GUIDE: Governed Unified Intelligence for Document-to-Artifact Generation in Enterprise Settings
The paper introduces GUIDE, a governed multi-agent framework that automates the extraction, validation, and synthesis of complex enterprise guideline documents into deployment-ready artifacts, significantly reducing processing time from days to under two hours while achieving high accuracy and rule extraction rates.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot how to do a very specific, complicated job, like sorting mail or painting a picture. You have a giant, messy instruction manual written for humans. It has long paragraphs, confusing tables that stretch across pages, and pictures with arrows pointing to things. Now, imagine you need to turn that messy manual into a perfect, step-by-step checklist that the robot can actually follow without getting confused. This is the world of "data management" and "artificial intelligence" (AI). In this field, scientists build computer programs called "Large Language Models" (LLMs) that can read and understand text, and "Vision-Language Models" (VLMs) that can look at pictures and text together. The big challenge is that these smart computers sometimes make things up (a problem called "hallucination") or get confused by messy tables, leading to instructions that are wrong or dangerous. If a company tries to do this conversion by hand, it takes days and is full of mistakes. This paper asks: Can we build a team of AI robots that work together, check each other's work, and follow strict rules to turn those messy manuals into perfect robot instructions quickly and safely?
The authors of this paper, a team from Centific Research, say "Yes, but only if we add a strict manager." They introduce a new system called GUIDE (Governed Unified Intelligence for Document-to-Artifact Generation). Think of GUIDE not as a single super-smart robot, but as a well-organized factory with six different specialized workers, all connected by a shared, version-controlled notebook.
Here is how the factory works:
- The Parser: The first worker reads the messy document. If it's text, they type it out perfectly. If it's a picture or a complex table, they use a special "eye" (a VLM) to see what's happening.
- The Rule Extractor: The second worker looks at that clean text and pulls out the important rules, turning them into a strict format with 26 specific fields, like a fill-in-the-blank form.
- The Consistency Checker: The third worker makes sure no two rules contradict each other and that no rules are repeated. They act like a librarian ensuring every book is in the right place.
- The Evaluator: This is the quality control inspector. They use a two-step check. First, they check if the form is filled out correctly (L1). Second, they ask a smart AI judge to grade the meaning of the rules on a scale of 1 to 5 (L2).
- The Human-in-the-Loop (HITL) Controller: If the AI inspector finds a rule that is confusing, missing a detail, or scores low, it doesn't just throw it away. It flags it and sends it to a human expert (like a Quality Manager) to fix. This is like a safety net.
- The Artifact Generator: Finally, the last worker takes all the approved rules and turns them into real-world documents, like job descriptions, training guides, or quality checklists, ready for the team to use.
The paper tested this system on 120 real-world enterprise guideline documents from industrial clients. These were tough documents, some with text, some with speech, and some with images. The results were impressive: GUIDE successfully processed 96% of the documents. It extracted 3,896 rules, and amazingly, 71.4% of those rules were so good that the system approved them automatically without needing a human to touch them. It even created 812 ready-to-use documents (artifacts).
Before this system, doing this work by hand took 2–3 days per document and was prone to errors. With GUIDE, the time dropped to between 40 and 125 minutes per document. The authors compared their team-based approach to a "one-shot" approach where a single AI tries to do everything at once. The single AI made many more mistakes: its "hallucination" rate (making things up) jumped from 3.2% to 15.7%, and it created many more duplicate or contradictory rules. This proves that having a team of specialized agents checking each other's work is much safer and more accurate than relying on one big brain.
However, the paper is honest about its limits. The system still struggles a bit with very messy, low-quality scans or tables that have no borders. Also, while the AI is great at following rules, it sometimes finds it hard to rewrite those technical rules into language that a non-expert (like a new employee) can easily understand. The authors suggest that in the future, they will teach the system to learn from human corrections to get even better at this.
In short, GUIDE shows that when you want to turn messy, complex human instructions into clean, reliable machine instructions, you don't just need a smarter AI; you need a smarter process. By breaking the job down, checking the work at every step, and knowing exactly when to ask a human for help, you can turn a task that used to take days into one that takes less than two hours, all while keeping the quality high.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.