to achieve millisecond-level responses.

3. Voting Strategy and Confidence Evaluation

Retrieve evidence from multiple knowledge sources, perform independent NLI judgments, calculate comprehensive confidence based on evidence quality and consistency, and set thresholds to mark high-risk content.

4. Stream Processing Architecture

Supports detection while generating, with sentence/paragraph-level granularity. Detection frequency and trigger strategies are configurable.

• Enterprise knowledge base Q&A: Detect consistency between answers and internal documents, prevent fabrication of policies and processes, and provide a traceable evidence chain;
• Medical and legal consultation: Verify the accuracy of professional terms and legal provisions, mark risky statements, and assist manual review;
• Content generation platforms: Real-time detection of factual accuracy to reduce the risk of false information spread;
• Educational auxiliary tools: Ensure the accuracy of teaching content and detect mathematical derivations and factual statements.

Feature	Traditional Methods	HalluGuard
LLM Calls	Requires additional calls	Zero LLM
Inference Latency	High (seconds)	Low (milliseconds)
Deployment Cost	High (API fees)	Low (local deployment)
Interpretability	Weak	Strong (evidence chain)
Real-time Performance	Limited	Stream-supported
HalluGuard eliminates additional LLM overhead, its low latency does not affect user experience, multi-layer verification reduces missed detection rates, and the complete evidence chain supports compliance.

HalluGuard offers multiple deployment options:

• Localcal deployment: Suitable for data-sensitive scenarios;
• Cloud service: Supports high-concurrency scaling;
• API integration: RESTful API for easy access to existing systems;
• Plugin extension: Supports integration with mainstream LLM frameworks.

HalluGuard represents important progress in the field of large model security, providing cost-effectiveness, performance guarantees, security enhancement, and auditability for LLM applications. Future directions include: supporting multi-language detection, integrating with more knowledge graphs, adaptive threshold learning, and training domain-specific NLI models.