Rummage Sale Bot Agent: An LLM-based Intelligent Bargaining Sales Robot

This article introduces the open-source project Rummage Sale Bot Agent, an interactive chatbot that simulates a flea market seller. It can manage inventory, handle transactions, and achieve dynamic price negotiation through the tool-calling capabilities of large language models (LLMs), demonstrating the innovative application of LLM tool calling in e-commerce scenarios.

In the rummage sale (flea market) scenario, sellers are mostly individuals, with a wide variety of goods and high price elasticity. Transactions are full of interpersonal interactions and bargaining, which poses challenges to automated sales systems: How to enable AI to complete the entire process from display to transaction while maintaining flexibility? This project is an exploration of this challenge, building a fully functional AI sales agent to demonstrate the application of LLM tool calling in e-commerce scenarios.

The project includes three core modules:

1. Inventory Management System: Product information modeling (including name, description, initial price, reserve price, etc.), real-time status tracking, dynamic listing/delisting, and proactive recommendation;
2. Secure Transaction Processing: Quotation management, transaction confirmation, receipt generation, and fraud prevention;
3. Dynamic Price Negotiation Engine: Context awareness (analyzing user purchase intention, budget, etc.), strategy selection (adjusting strategies for different buyer types), concession rhythm (simulating human concessions without going below the reserve price), and relationship maintenance (treating non-transaction users friendly).

1. Tool Calling Architecture: The AI can call tool functions such as get_inventory() and get_item_details(), decoupling business logic from AI decision-making;
2. Prompt Engineering: Carefully designed system prompts to set the AI's role (friendly seller), behavioral guidelines (honest and flexible), decision framework, and safety boundaries (not accepting prices below the reserve price);
3. State Management: Maintaining conversation stages, user-interested products, quotation status, conversation history summaries, etc., in memory.

Typical interaction flow: User asks about used books → AI calls the inventory tool to filter and recommend books → User offers $15 for a Python book → AI makes a small concession to $18 (not reaching the reserve price of $16 but the user has purchase intention) → User offers $17 → AI accepts the deal (above the reserve price and the user pays immediately). This example demonstrates the AI's bargaining strategy and tool-calling capabilities.

The core model of the project can be extended to:

• AI operation agency for C2C second-hand trading platforms (such as Xianyu, Zhuanzhuan);
• Customer service for small retailers (independent bookstores, antique shops, etc.);
• Auction auxiliary tools (pre-auction communication to collect intended prices);
• Price intelligence collection (simulating buyer interactions to obtain market prices).

Limitations of the current version:

1. Multi-turn memory limitation (long conversations tend to forget early details);
2. Multi-user concurrency (inventory synchronization needs to be strengthened);
3. Payment method expansion (only simulates cash transactions, needs to connect to real payment gateways);
4. Depth of emotional understanding (perception of user emotions is relatively basic).

Rummage Sale Bot Agent focuses on the flea market scenario and demonstrates the practical application of LLM tool calling in e-commerce. The core insight is that effective AI agents need to exert value under clear boundaries and tool support. The tech stack includes LLM backend, Python dialogue framework, and memory storage. The learning value for developers includes: tool calling patterns, state machine design, prompt engineering, and business logic integration.