VCL: Bridging Natural Language and Structured Control for Domain-Specific Document Analysis

Querying unstructured document corpora in knowledge-intensive domains including law and medicine requires practitioners to perform structured cognitive operations such as extraction, summarization, comparison, that go well beyond generic question answering and demand transparency and control over execution. Yet no existing approach fully meets this need: Unstructured Document Analytics (UDA) systems address the need for structured control but most expose SQL interfaces designed for data engineers, rather than domain practitioners. At the same time, long context-LLMs and RAG-based approaches leave multi-step query decomposition to implicit model reasoning, yielding opaque pipelines. To address this gap, we propose the Verbal Command Language (VCL), a formal declarative language that accepts natural language questions while providing SQL-like structural rigor at the processing level. VCL provides six cognitively grounded commands (SEARCH, EXTRACT, TRACE, SUMMARIZE, COMPARE, INTEGRATE), derived from field observation of judicial workflows and grounded in Speech Act Theory. Moreover, domain-specific elements are defined through natural language descriptions, following the definition-driven paradigm of modern information extraction approaches. VCL is executed by an LLM-based multi-agent engine, ensuring full traceability of every reasoning step, and evaluated on a real-world legal corpus against five baselines, including commercial systems (NotebookLM, Microsoft Copilot) and RAG variants. VCL with gpt-5.1 achieves the highest Recall (0.70) and F1 (0.65) across all systems, while maintaining Faithfulness above 0.93, suggesting that cognitively grounded, declarative query abstractions over agentic pipelines with natural language interfaces are a promising direction for document analytics in high-stakes professional domains.