--- title: 'RecBot Framework: Dual-Agent Recommender' url: https://www.emergentmind.com/topics/recbot-framework type: topic --- # RecBot Framework: Dual-Agent Recommender The RecBot Framework is a dual-agent conversational recommendation architecture specifically designed to enable explicit, natural language user control over real-time recommendation policies within mainstream recommendation feeds. Distinct from traditional recommender systems, which rely primarily on passive, implicit feedback signals, RecBot allows users to influence item selection instantly by expressing constraints, preferences, and exclusions through unconstrained linguistic commands. The architecture integrates large language model (LLM)-enhanced intent parsing, tool-chain orchestration for policy adaptation, and simulation-augmented knowledge distillation for efficient deployment, demonstrating significant improvements in user satisfaction and business metrics over conventional approaches [2509.21317]. ## 1. Dual-Agent Architecture and Workflow RecBot operates through the coordinated action of two agents—a Parser Agent and a Planner Agent. The Parser Agent ingests the current recommendation feed $R_t$, raw natural language command $c_t$, and previous preference memory $P_t$, employing advanced LLM reasoning and dynamic memory consolidation to transduce free-form user feedback into structured, domain-specific preference representations $P_{t+1}$. The parsing process decomposes preferences into: - **Positive preferences ($P_{t+1}^{+}$):** Encompassing both hard, rule-based constraints (e.g., "below \$200") and soft, semantic preferences (e.g., "prefer romantic movies"). - **Negative preferences ($P_{t+1}^{-}$):** Discriminating strict exclusions (e.g., "not floral pattern") and soft disinclinations. Formally, $$ P_{t+1} = \mathcal{P}(R_t, c_t, P_t) = \{ P_{t+1}^{+},\; P_{t+1}^{-}\} $$ The Planner Agent receives $P_{t+1}$ alongside implicit behavioral history $H_t$, and orchestrates an adaptive tool-chain: - **Filter Tool:** Prunes candidate set $I$ to satisfy positive hard constraints and exclude negative hard constraints. - **Matcher Tool:** Computes relevance via two parallel mechanisms: - Semantic similarity ($s_{\text{sem}}$): Cosine similarity between item descriptions and positive intent, using contextual embedding models (e.g., Sentence-BERT, BGE). - Active-Intent-Aware Collaborative Filtering ($S_{\text{aia}}$): Captures personalized sequential patterns using multi-head self-attention (MHSA) and multi-head cross-attention (MHCA) between multimodal user history and parsed intent. - **Attenuator Tool:** Applies semantic penalties to items resembling user's negative instructions. - **Aggregator:** Combines Matcher and Attenuator scores into final ranking: $$ s_{\text{final}}(i) = s_{\text{match}}(i) + s_{\text{atten}}(i) $$ with $$ s_{\text{match}}(i) = \alpha \cdot s_{\text{sem}}(i, P_{t+1}^{+}) + (1-\alpha) \cdot S_{\text{aia}}(i, P_{t+1}^{+}, H_t) $$ $$ s_{\text{atten}}(i) = -\beta \cdot \mathrm{sim}(e_{\text{item}}(i), e(P_{t+1}^{-})) $$ This agent-driven pipeline allows RecBot to continuously update the top-$K$ feed $R_{t+1}$ in direct response to explicit user commands, as formalized by: $$ R_{t+1} = \mathrm{TopK}(\{s_{\text{final}}(i) | i \in I\}) $$ ## 2. Active User Commands and Semantic Parsing The framework facilitates direct, granular user control. Users may issue commands such as, "Show me blue dresses under \$100, exclude sleeveless styles," which the Parser Agent interprets into hard inclusion, soft inclusion, hard exclusion, and soft exclusion signals. The parsing process, leveraging both LLM semantic inference and domain-specific vocabulary mapping, enables nuanced extraction and separation of complex multi-attribute constraints within unconstrained natural language. Redundant or noisy feedback is robustly handled, whereas multi-constraint queries are disambiguated and split into atomic elements for downstream enforcement and matching. This paradigm shift from implicit, behavior-driven feedback to explicit command-based control improves model interpretability, user alignment, and precludes misattribution of motivation behind clicks or dislikes. ## 3. Adaptive Tool-Chain Orchestration and Scoring Upon receiving the structured intent, the Planner Agent invokes a dynamic tool chain: - The **Filter Tool** enforces strict attribute adherence (e.g., price, date, category exclusion). - The **Matcher Tool** computes semantic relevance by embedding both item descriptions and structured intents, using pre-trained transformers, and assesses collaborative signals through attention mechanisms over multimodal historical records. - The **Attenuator Tool** penalizes candidates with high semantic similarity to expressed exclusions. - The **Aggregator** synthesizes outputs from all tools to yield the ranked feed. Adjustable weighting parameters ($\alpha$ for match modes, $\beta$ for attenuation) provide fine-grained control over the balance between explicit command satisfaction and collaborative historical personalization. This agent composition enables the RecBot framework to adapt recommendation policies in real time in response to the full spectrum of user linguistic commands, mitigating delayed model updates and information cocooning. ## 4. Simulation-Augmented Knowledge Distillation Practical RecBot deployment faces challenges from the inference cost and latency of large, closed-source models. To address this, the framework implements simulation-augmented knowledge distillation: - **User Simulation Agent ($\mathcal{U}_{\text{sim}}$):** Generates multi-turn interaction trajectories, simulating realistic feedback, target setting, and interest drift scenarios. - **Teacher-Student Training:** A powerful teacher agent (e.g., GPT-4.1-powered RecBot) interacts with simulated users, generating training trajectories. Lightweight, open-source student agents distill both parsing and planning behaviors through next-token prediction objectives—the cross-entropy loss being formulated as: $$ \mathcal{L}(\theta) = \sum_{(x, y) \in \mathcal{D}_{\text{Mixed}}} \sum_{j} -\log P_\theta(y_j | x, y_{