some changes

app.py

@@ -119,6 +119,94 @@ def _parse_recommendation_json(raw: str):
     except Exception:
         return None
 
+
+def _build_recommendation_sections(rec_data, candidate_lookup):
+    if not rec_data:
+        return None, None
+
+    recommend_flag = rec_data.get("recommend")
+    if isinstance(recommend_flag, str):
+        recommend_flag = recommend_flag.strip().lower() in {"yes", "true", "1"}
+    elif isinstance(recommend_flag, (int, float)):
+        recommend_flag = bool(recommend_flag)
+
+    if not recommend_flag:
+        return None, None
+
+    recommendations = rec_data.get("recommendations", [])
+    if not isinstance(recommendations, list):
+        return None, None
+
+    lines = ["### Suggested Products", ""]
+    products_payload = []
+
+    for idx, item in enumerate(recommendations[:3], start=1):
+        if not isinstance(item, dict):
+            continue
+        raw_id = item.get("id")
+        if raw_id is None:
+            continue
+        pid = str(raw_id).strip()
+        if not pid:
+            continue
+
+        candidate = candidate_lookup.get(pid, {})
+
+        brand = (
+            candidate.get("brand_en")
+            or candidate.get("brand_zh")
+            or item.get("brand")
+            or ""
+        )
+        name = (
+            candidate.get("product_name_en")
+            or candidate.get("product_name_zh")
+            or item.get("name")
+            or f"Product {idx}"
+        )
+        category = (
+            candidate.get("category_en")
+            or candidate.get("category_zh")
+            or item.get("category")
+            or None
+        )
+        price_value = candidate.get("price_value")
+        price_currency = candidate.get("price_currency")
+        why = item.get("why") or "Supports the user’s concern."
+        how = item.get("how") or "Use as directed on the product label."
+        url = candidate.get("source_url") or item.get("url")
+        image_url = candidate.get("image_url") or item.get("image_url")
+
+        lines.extend([
+            f"{idx}. **{name}**",
+            f"- **Why it helps:** {why}",
+            f"- **How to use:** {how}",
+            "",
+        ])
+
+        products_payload.append({
+            "id": pid,
+            "brand": brand,
+            "name": name,
+            "category": category,
+            "price_value": price_value,
+            "price_currency": price_currency,
+            "why": why,
+            "how": how,
+            "url": url,
+            "image_url": image_url,
+        })
+
+    if not products_payload:
+        return None, None
+
+    suggestion_text = "\n".join(lines).strip()
+    product_json_payload = json.dumps(
+        {"version": 1, "products": products_payload},
+        ensure_ascii=False,
+    )
+    return suggestion_text, product_json_payload
+
 # ---- Inference on GPU (ZeroGPU pattern) ----
 @spaces.GPU(duration=120)
 def generate_answer(image, question, temperature=0.7, top_p=0.95, max_tokens=256):
@@ -183,7 +271,10 @@ def pet_answer_with_recs(image, question, temperature=0.7, top_p=0.95, max_token
 
     # Step 2: retrieve product candidates (humans/skincare; model will decide relevance)
     cands = product_search(question, k=8)
-    cand_block = format_candidates_for_llm(cands, budget_twd=budget_twd)
+    cand_block_json, cand_list = format_candidates_for_llm(cands, budget_twd=budget_twd)
+    candidate_lookup = {
+        str(c.get("id")).strip(): c for c in cand_list if c.get("id") is not None
+    }
 
     # Step 3: build a small, text-only prompt for suggestions
     # IMPORTANT: we use the same Qwen2.5-VL model in text mode
@@ -192,22 +283,16 @@ def pet_answer_with_recs(image, question, temperature=0.7, top_p=0.95, max_token
         "content": [
             {"type": "text", "text":
              "You are DermalCare's assistant.\n"
-             "You will receive (1) a user question and (2) a JSON array named candidate_products containing skincare products.\n"
-             "If one or more products genuinely help the user, respond with: \n"
-             "### Suggested Products\n"
-             "1. **Product name**\n- **Why it helps:** ...\n- **How to use:** ...\n"
-             "(repeat up to 3 items)\n"
-             "\n<DERMACARE_PRODUCTS_JSON>{\"version\": 1, \"products\": [...]}</DERMACARE_PRODUCTS_JSON>\n"
-             "The JSON must list exactly the recommended products with keys: id, brand, name, category, price_value, price_currency, why, how, url, image_url.\n"
-             "Use values from candidate_products: use brand_en/brand_zh to compose 'brand', prefer English names but fall back to Chinese when absent.\n"
-             "If a field is missing, set it to null. Do NOT invent ids or products outside the provided list.\n"
-             "If no products are relevant, output exactly 'No relevant products.' with nothing else."}
+             "Respond ONLY with valid JSON (no markdown, no explanations).\n"
+             "Expected schema: {\"recommend\": bool, \"recommendations\": [ {\"id\": str, \"why\": str, \"how\": str } ], \"notes\": str }.\n"
+             "Use candidate_products as the exclusive source of items. If a product is recommended, its id must exist in candidate_products.\n"
+             "If no products are relevant, return {\"recommend\": false, \"recommendations\": [], \"notes\": \"No relevant products.\"}."}
         ]
     },{
         "role": "user",
         "content": [
             {"type": "text", "text": f"User message:\n{question}"},
-            {"type": "text", "text": f"candidate_products = {cand_block}"}
+            {"type": "text", "text": f"candidate_products = {cand_block_json}"}
         ]
     }]