change to model bge visual

app.py

@@ -1,7 +1,8 @@
 import gradio as gr
 import time
 from datetime import datetime
-from sentence_transformers import SentenceTransformer
+from visual_bge.modeling import Visualized_BGE
+from huggingface_hub import hf_hub_download
 from qdrant_client import QdrantClient
 from qdrant_client.models import Filter, FieldCondition, MatchValue
 import os
@@ -21,86 +22,91 @@ qdrant_client = QdrantClient(
 # Airtable Config
 AIRTABLE_API_KEY = os.environ.get("airtable_api")
 BASE_ID = os.environ.get("airtable_baseid")
-TABLE_NAME = "Feedback_search"
-api = Api(AIRTABLE_API_KEY)
-table = api.table(BASE_ID, TABLE_NAME)
+TABLE_NAME = "Feedback_search" # use table name
+api = Api(AIRTABLE_API_KEY) # api to airtable
+table = api.table(BASE_ID, TABLE_NAME) # choose table
 
 # Preload Models
-model = SentenceTransformer("BAAI/bge-m3")
-collection_name = "product_bge-m3"
-threshold = 0.5
+model_weight = hf_hub_download(repo_id="BAAI/bge-visualized", filename="Visualized_m3.pth")
+# Load model
+model = Visualized_BGE(
+    model_name_bge="BAAI/bge-m3",
+    model_weight=model_weight
+)
+collection_name = "product_visual_bge" # setup collection name in qdrant
+threshold = 0.5 # threshold use when rerank
 
 # Utils
-def is_non_thai(text):
+def is_non_thai(text): # check if english retune true
     return re.match(r'^[A-Za-z0-9&\-\s]+$', text) is not None
 
 def normalize(text: str) -> str:
-    if is_non_thai(text):
+    if is_non_thai(text): # send text to check english
         return text.strip()
-    text = unicodedata.normalize("NFC", text)
-    return text.replace("เแ", "แ").replace("เเ", "แ").strip().lower()
+    text = unicodedata.normalize("NFC", text) # change text to unicode
+    return text.replace("เแ", "แ").replace("เเ", "แ").strip().lower() # เเก้กรณีกด เ หลายที
 
 # Global state
-latest_query_result = {"query": "", "result": "", "raw_query": "", "time": ""}
+latest_query_result = {"query": "", "result": "", "raw_query": "", "time": ""} # create for send to airtable
 
 # Search Function
 def search_product(query):
-    yield gr.update(value="🔄 กำลังค้นหา..."), ""
+    yield gr.update(value="🔄 กำลังค้นหา..."), "" # when user search
 
-    start_time = time.time()
-    latest_query_result["raw_query"] = query
+    start_time = time.time() # start timer
+    latest_query_result["raw_query"] = query # collect user qeary
 
-    corrected_query = normalize(query)
-    query_embed = model.encode(corrected_query)
+    corrected_query = normalize(query) # change query to normalize query
+    query_embed = model.encode(text=corrected_query)[0] # embed corrected_query to vector
 
     try:
+        #use qdrant search
         result = qdrant_client.query_points(
-            collection_name=collection_name,
-            query=query_embed.tolist(),
-            with_payload=True,
-            query_filter=Filter(must=[FieldCondition(key="type", match=MatchValue(value="product"))]),
-            limit=50
+            collection_name=collection_name, # choose collection in qdrant
+            query=query_embed.tolist(), # vector query
+            with_payload=True, # see payload
+            limit=50 # need 50 product
         ).points
     except Exception as e:
-        yield gr.update(value="❌ Qdrant error"), f"<p>❌ Qdrant error: {str(e)}</p>"
+        yield gr.update(value="❌ Qdrant error"), f"<p>❌ Qdrant error: {str(e)}</p>" # have problem when search
         return
 
     if len(result) > 0:
       topk = 50  # ดึงมา rerank แค่ 50 อันดับแรกจาก Qdrant
       result = result[:topk]
 
-      scored = []
+      scored = [] # use to collect product and score
       for r in result:
-          name = str(r.payload.get("name", "")).lower()
-          brand = str(r.payload.get("brand", "")).lower()
-          query_lower = corrected_query.lower()
+          name = str(r.payload.get("name", "")).lower() # get name in payload and lowercase
+          brand = str(r.payload.get("brand", "")).lower() # get brand in payload and lowercase
+          query_lower = corrected_query.lower() # lowercase corected_quey
 
           # ถ้า query สั้นเกินไป ให้ fuzzy_score = 0 เพื่อกันเพี้ยน
           if len(corrected_query) >= 3 and name:
-            fuzzy_name_score = fuzz.partial_ratio(query_lower, name) / 100.0
-            fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0
+            fuzzy_name_score = fuzz.partial_ratio(query_lower, name) / 100.0 # query compare name score
+            fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0 # query compare brand score
           else:
             fuzzy_name_score = 0.0
             fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0
 
           # รวม hybrid score
           if fuzzy_name_score < 0.5:
-            hybrid_score = r.score
+            hybrid_score = r.score # not change qdrant score
           else:
-            hybrid_score = 0.7 * r.score + 0.3 * fuzzy_name_score
+            hybrid_score = 0.7 * r.score + 0.3 * fuzzy_name_score # use qdrant score 70% and fuzzy name score 30%
           if fuzzy_brand_score >= 0.8:
-            hybrid_score = hybrid_score*1.2
+            hybrid_score = hybrid_score*1.2 # มั่นใจว่าถูกเเบรนด์ เพิ่ม score 120%
           r.payload["score"] = hybrid_score  # เก็บลง payload ใช้เทียบ treshold ตอนเเสดงผล
           r.payload["fuzzy_name_score"] = fuzzy_name_score # เก็บไว้เผื่อ debug
           r.payload["fuzzy_brand_score"] = fuzzy_brand_score # เก็บไว้เผื่อ debug
           r.payload['semantic_score'] = r.score # เก็บไว้เผื่อ debug
-          scored.append((r, hybrid_score))
+          scored.append((r, hybrid_score)) # collect product and hybrid score
 
       # เรียงตาม hybrid score แล้วกรองผลลัพธ์ที่ hybrid score ต่ำเกิน
-      scored = sorted(scored, key=lambda x: x[1], reverse=True)
-      result = [r[0] for r in scored]
+      scored = sorted(scored, key=lambda x: x[1], reverse=True) # sort
+      result = [r[0] for r in scored] # collect new sort product
 
-    elapsed = time.time() - start_time
+    elapsed = time.time() - start_time # stop search time
     html_output = f"<p>⏱ <strong>{elapsed:.2f} วินาที</strong></p>"
     if corrected_query != query:
         html_output += f"<p>🔧 แก้คำค้นจาก: <code>{query}</code> → <code>{corrected_query}</code></p>"
@@ -108,11 +114,11 @@ def search_product(query):
     result_summary, found = "", False
 
     for res in result:
-        if res.payload["score"] >= threshold:
-            found = True
+        if res.payload["score"] >= threshold: # choose only product score more than threshold
+            found = True # find product
             name = res.payload.get("name", "ไม่ทราบชื่อสินค้า")
             score = f"{res.payload['score']:.4f}"
-            img_url = res.payload.get("imageUrl", "")
+            img_url = res.payload.get("image_url", "")
             price = res.payload.get("price", "ไม่ระบุ")
             brand = res.payload.get("brand", "")
 
@@ -146,6 +152,8 @@ def search_product(query):
 def log_feedback(feedback):
     try:
         now = datetime.now().strftime("%Y-%m-%d")
+        # create table for send to airtable
+        # คอลัมน์ต้องตรงกับบน airtable
         table.create({
             "model": "BGE M3",
             "timestamp": now,