Extracting structured information from visually rich documents at enterprise scale demands both the reasoning capability of large language models and the efficiency of deterministic execution. Current approaches either deploy LLMs as instance-level analysts, incurring per-document inference costs that are prohibitive for repetitive templates, or rely on manually authored vendor-specific prompts that do not scale beyond a handful of high-volume vendors. We present CACHE-ED2, a framework that repositions the LLM as a system-level developer: given a novel document format, a multimodal ReAct-based agent synthesizes a reusable Document Extraction DSL (DocExDSL) program encoding the complete extraction logic (strategies, disambiguation, transforms, and automated reasoning validation), then caches it for deterministic, LLM-free execution on all subsequent documents of the same format. A contrastive Document Format Encoder routes incoming documents to cached programs with perfect hit rates, and a HITL Feedback Analysis Agent incorporates analyst corrections to refine programs over time. On public benchmarks (CORD v2, FATURA) and production streams spanning seven business verticals (580,000+ documents, 11,500+ vendors), CACHE-ED2 achieves near-perfect extraction accuracy (∼0.99) with a 1.0 extraction rate on document-present fields and perfect format matching across 50+ benchmark templates, matching or exceeding manually authored vendor-specific prompts, while reducing cumulative token consumption by ∼2.6× over first 1,000 documents.

Information extraction (IE) from Visually Rich Documents (VRDs) containing layout features along with text is a critical and well-studied task. Specialized non-LLM NLP-based solutions typically involve training models using both textual and geometric information to label sequences/tokens as named entities or answers to specific questions. However, these approaches lack reasoning, are not able to infer values not explicitly present in documents, and do not generalize well to new formats. Generative LLMs-based approaches proposed recently are capable of reasoning, but struggle to comprehend clues from document layout especially in previously unseen document formats, and do not show competitive performance in heterogeneous VRD benchmark datasets. In this paper, we propose BLOCKIE, a novel LLM-based approach that organizes VRDs into localized, reusable semantic textual segments called semantic blocks, which are processed independently. Through focused and more generalizable reasoning,our approach outperforms the state-of-the-art on public VRD benchmarks by 1-3% in F1 scores, is resilient to document formats previously not encountered and shows abilities to correctly extract information not explicitly present in documents.