Aether: A Hybrid RAG Workflow Reasoning Engine for Financial Compliance Documents

Aether is an AI workflow engine designed specifically for financial compliance document processing. It adopts a planner-executor-judger loop architecture, combines hybrid RAG retrieval and complete audit trails, and enables automated processing from unstructured documents to auditable results. Its design focuses on reliability, auditability, and cost control, making it suitable for highly regulated financial scenarios.

Financial compliance analysts often face unstructured documents in multiple formats such as CSV and PDF, needing to perform tasks like number verification, rule cross-referencing, and violation marking, relying on experiential judgment. Traditional tools are either rigid and unable to adapt to changing inputs, or too loose to meet compliance audit requirements, creating a gap in both flexibility and rigor.

Three-Tier Agent Architecture: The Planner (Claude Opus) generates structured execution plans; the Executor (no LLM involved) mechanizes tool scheduling (e.g., DuckDB for data loading, SQL queries); the Judger (Claude Haiku) compares outputs against goals—if partially passed, re-plan (max 2 rounds), if failed, escalate to manual review.

Hybrid RAG Retrieval: Combines BM25 sparse retrieval and dense vector retrieval (all-MiniLM-L6-v2), merged via RRF and re-ranked using flashrank. The retrieval precision@5 reaches 96%, and the pipeline has high determinism (standard deviation of 0.00 over 5 runs).

Success rate is 87% (13/15) across 15 test cases, with a reliable lower bound proving practical feasibility. Each run costs approximately $0.65, as the model routing strategy (Opus for planning, Haiku for judging) balances capability and cost. Key failure modes: Planner SQL hallucinations, intermittent over-escalation by the Judger.

1. Directly use the Anthropic SDK without framework abstraction to ensure transparency; 2. Agent outputs are validated via Pydantic with error retries (max 3 times); 3. The engine is separated from the domain—adapt to different domains by switching the PROMPTS_DIR; 4. SQLite tracks all LLM/tool calls, with auditability as a core design principle.

Limitations: Not a general AI assistant—focused on document processing workflows; not a framework wrapper—built from scratch; frontend is only MVP-level Streamlit.

Future Directions: Intelligent retrieval (planner-driven query decomposition), query-aware re-ranking, deeper retrieval evaluation (cross-document recall, ablation studies, etc.), record and fix each retrieval failure mode.

Practical Significance: Addresses the pain point of structured extraction from financial compliance documents and meets audit trail requirements; the architecture can be referenced in other domains.

Summary: Aether is a deeply optimized financial document processing engine. Developed by an individual over 6 weeks, it is a fully functional project with rigorous evaluation, providing a reference for AI system deployment in regulatory environments. Its design decisions have universal guiding significance for production-level AI systems.