Spaces:
Sleeping
Sleeping
change to model bge visual
Browse files
app.py
CHANGED
@@ -1,7 +1,8 @@
|
|
1 |
import gradio as gr
|
2 |
import time
|
3 |
from datetime import datetime
|
4 |
-
from
|
|
|
5 |
from qdrant_client import QdrantClient
|
6 |
from qdrant_client.models import Filter, FieldCondition, MatchValue
|
7 |
import os
|
@@ -21,86 +22,91 @@ qdrant_client = QdrantClient(
|
|
21 |
# Airtable Config
|
22 |
AIRTABLE_API_KEY = os.environ.get("airtable_api")
|
23 |
BASE_ID = os.environ.get("airtable_baseid")
|
24 |
-
TABLE_NAME = "Feedback_search"
|
25 |
-
api = Api(AIRTABLE_API_KEY)
|
26 |
-
table = api.table(BASE_ID, TABLE_NAME)
|
27 |
|
28 |
# Preload Models
|
29 |
-
|
30 |
-
|
31 |
-
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
# Utils
|
34 |
-
def is_non_thai(text):
|
35 |
return re.match(r'^[A-Za-z0-9&\-\s]+$', text) is not None
|
36 |
|
37 |
def normalize(text: str) -> str:
|
38 |
-
if is_non_thai(text):
|
39 |
return text.strip()
|
40 |
-
text = unicodedata.normalize("NFC", text)
|
41 |
-
return text.replace("เแ", "แ").replace("เเ", "แ").strip().lower()
|
42 |
|
43 |
# Global state
|
44 |
-
latest_query_result = {"query": "", "result": "", "raw_query": "", "time": ""}
|
45 |
|
46 |
# Search Function
|
47 |
def search_product(query):
|
48 |
-
yield gr.update(value="🔄 กำลังค้นหา..."), ""
|
49 |
|
50 |
-
start_time = time.time()
|
51 |
-
latest_query_result["raw_query"] = query
|
52 |
|
53 |
-
corrected_query = normalize(query)
|
54 |
-
query_embed = model.encode(corrected_query)
|
55 |
|
56 |
try:
|
|
|
57 |
result = qdrant_client.query_points(
|
58 |
-
collection_name=collection_name,
|
59 |
-
query=query_embed.tolist(),
|
60 |
-
with_payload=True,
|
61 |
-
|
62 |
-
limit=50
|
63 |
).points
|
64 |
except Exception as e:
|
65 |
-
yield gr.update(value="❌ Qdrant error"), f"<p>❌ Qdrant error: {str(e)}</p>"
|
66 |
return
|
67 |
|
68 |
if len(result) > 0:
|
69 |
topk = 50 # ดึงมา rerank แค่ 50 อันดับแรกจาก Qdrant
|
70 |
result = result[:topk]
|
71 |
|
72 |
-
scored = []
|
73 |
for r in result:
|
74 |
-
name = str(r.payload.get("name", "")).lower()
|
75 |
-
brand = str(r.payload.get("brand", "")).lower()
|
76 |
-
query_lower = corrected_query.lower()
|
77 |
|
78 |
# ถ้า query สั้นเกินไป ให้ fuzzy_score = 0 เพื่อกันเพี้ยน
|
79 |
if len(corrected_query) >= 3 and name:
|
80 |
-
fuzzy_name_score = fuzz.partial_ratio(query_lower, name) / 100.0
|
81 |
-
fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0
|
82 |
else:
|
83 |
fuzzy_name_score = 0.0
|
84 |
fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0
|
85 |
|
86 |
# รวม hybrid score
|
87 |
if fuzzy_name_score < 0.5:
|
88 |
-
hybrid_score = r.score
|
89 |
else:
|
90 |
-
hybrid_score = 0.7 * r.score + 0.3 * fuzzy_name_score
|
91 |
if fuzzy_brand_score >= 0.8:
|
92 |
-
hybrid_score = hybrid_score*1.2
|
93 |
r.payload["score"] = hybrid_score # เก็บลง payload ใช้เทียบ treshold ตอนเเสดงผล
|
94 |
r.payload["fuzzy_name_score"] = fuzzy_name_score # เก็บไว้เผื่อ debug
|
95 |
r.payload["fuzzy_brand_score"] = fuzzy_brand_score # เก็บไว้เผื่อ debug
|
96 |
r.payload['semantic_score'] = r.score # เก็บไว้เผื่อ debug
|
97 |
-
scored.append((r, hybrid_score))
|
98 |
|
99 |
# เรียงตาม hybrid score แล้วกรองผลลัพธ์ที่ hybrid score ต่ำเกิน
|
100 |
-
scored = sorted(scored, key=lambda x: x[1], reverse=True)
|
101 |
-
result = [r[0] for r in scored]
|
102 |
|
103 |
-
elapsed = time.time() - start_time
|
104 |
html_output = f"<p>⏱ <strong>{elapsed:.2f} วินาที</strong></p>"
|
105 |
if corrected_query != query:
|
106 |
html_output += f"<p>🔧 แก้คำค้นจาก: <code>{query}</code> → <code>{corrected_query}</code></p>"
|
@@ -108,11 +114,11 @@ def search_product(query):
|
|
108 |
result_summary, found = "", False
|
109 |
|
110 |
for res in result:
|
111 |
-
if res.payload["score"] >= threshold:
|
112 |
-
found = True
|
113 |
name = res.payload.get("name", "ไม่ทราบชื่อสินค้า")
|
114 |
score = f"{res.payload['score']:.4f}"
|
115 |
-
img_url = res.payload.get("
|
116 |
price = res.payload.get("price", "ไม่ระบุ")
|
117 |
brand = res.payload.get("brand", "")
|
118 |
|
@@ -146,6 +152,8 @@ def search_product(query):
|
|
146 |
def log_feedback(feedback):
|
147 |
try:
|
148 |
now = datetime.now().strftime("%Y-%m-%d")
|
|
|
|
|
149 |
table.create({
|
150 |
"model": "BGE M3",
|
151 |
"timestamp": now,
|
|
|
1 |
import gradio as gr
|
2 |
import time
|
3 |
from datetime import datetime
|
4 |
+
from visual_bge.modeling import Visualized_BGE
|
5 |
+
from huggingface_hub import hf_hub_download
|
6 |
from qdrant_client import QdrantClient
|
7 |
from qdrant_client.models import Filter, FieldCondition, MatchValue
|
8 |
import os
|
|
|
22 |
# Airtable Config
|
23 |
AIRTABLE_API_KEY = os.environ.get("airtable_api")
|
24 |
BASE_ID = os.environ.get("airtable_baseid")
|
25 |
+
TABLE_NAME = "Feedback_search" # use table name
|
26 |
+
api = Api(AIRTABLE_API_KEY) # api to airtable
|
27 |
+
table = api.table(BASE_ID, TABLE_NAME) # choose table
|
28 |
|
29 |
# Preload Models
|
30 |
+
model_weight = hf_hub_download(repo_id="BAAI/bge-visualized", filename="Visualized_m3.pth")
|
31 |
+
# Load model
|
32 |
+
model = Visualized_BGE(
|
33 |
+
model_name_bge="BAAI/bge-m3",
|
34 |
+
model_weight=model_weight
|
35 |
+
)
|
36 |
+
collection_name = "product_visual_bge" # setup collection name in qdrant
|
37 |
+
threshold = 0.5 # threshold use when rerank
|
38 |
|
39 |
# Utils
|
40 |
+
def is_non_thai(text): # check if english retune true
|
41 |
return re.match(r'^[A-Za-z0-9&\-\s]+$', text) is not None
|
42 |
|
43 |
def normalize(text: str) -> str:
|
44 |
+
if is_non_thai(text): # send text to check english
|
45 |
return text.strip()
|
46 |
+
text = unicodedata.normalize("NFC", text) # change text to unicode
|
47 |
+
return text.replace("เแ", "แ").replace("เเ", "แ").strip().lower() # เเก้กรณีกด เ หลายที
|
48 |
|
49 |
# Global state
|
50 |
+
latest_query_result = {"query": "", "result": "", "raw_query": "", "time": ""} # create for send to airtable
|
51 |
|
52 |
# Search Function
|
53 |
def search_product(query):
|
54 |
+
yield gr.update(value="🔄 กำลังค้นหา..."), "" # when user search
|
55 |
|
56 |
+
start_time = time.time() # start timer
|
57 |
+
latest_query_result["raw_query"] = query # collect user qeary
|
58 |
|
59 |
+
corrected_query = normalize(query) # change query to normalize query
|
60 |
+
query_embed = model.encode(text=corrected_query)[0] # embed corrected_query to vector
|
61 |
|
62 |
try:
|
63 |
+
#use qdrant search
|
64 |
result = qdrant_client.query_points(
|
65 |
+
collection_name=collection_name, # choose collection in qdrant
|
66 |
+
query=query_embed.tolist(), # vector query
|
67 |
+
with_payload=True, # see payload
|
68 |
+
limit=50 # need 50 product
|
|
|
69 |
).points
|
70 |
except Exception as e:
|
71 |
+
yield gr.update(value="❌ Qdrant error"), f"<p>❌ Qdrant error: {str(e)}</p>" # have problem when search
|
72 |
return
|
73 |
|
74 |
if len(result) > 0:
|
75 |
topk = 50 # ดึงมา rerank แค่ 50 อันดับแรกจาก Qdrant
|
76 |
result = result[:topk]
|
77 |
|
78 |
+
scored = [] # use to collect product and score
|
79 |
for r in result:
|
80 |
+
name = str(r.payload.get("name", "")).lower() # get name in payload and lowercase
|
81 |
+
brand = str(r.payload.get("brand", "")).lower() # get brand in payload and lowercase
|
82 |
+
query_lower = corrected_query.lower() # lowercase corected_quey
|
83 |
|
84 |
# ถ้า query สั้นเกินไป ให้ fuzzy_score = 0 เพื่อกันเพี้ยน
|
85 |
if len(corrected_query) >= 3 and name:
|
86 |
+
fuzzy_name_score = fuzz.partial_ratio(query_lower, name) / 100.0 # query compare name score
|
87 |
+
fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0 # query compare brand score
|
88 |
else:
|
89 |
fuzzy_name_score = 0.0
|
90 |
fuzzy_brand_score = fuzz.partial_ratio(query_lower, brand) / 100.0
|
91 |
|
92 |
# รวม hybrid score
|
93 |
if fuzzy_name_score < 0.5:
|
94 |
+
hybrid_score = r.score # not change qdrant score
|
95 |
else:
|
96 |
+
hybrid_score = 0.7 * r.score + 0.3 * fuzzy_name_score # use qdrant score 70% and fuzzy name score 30%
|
97 |
if fuzzy_brand_score >= 0.8:
|
98 |
+
hybrid_score = hybrid_score*1.2 # มั่นใจว่าถูกเเบรนด์ เพิ่ม score 120%
|
99 |
r.payload["score"] = hybrid_score # เก็บลง payload ใช้เทียบ treshold ตอนเเสดงผล
|
100 |
r.payload["fuzzy_name_score"] = fuzzy_name_score # เก็บไว้เผื่อ debug
|
101 |
r.payload["fuzzy_brand_score"] = fuzzy_brand_score # เก็บไว้เผื่อ debug
|
102 |
r.payload['semantic_score'] = r.score # เก็บไว้เผื่อ debug
|
103 |
+
scored.append((r, hybrid_score)) # collect product and hybrid score
|
104 |
|
105 |
# เรียงตาม hybrid score แล้วกรองผลลัพธ์ที่ hybrid score ต่ำเกิน
|
106 |
+
scored = sorted(scored, key=lambda x: x[1], reverse=True) # sort
|
107 |
+
result = [r[0] for r in scored] # collect new sort product
|
108 |
|
109 |
+
elapsed = time.time() - start_time # stop search time
|
110 |
html_output = f"<p>⏱ <strong>{elapsed:.2f} วินาที</strong></p>"
|
111 |
if corrected_query != query:
|
112 |
html_output += f"<p>🔧 แก้คำค้นจาก: <code>{query}</code> → <code>{corrected_query}</code></p>"
|
|
|
114 |
result_summary, found = "", False
|
115 |
|
116 |
for res in result:
|
117 |
+
if res.payload["score"] >= threshold: # choose only product score more than threshold
|
118 |
+
found = True # find product
|
119 |
name = res.payload.get("name", "ไม่ทราบชื่อสินค้า")
|
120 |
score = f"{res.payload['score']:.4f}"
|
121 |
+
img_url = res.payload.get("image_url", "")
|
122 |
price = res.payload.get("price", "ไม่ระบุ")
|
123 |
brand = res.payload.get("brand", "")
|
124 |
|
|
|
152 |
def log_feedback(feedback):
|
153 |
try:
|
154 |
now = datetime.now().strftime("%Y-%m-%d")
|
155 |
+
# create table for send to airtable
|
156 |
+
# คอลัมน์ต้องตรงกับบน airtable
|
157 |
table.create({
|
158 |
"model": "BGE M3",
|
159 |
"timestamp": now,
|