Product Recommender

Intelligent product recommendation engine for retail digital employees. Recommends products based on customer needs, budget, recipient, occasion, preferences...

Security Scan

The skill's code, instructions, and requirements are coherent with a product recommendation engine and do not request unrelated credentials or perform obvious network/exfiltration actions — but the provided script output was truncated in the submission, so review of the file footer is recommended.

Purpose & Capability

Name/description (product recommendation) match the included SKILL.md and the recommend.py logic: intent extraction, budget/constraint filtering, scoring, upsell logic and presentation. Required resources (a products[] knowledge base) are consistent with the stated purpose; no unrelated credentials or binaries are requested.

Instruction Scope

SKILL.md stays within the recommender domain: it instructs extracting intent signals, running scripts/recommend.py, and returning 3 curated items. It does mention logging 'feature_request' and session state (e.g., upsell_declined) which are not fully implemented in the visible code — minor scope mismatch but not evidence of malicious behavior. The runtime instructions do require a knowledge_base.json path; ensure that path only points to intended product data (the script will read any file given).

Install Mechanism

No install spec (instruction-only with an included script). That is low-risk: nothing is downloaded or installed automatically by the skill.

Credentials

The skill requests no environment variables, credentials, or config paths. The code likewise does not reference secrets or external tokens in the visible portion.

Persistence & Privilege

always is false and the skill does not request persistent system-wide privileges. The script appears to be a transient CLI utility that reads a KB and returns recommendations; it does not modify other skills or system settings in the visible code.

Assessment

This skill appears coherent for recommending products: it runs a local Python script against a provided knowledge_base.json and uses deterministic filtering/scoring. Before installing/using: 1) Review the complete scripts/recommend.py file (the submission truncated the file end) to confirm the main() function does not perform unexpected I/O or network calls. 2) Ensure the agent will be given only trusted knowledge_base JSON files (don't point --kb at sensitive local files). 3) If you rely on inventory/live API behavior, check how 'stock_status: live_api' is handled elsewhere — the skill assumes live availability if that flag is present. 4) Confirm how 'feature_request' logging is implemented in your environment (the docs mention logging but the visible code doesn't show where it goes). If you want a higher assurance, ask for the full recommend.py file (complete) and any runtime wrapper the agent uses to invoke it so you can verify there are no hidden network endpoints or credential usage.

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	"送妈妈", "给男朋友", "自用"	Named or implied
Occasion	"生日", "面试", "日常穿", "夏天用"	Event or context
Preferences	"素色", "轻便", "不要太甜", "简约风"	Style/attribute keywords
Age/Gender	"30岁女性", "老年人", "男生"	Demographic
Constraints	"不含酒精", "纯棉", "防水"	Hard requirements
Quantity	"买一套", "各来一个"	Number intent

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)

为您推荐 3 款最适合的选择：

1️⃣ [产品名] ¥[price]
   [1句话说明为什么适合这个场景/人群]
   [关键亮点：1-2个最相关的属性]

2️⃣ [产品名] ¥[price]
   [...]

3️⃣ [产品名] ¥[price]
   [...]

[可选] 您更倾向哪款？我可以帮您查一下库存～

Gift recommendation (add wrapping note)

送礼推荐：[产品名] ¥[price]
[为什么适合作为礼物 — 1句话]
[礼盒包装是否可用 if known]

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
2. Suggest the closest available option
3. 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