Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
|
@@ -24,18 +24,17 @@ firebase_admin.initialize_app(cred, {'databaseURL': database_url})
|
|
| 24 |
ref = db.reference('/')
|
| 25 |
print("β
Firebase connected")
|
| 26 |
|
| 27 |
-
# ---------- Sentence-Transformer Model
|
| 28 |
print("π Loading sentence-transformers...")
|
| 29 |
embedding_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
|
| 30 |
print("β
Embedding model loaded")
|
| 31 |
|
| 32 |
-
# ---------- Local GPT-2 Model (Fallback) ----------
|
| 33 |
-
print("π Loading local GPT-2 model
|
| 34 |
try:
|
| 35 |
gpt2_tokenizer = AutoTokenizer.from_pretrained("gpt2")
|
| 36 |
gpt2_model = AutoModelForCausalLM.from_pretrained("gpt2")
|
| 37 |
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token
|
| 38 |
-
# Move to GPU if available
|
| 39 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 40 |
gpt2_model = gpt2_model.to(device)
|
| 41 |
print(f"β
GPT-2 loaded on {device}")
|
|
@@ -44,80 +43,107 @@ except Exception as e:
|
|
| 44 |
gpt2_model = None
|
| 45 |
gpt2_tokenizer = None
|
| 46 |
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
"product photography": "High-end product photography for jewelry, cars, luxury goods.",
|
| 50 |
-
"architecture renders": "3D architectural visualization and rendering.",
|
| 51 |
-
"corporate videos": "4K/8K corporate video production.",
|
| 52 |
-
"drone cinematography": "Drone videography for events, real estate, commercials.",
|
| 53 |
-
"3d rendering": "3D rendering for products and architecture.",
|
| 54 |
-
"wedding photography": "Wedding photography & cinematography packages.",
|
| 55 |
-
"real estate visuals": "Real estate photo, virtual tours, drone shoots.",
|
| 56 |
-
"ai editing": "AI-enhanced photo editing and retouching.",
|
| 57 |
-
"headshots": "Professional corporate headshots and portraits."
|
| 58 |
-
}
|
| 59 |
-
|
| 60 |
-
SERVICES_LIST = list(SERVICES_DESC.keys())
|
| 61 |
-
|
| 62 |
-
# ---------- Intent Classification ----------
|
| 63 |
-
IRRELEVANT_KEYWORDS = ["cricket", "weather", "politics", "imran", "nawaz", "election", "stock", "bitcoin", "covid", "corona"]
|
| 64 |
-
|
| 65 |
-
def classify_intent(user_text):
|
| 66 |
-
text_lower = user_text.lower()
|
| 67 |
-
if any(kw in text_lower for kw in IRRELEVANT_KEYWORDS):
|
| 68 |
-
return "irrelevant"
|
| 69 |
-
if any(x in text_lower for x in ["contact", "phone", "number", "whatsapp", "email", "call", "raabta"]):
|
| 70 |
-
return "contact"
|
| 71 |
-
if any(x in text_lower for x in ["price", "cost", "rate", "kitne", "quote", "charges", "fee"]):
|
| 72 |
-
return "pricing"
|
| 73 |
-
if any(x in text_lower for x in ["portfolio", "sample", "gallery", "kaam dikhao"]):
|
| 74 |
-
return "portfolio"
|
| 75 |
-
if any(x in text_lower for x in ["salam", "hello", "hi", "hey"]):
|
| 76 |
-
return "greeting"
|
| 77 |
-
return "service_inquiry"
|
| 78 |
-
|
| 79 |
-
def handle_negation(text):
|
| 80 |
-
neg_patterns = [r"can(not|'t) you", r"do you not", r"don't you", r"nhi karte", r"nahi karte"]
|
| 81 |
-
return any(re.search(pat, text.lower()) for pat in neg_patterns)
|
| 82 |
-
|
| 83 |
-
# ---------- Local GPT-2 Generation (Personalized, No Copy-Paste) ----------
|
| 84 |
-
def generate_with_gpt2(prompt):
|
| 85 |
-
"""Generate a response using local GPT-2 model."""
|
| 86 |
if gpt2_model is None or gpt2_tokenizer is None:
|
| 87 |
-
return
|
| 88 |
|
| 89 |
-
#
|
| 90 |
-
|
| 91 |
-
|
| 92 |
|
| 93 |
-
|
|
|
|
|
|
|
| 94 |
Assistant:"""
|
| 95 |
|
| 96 |
-
inputs = gpt2_tokenizer(
|
| 97 |
device = next(gpt2_model.parameters()).device
|
| 98 |
inputs = {k: v.to(device) for k, v in inputs.items()}
|
| 99 |
|
| 100 |
with torch.no_grad():
|
| 101 |
outputs = gpt2_model.generate(
|
| 102 |
**inputs,
|
| 103 |
-
max_new_tokens=
|
| 104 |
-
temperature=0.
|
| 105 |
do_sample=True,
|
| 106 |
top_p=0.9,
|
|
|
|
|
|
|
| 107 |
pad_token_id=gpt2_tokenizer.eos_token_id
|
| 108 |
)
|
| 109 |
response = gpt2_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
|
| 110 |
# Extract only the assistant part
|
| 111 |
if "Assistant:" in response:
|
| 112 |
response = response.split("Assistant:")[-1].strip()
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
#
|
| 117 |
-
|
| 118 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
return response
|
| 120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
# ---------- OpenRouter Setup ----------
|
| 122 |
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
|
| 123 |
MODEL = "google/gemini-2.0-flash-lite-preview-02-05"
|
|
@@ -127,8 +153,8 @@ SYSTEM_PROMPT = """You are "VisionCraft AI" β assistant for "Elite Visuals", a
|
|
| 127 |
Services: product photography, architecture renders, corporate videos, drone cinematography, 3D rendering, wedding photography, real estate visuals, AI editing.
|
| 128 |
|
| 129 |
Always answer in bullet points. Be concise, helpful, and personalized to the user's exact question.
|
| 130 |
-
If user asks "can you not provide X", clarify that you DO provide it
|
| 131 |
-
Never
|
| 132 |
|
| 133 |
def call_openrouter(prompt):
|
| 134 |
if not OPENROUTER_API_KEY:
|
|
@@ -142,7 +168,7 @@ def call_openrouter(prompt):
|
|
| 142 |
{"role": "user", "content": prompt}
|
| 143 |
],
|
| 144 |
"temperature": 0.7,
|
| 145 |
-
"max_tokens":
|
| 146 |
}
|
| 147 |
try:
|
| 148 |
resp = requests.post(url, json=payload, headers=headers, timeout=15)
|
|
@@ -153,15 +179,21 @@ def call_openrouter(prompt):
|
|
| 153 |
return None
|
| 154 |
|
| 155 |
def get_ai_response(user_message, intent):
|
| 156 |
-
# Priority 1: OpenRouter
|
| 157 |
-
|
| 158 |
-
if
|
| 159 |
print("β
OpenRouter succeeded")
|
| 160 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
|
| 162 |
-
# Priority
|
| 163 |
-
print("β οΈ
|
| 164 |
-
return
|
| 165 |
|
| 166 |
# ---------- Firebase Helpers ----------
|
| 167 |
def save_message(session_id, role, content, intent=None):
|
|
|
|
| 24 |
ref = db.reference('/')
|
| 25 |
print("β
Firebase connected")
|
| 26 |
|
| 27 |
+
# ---------- Sentence-Transformer Model ----------
|
| 28 |
print("π Loading sentence-transformers...")
|
| 29 |
embedding_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
|
| 30 |
print("β
Embedding model loaded")
|
| 31 |
|
| 32 |
+
# ---------- Local GPT-2 Model (Fallback) with Fixes ----------
|
| 33 |
+
print("π Loading local GPT-2 model...")
|
| 34 |
try:
|
| 35 |
gpt2_tokenizer = AutoTokenizer.from_pretrained("gpt2")
|
| 36 |
gpt2_model = AutoModelForCausalLM.from_pretrained("gpt2")
|
| 37 |
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token
|
|
|
|
| 38 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 39 |
gpt2_model = gpt2_model.to(device)
|
| 40 |
print(f"β
GPT-2 loaded on {device}")
|
|
|
|
| 43 |
gpt2_model = None
|
| 44 |
gpt2_tokenizer = None
|
| 45 |
|
| 46 |
+
def generate_with_gpt2(user_query):
|
| 47 |
+
"""Generate response using GPT-2 with repetition penalties and structured prompt."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
if gpt2_model is None or gpt2_tokenizer is None:
|
| 49 |
+
return None
|
| 50 |
|
| 51 |
+
# Clean prompt β tell GPT-2 it's an assistant for Elite Visuals
|
| 52 |
+
prompt = f"""You are a helpful assistant for Elite Visuals, a premium imagery business.
|
| 53 |
+
Services: product photography, architecture renders, corporate videos, drone cinematography, wedding photography, real estate visuals, AI editing.
|
| 54 |
|
| 55 |
+
Answer the user's question in 2-3 short bullet points. Be direct and helpful. Do not repeat yourself.
|
| 56 |
+
|
| 57 |
+
User: {user_query}
|
| 58 |
Assistant:"""
|
| 59 |
|
| 60 |
+
inputs = gpt2_tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
|
| 61 |
device = next(gpt2_model.parameters()).device
|
| 62 |
inputs = {k: v.to(device) for k, v in inputs.items()}
|
| 63 |
|
| 64 |
with torch.no_grad():
|
| 65 |
outputs = gpt2_model.generate(
|
| 66 |
**inputs,
|
| 67 |
+
max_new_tokens=100,
|
| 68 |
+
temperature=0.7,
|
| 69 |
do_sample=True,
|
| 70 |
top_p=0.9,
|
| 71 |
+
repetition_penalty=1.5, # Penalty for repetition
|
| 72 |
+
no_repeat_ngram_size=3, # No repeating 3-grams
|
| 73 |
pad_token_id=gpt2_tokenizer.eos_token_id
|
| 74 |
)
|
| 75 |
response = gpt2_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 76 |
+
|
| 77 |
# Extract only the assistant part
|
| 78 |
if "Assistant:" in response:
|
| 79 |
response = response.split("Assistant:")[-1].strip()
|
| 80 |
+
else:
|
| 81 |
+
response = response[len(prompt):].strip()
|
| 82 |
+
|
| 83 |
+
# Remove any trailing repetition or incomplete sentences
|
| 84 |
+
lines = response.split('\n')
|
| 85 |
+
unique_lines = []
|
| 86 |
+
for line in lines:
|
| 87 |
+
if line.strip() and line not in unique_lines:
|
| 88 |
+
unique_lines.append(line.strip())
|
| 89 |
+
response = '\n'.join(unique_lines[:4]) # max 4 lines
|
| 90 |
+
|
| 91 |
+
# If response is empty or too short, use fallback
|
| 92 |
+
if len(response) < 10 or response.count('-') == 0:
|
| 93 |
+
return None
|
| 94 |
return response
|
| 95 |
|
| 96 |
+
# ---------- Intent-Based Smart Fallback (No GPT-2 nonsense) ----------
|
| 97 |
+
SERVICES_LIST = ["product photography", "architecture renders", "corporate videos", "drone cinematography", "wedding photography", "real estate visuals", "AI editing", "headshots"]
|
| 98 |
+
|
| 99 |
+
def intent_based_reply(user_message, intent):
|
| 100 |
+
"""Safe fallback when GPT-2 fails."""
|
| 101 |
+
msg_lower = user_message.lower()
|
| 102 |
+
|
| 103 |
+
# Contact
|
| 104 |
+
if "contact" in msg_lower or "phone" in msg_lower or "number" in msg_lower or "email" in msg_lower:
|
| 105 |
+
return "- π Phone: +92 300 1234567\n- βοΈ Email: hello@elitevisuals.com\n- π Website: www.elitevisuals.com\n- β° Timings: Mon-Sat, 10am-7pm"
|
| 106 |
+
|
| 107 |
+
# Drone specific
|
| 108 |
+
if "drone" in msg_lower or "videography" in msg_lower:
|
| 109 |
+
return "- β
Yes, we offer drone cinematography for events, real estate, and commercials.\n- π° Pricing starts at PKR 50,000 per hour.\n- π₯ Sample video available on request."
|
| 110 |
+
|
| 111 |
+
# Wedding
|
| 112 |
+
if "wedding" in msg_lower:
|
| 113 |
+
return "- β
Yes, we provide wedding photography and cinematography.\n- π° Packages from PKR 50,000 to PKR 1,50,000.\n- πΈ Portfolio available on Instagram @elitevisuals."
|
| 114 |
+
|
| 115 |
+
# Pricing
|
| 116 |
+
if any(x in msg_lower for x in ["price", "cost", "rate", "kitne"]):
|
| 117 |
+
return "- πΈ Product photography: PKR 15,000β25,000/session\n- ποΈ Architecture renders: PKR 30,000β50,000/project\n- π₯ Corporate video: PKR 50,000β2,00,000\n- π¬ Wedding photography: PKR 50,000β1,50,000\n- βοΈ Drone services: PKR 25,000β60,000/hour"
|
| 118 |
+
|
| 119 |
+
# Greeting
|
| 120 |
+
if any(x in msg_lower for x in ["salam", "hello", "hi"]):
|
| 121 |
+
return "- Wa-Alaikum-Assalam! π\n- Main Elite Visuals ka assistant hoon.\n- Main in services mein madad kar sakta hoon: " + ", ".join(SERVICES_LIST[:5]) + "."
|
| 122 |
+
|
| 123 |
+
# Out of scope (weather, politics, etc.)
|
| 124 |
+
if any(x in msg_lower for x in ["weather", "cricket", "politics", "imran", "nawaz", "stock", "bitcoin"]):
|
| 125 |
+
return "- β Maazrat, yeh sawaal mere scope mein nahi hai.\n- Main sirf Elite Visuals ki imagery services ke baare mein jaankaari de sakta hoon.\n- Kya aap in mein se kisi ke baare mein poochna chahenge?"
|
| 126 |
+
|
| 127 |
+
# General service inquiry
|
| 128 |
+
return f"- Main Elite Visuals ki in services mein madad kar sakta hoon:\n β’ " + "\n β’ ".join(SERVICES_LIST[:6]) + "\n- Kya aap kisi specific ke baare mein jaanna chahenge?"
|
| 129 |
+
|
| 130 |
+
# ---------- Intent Classification ----------
|
| 131 |
+
IRRELEVANT_KEYWORDS = ["weather", "cricket", "politics", "imran", "nawaz", "election", "stock", "bitcoin", "covid", "corona"]
|
| 132 |
+
|
| 133 |
+
def classify_intent(user_text):
|
| 134 |
+
text_lower = user_text.lower()
|
| 135 |
+
if any(kw in text_lower for kw in IRRELEVANT_KEYWORDS):
|
| 136 |
+
return "irrelevant"
|
| 137 |
+
if any(x in text_lower for x in ["contact", "phone", "number", "whatsapp", "email", "call", "raabta"]):
|
| 138 |
+
return "contact"
|
| 139 |
+
if any(x in text_lower for x in ["price", "cost", "rate", "kitne", "quote", "charges", "fee"]):
|
| 140 |
+
return "pricing"
|
| 141 |
+
if any(x in text_lower for x in ["portfolio", "sample", "gallery", "kaam dikhao"]):
|
| 142 |
+
return "portfolio"
|
| 143 |
+
if any(x in text_lower for x in ["salam", "hello", "hi", "hey"]):
|
| 144 |
+
return "greeting"
|
| 145 |
+
return "service_inquiry"
|
| 146 |
+
|
| 147 |
# ---------- OpenRouter Setup ----------
|
| 148 |
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
|
| 149 |
MODEL = "google/gemini-2.0-flash-lite-preview-02-05"
|
|
|
|
| 153 |
Services: product photography, architecture renders, corporate videos, drone cinematography, 3D rendering, wedding photography, real estate visuals, AI editing.
|
| 154 |
|
| 155 |
Always answer in bullet points. Be concise, helpful, and personalized to the user's exact question.
|
| 156 |
+
If user asks "can you not provide X", clarify that you DO provide it.
|
| 157 |
+
Never repeat the same sentence. Keep answers to 2-4 bullet points."""
|
| 158 |
|
| 159 |
def call_openrouter(prompt):
|
| 160 |
if not OPENROUTER_API_KEY:
|
|
|
|
| 168 |
{"role": "user", "content": prompt}
|
| 169 |
],
|
| 170 |
"temperature": 0.7,
|
| 171 |
+
"max_tokens": 300
|
| 172 |
}
|
| 173 |
try:
|
| 174 |
resp = requests.post(url, json=payload, headers=headers, timeout=15)
|
|
|
|
| 179 |
return None
|
| 180 |
|
| 181 |
def get_ai_response(user_message, intent):
|
| 182 |
+
# Priority 1: OpenRouter
|
| 183 |
+
resp = call_openrouter(user_message)
|
| 184 |
+
if resp and resp.strip() and len(resp) > 10:
|
| 185 |
print("β
OpenRouter succeeded")
|
| 186 |
+
return resp
|
| 187 |
+
|
| 188 |
+
# Priority 2: Local GPT-2 (now fixed)
|
| 189 |
+
gpt2_resp = generate_with_gpt2(user_message)
|
| 190 |
+
if gpt2_resp and len(gpt2_resp) > 10 and "view your clients" not in gpt2_resp.lower():
|
| 191 |
+
print("β
GPT-2 succeeded")
|
| 192 |
+
return gpt2_resp
|
| 193 |
|
| 194 |
+
# Priority 3: Intent-based fallback (safe, no nonsense)
|
| 195 |
+
print("β οΈ Using intent-based fallback")
|
| 196 |
+
return intent_based_reply(user_message, intent)
|
| 197 |
|
| 198 |
# ---------- Firebase Helpers ----------
|
| 199 |
def save_message(session_id, role, content, intent=None):
|