GUIDE is a governed multi-agent framework designed to automate the conversion of complex, multimodal enterprise guideline documents into structured, executable work artifacts. By replacing manual processes that typically take 2–3 days per document, the system reduces turnaround time to 40–125 minutes while ensuring data consistency and traceability.
The Problem: Manual Bottlenecks
Enterprise guidelines often contain a mix of narrative text, complex tables, and images. Current automated systems struggle with these documents because they frequently produce hallucinations, fail to interpret table structures, and lack a governed workflow to validate outputs. Consequently, project and quality managers must manually read, interpret, and convert these documents into instructions and statements of work, creating a significant operational bottleneck.
How GUIDE Works
The framework treats document-to-artifact conversion as a data management challenge, utilizing a central, versioned rule store that acts as a shared data layer. Six specialized agents coordinate through this store using schema-validated contracts:
Parsing Agent: Separates deterministic text extraction from VLM-based visual processing.
Rule Extraction Agent: Identifies candidate rules and normalizes them into a 26-field schema.
Consistency Module: Uses embedding-based similarity and natural language inference to deduplicate and align rules.
Evaluation Module: Performs two-stage validation—L1 (structural) and L2 (semantic)—to score objects and determine if they require human review.
HITL Controller: Manages a dependency-aware human-in-the-loop workflow, routing only flagged items to experts.
Artifact Generation Agent: Transforms approved rules into deployment-ready documents, such as annotator guidelines and QA rubrics.
Performance Results
In an evaluation using 120 real-world enterprise documents, the researchers reported that GUIDE achieved a 96% document success rate. The system extracted 3,896 rules, with 71.4% of them auto-approved by the semantic evaluation engine. Additionally, the framework produced 812 deployment-ready artifacts. Compared to a monolithic baseline that lacked the governed rule store and validation layers, GUIDE significantly reduced hallucination rates (from 15.7% to 3.2%) and contradictions (from 7.8% to 2.9%).
Limitations and Considerations
The researchers note that the system’s performance is sensitive to the quality of the input documents; VLM extraction stability decreases when processing low-quality scans or complex, borderless tables. Furthermore, the framework currently struggles with adapting technical rules for non-expert audiences, and its calibration mechanism requires accumulating zero-edit human approvals over time, meaning early deployment cycles may have limited scoring stability. Future work is intended to address these challenges, including expanding the system to support multilingual settings and broader regulatory or clinical domains.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!