Product Recommender

Overview

This skill handles all "help me choose" queries. It goes beyond listing products —
it understands the customer's situation, filters intelligently, and presents
a curated shortlist with reasons.

Depends on: products[] in knowledge base (Step 03).
Works better with: inventory data (to exclude out-of-stock items).

Intent Extraction

Before recommending, extract these signals from the conversation:

Signal	Examples	How to Extract
Budget	"500以内", "¥200左右", "不超过1000"	Parse number + direction
Recipient

If critical signals are missing (especially budget), ask one clarifying question.
Never ask for all missing fields at once.

Reference: intent-extraction.md

Filtering Logic

Apply filters in this order (hard → soft):

1. Hard filters (eliminate if not met):

- Budget: price ≤ budget_max (or sale_price if active) - Hard constraints: attribute must match (e.g., "纯棉" → filter by material tag) - Stock: exclude if stock_qty == 0 (when inventory data available)

2. Soft scoring (rank what remains):

- Recipient match: suitable_for overlap with recipient description - Occasion match: tags overlap with occasion keywords - Style/preference match: description + tags keyword overlap - Popularity signal: use sales_rank if available, else recency

3. Return top N (default: 3, configurable via max_recommendations)

Reference: filtering-logic.md

Recommendation Presentation

Standard format (3 recommendations)

CODEBLOCK0

Gift recommendation (add wrapping note)

CODEBLOCK1

Upsell (when appropriate)

If the customer's budget allows 20% more for a meaningfully better option:

"还有一款 ¥[price+] 的[产品名]，多了[key upgrade]，性价比也很高，要不要看看？"

Only suggest once per conversation. Never push if customer declines.

Special Flows

"帮我比较" (Comparison)

When customer names 2+ specific products:

- Fetch both from KB
Build a comparison table: price / key specs / suitable for / verdict
Give a clear recommendation, not just data

"搭配什么" (Outfit/Pairing)

When customer asks what goes with a product:

- Identify the anchor product
Filter KB for complementary items (matching category tags: "搭配", "配套")
Present as a complete set with total price

"再便宜一点" (Price objection)

When customer asks for cheaper options after seeing recommendations:

- Re-filter with lower budget
If nothing cheaper: explain value at current price, don't apologize for price

"没有我想要的" (No match)

When no product passes the filters:

1. Tell the customer honestly
Suggest the closest available option
Offer to notify when matching product arrives (log as feature request)

Script

Use scripts/recommend.py for deterministic filtering and scoring.
Reference: filtering-logic.md

产品推荐器

概述

本技能处理所有帮我选择类查询。它不仅仅是列出产品——它能理解客户的具体情况，智能筛选，并呈现精选推荐列表及推荐理由。

依赖项： 知识库中的products[]（步骤03）。
配合使用效果更佳： 库存数据（用于排除缺货商品）。

意图提取

在推荐之前，从对话中提取以下信号：

信号	示例	提取方式
预算	500以内, ¥200左右, 不超过1000	解析数字+方向
收礼人

如果缺少关键信号（尤其是预算），请提出一个澄清性问题。切勿一次性询问所有缺失信息。

参考： intent-extraction.md

筛选逻辑

按以下顺序应用筛选条件（硬性→软性）：

1. 硬性筛选（不满足则排除）：

- 预算：price ≤ budgetmax（如有促销价则用saleprice） - 硬性限制：属性必须匹配（例如纯棉→按材质标签筛选） - 库存：若stock_qty == 0则排除（有库存数据时）

2. 软性评分（对剩余商品排序）：

- 收礼人匹配：suitable_for与收礼人描述的重叠度 - 场合匹配：tags与场合关键词的重叠度 - 风格/偏好匹配：description+tags关键词重叠度 - 热度信号：有sales_rank则用，否则按更新时间

3. 返回前N个（默认：3，可通过max_recommendations配置）

参考： filtering-logic.md

特殊流程

帮我比较（比较）

当客户提到2个以上具体产品时：

- 从知识库获取所有产品信息
构建对比表格：价格/关键规格/适用人群/结论
给出明确推荐，而非仅提供数据

搭配什么（搭配/组合）

当客户询问某产品的搭配时：

- 确定核心产品
从知识库筛选互补商品（匹配品类标签：搭配, 配套）
以套装形式呈现，包含总价

再便宜一点（价格异议）

当客户看到推荐后要求更便宜选项时：

- 用更低预算重新筛选
若无更便宜选项：解释当前价格的价值，不要为价格道歉

没有我想要的（无匹配）

当没有产品通过筛选时：

1. 如实告知客户
推荐最接近的可用选项
提供到货通知（记录为功能需求）

脚本

使用scripts/recommend.py进行确定性筛选和评分。
参考： filtering-logic.md

product-recommender产品推荐器