Goal‑oriented conversational systems must answer factual queries, comprehend visitor‑provided information, and advance business objectives without turning into rigid questionnaires.
This paper introduces a Symbolic‑RAG‑Generative architecture centered on the Goal‑oriented Retrieval‑Augmented Conversation Engine (GRACE).
An instruction‑constrained Business Goal Compiler translates business intent into an immutable objective set, a normalized priority vector, canonical questions, and an initial state vector.
At runtime GRACE consumes the full conversation history, the latest visitor message, the current state, and a grounded answer produced by a separate RAG component. It updates completion solely from visitor‑authored evidence and selects a single context‑modulated follow‑up question.
The core policy maximizes expected business progress under a minimum visitor‑utility constraint: $$\max_{\pi}\ \mathbb{E}[\text{business\_progress}]\quad\text{s.t.}\ \mathbb{E}[\text{visitor\_utility}]\geq\tau$$, where \tau is the utility threshold.
We formalize the state space, monotonic transitions, source separation, question modulation, and constrained policy, and present a reference layered architecture.
Evaluation comprises 24 English real‑estate and 10 Spanish professional‑cleaning conversations, totaling 119 visitor turns, covering standard, multi‑goal, RAG‑detour, validation, refusal, and robustness scenarios.
Results report 84.9% exact state‑transition accuracy, 91.6% evidence precision, 89.6% evidence recall, 100% monotonicity, and 94.1% terminal‑state accuracy, demonstrating strong auditability and resilience of symbolic states in real‑world business dialogues.
The study validates the synergy of symbolic reasoning and retrieval‑augmented generation, offering a pathway to interpretable, auditable, and business‑driven conversational agents.
Review