retail-knowledge

# Retail Knowledge — Q&A Engine ## Overview This skill answers questions using the store's configured knowledge base. It is the foundational skill for all retail digital employee roles. **Depends on:** Knowledge base populated via `retail-agent-setup` Step 03. If no knowledge base is configured, guide the user to run `retail agent setup` first. --- ## Query Routing When a question arrives, classify it and route to the correct knowledge domain: | Query Type | Keywords | Knowledge Domain | |------------|----------|-----------------| | Product info | 产品/商品/成分/规格/面料/尺寸/功效 | `products` | | Policy | 退货/换货/退款/保修/三包/质保 | `policies` | | Promotion | 活动/优惠/折扣/满减/赠品/促销 | `promotions` | | FAQ | 怎么/如何/可以/能否/多久 | `faqs` | | Store info | 地址/营业时间/几点/电话/停车 | `store_info` | | Membership | 积分/会员/等级/VIP/余额 | `membership` | | Recommendation | 推荐/适合/送礼/比较/哪个好 | → hand off to `product-recommender` skill | | Inventory | 有没有/还有/库存/现货 | → hand off to `inventory-query` skill | | Complaint | 坏了/质量问题/投诉/要退 | → hand off to `complaint-handler` skill | --- ## Answer Construction Rules ### Rule 1: Always ground answers in the knowledge base Never invent product specs, policy terms, or promotion details. If the knowledge base doesn't have the answer, use the configured `unknown_response`. ### Rule 2: Be specific Bad: "我们有退货政策" Good: "购买后7天内，商品未使用且保留吊牌，可申请无理由退货。退款将在3个工作日内到账。" ### Rule 3: Cite conditions when relevant For policies and promotions, always mention key conditions and exceptions. Example: "满300减50，不与其他优惠叠加，促销商品除外。" ### Rule 4: Match persona tone Apply the configured `persona_config` (name, tone, address form, emoji usage). **Reference:** [answer-style-guide.md](references/answer-style-guide.md) ### Rule 5: Handle unknowns gracefully If no matching knowledge base entry exists: 1. Say so honestly (use configured `unknown_response`) 2. Offer an alternative: escalate, or suggest the user contact staff 3. Log the query internally for Step 12 gap digest Never say "I don't know" bluntly — soften it while staying honest. --- ## Multi-Turn Conversation Maintain context across turns within a session: - Remember what product was mentioned earlier ("那款" / "刚才说的那个") - Remember stated preferences ("她喜欢素色" → filter subsequent answers) - If user backtracks or changes topic, reset context gracefully **Reference:** [conversation-patterns.md](references/conversation-patterns.md) --- ## Knowledge Base Structure Expect the knowledge base (populated by `retail-agent-setup`) in this format: ```json { "products": [ { "sku": "...", "name": "...", "description": "...", ... } ], "policy_entries": [ { "policy_id": "...", "title": "...", "full_text": "...", ... } ], "promotions": [ { "promo_id": "...", "title": "...", "rules": "...", ... } ], "faqs": [ { "faq_id": "...", "question": "...", "answer": "...", ... } ], "store_info": { "name": "...", "address": "...", "hours": "...", "phone": "..." }, "membership": { "levels": [...], "points_rules": "...", "query_method": "..." } } ``` **Reference:** [kb-schema.md](references/kb-schema.md) — full schema with field descriptions. --- ## Fallback Behavior If the knowledge base is empty or missing a domain: | Missing Domain | Fallback Response | |---------------|-------------------| | No products | "我们的商品信息正在整理中，请联系店员了解详情。" | | No policies | "退换货政策请联系门店工作人员确认。" | | No promotions | "目前暂无特别优惠活动，欢迎关注我们的公众号获取最新信息。" | | No store info | Escalate to configured L1 contact | --- ## Script: Knowledge Base Search Use `scripts/kb_search.py` when the knowledge base is a local JSON file and a direct keyword/semantic search is needed before constructing an answer. **Reference:** [search-strategy.md](references/search-strategy.md) — when to use exact match vs. fuzzy match vs. LLM synthesis.