CIFQA: A Deterministic Tool-Grounded Multi-Agent LLM Framework for Financial Query Answering

Battle Count: 50

The paper focuses on financial query answering (fixed deposit calculations, interest computation, TDS rules) rather than quantitative trading, market prediction, or portfolio optimization. While the deterministic computation paradigm could theoretically be applied to trading systems, the paper does not address trading strategies, market microstructure, or algorithmic execution. The relevance is indirect through the general principle of separating language understanding from deterministic computation.

The system requires orchestrating multiple LLM agents (router, extractor, planner, response generator) with deterministic Python computation engines (rate lookup, interest computation, calendar engine, rule engine). The multi-agent pipeline with structured inter-agent communication, domain-specific financial logic, calendar-aware computations, and strict evaluation protocols adds significant engineering complexity. However, the modular design and clear separation of concerns make individual components manageable.

The paper describes the architecture, pipeline stages, and deterministic computation engines in detail. However, no code repository or specific implementation details (e.g., exact prompts, rate table formats, API configurations) are provided. The evaluation dataset of 126 FD queries is described but not explicitly stated as publicly available. The system uses API-based LLM inference, making exact reproduction dependent on model versions and API parameters.