A Multi-Agent System for Extracting and Querying Financial Documents

Financial documents—10-K/10-Q filings, earnings releases, prospectuses, invoices, and bank statements—combine dense prose, complex tables, and heterogeneous layouts. We propose a multi-agent architecture that orchestrates document ingestion, layout-aware extraction, knowledge graph construction, retrieval-augmented reasoning, and auditable question answering. Our design integrates advances in agentic LLM collaboration (AutoGen, CAMEL), reasoning-and-acting prompts (ReAct), and document AI (LayoutLMv3, Donut) with domain standards (FIBO ontology, ISO 20022 messaging) and public data interfaces (SEC EDGAR, XBRL). We specify components, routing policies, evaluation protocols (FinQA, TAT-QA), and governance for factuality, lineage, and compliance. This paper contributes: (1) a reference architecture and coordination patterns for financial-document intelligence, (2) a schema-aligned extraction pipeline that yields structured, explainable facts, and (3) a reproducible evaluation plan emphasizing numeric reasoning, table/text fusion, and citation fidelity.