Rebate Assistant

Problem

Rebate assistants often mix authorization, product search, link conversion, and withdrawal into ad hoc replies. Agents may add explanations around tool output or invent rebate links when an API call fails. Rebate Assistant keeps the user path to one chain: provide a product, recognize intent, search or generate a rebate link, and return the result.

How it works

• S01 Authorization and tutorial: handles entries such as rebate, tutorial, withdrawal tutorial, and account balance. It explains the required WeChat authorization flow and supports fixed actions like start_auth, confirm_auth, withdraw_prepare, withdraw_confirm, and account_balance.
• S02 Link rebate: recognizes Taobao, JD, and Pinduoduo product links, then proceeds through platform recognition, product lookup, lease request, and rebate link generation.
• S03 Product search: when no link is provided, extracts raw_text and an identifiable platform from inputs such as brand plus model, category plus budget, or gift target. It outputs product_search_intent and search_request for downstream search.
  The conversation layer routes through scripts/cli/rebate_assistant_router.js; script stdout is returned as Markdown, avoiding extra polishing, paraphrasing, or product recommendations.

Boundaries and notes

• Only Taobao, JD, and Pinduoduo product links are supported for rebate generation.
• Do not generate rebate links, rebate codes, or payout promises when authorization is missing, no product is found, or the API fails.
• Withdrawal first checks available balance, asks for confirmation when the amount is valid, and submits /v1/withdraw/apply only after user confirmation.