Spaces:
Sleeping
Sleeping
Update server.py
Browse files
server.py
CHANGED
|
@@ -14,7 +14,8 @@ from pydantic import BaseModel
|
|
| 14 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 15 |
|
| 16 |
# --- Configuration ---
|
| 17 |
-
|
|
|
|
| 18 |
PORT = int(os.environ.get("PORT", 8000))
|
| 19 |
HOST = "0.0.0.0"
|
| 20 |
DB_FILE = "chatbot.db"
|
|
@@ -23,7 +24,7 @@ JSON_SOURCE = "database.json"
|
|
| 23 |
logging.basicConfig(level=logging.INFO)
|
| 24 |
logger = logging.getLogger(__name__)
|
| 25 |
|
| 26 |
-
app = FastAPI(title="Hebrew Smart Home Technician", version="
|
| 27 |
|
| 28 |
app.add_middleware(
|
| 29 |
CORSMiddleware,
|
|
@@ -86,6 +87,7 @@ def expand_query(query: str) -> List[str]:
|
|
| 86 |
return list(search_terms)
|
| 87 |
|
| 88 |
def simplify_schematic_text(text: str) -> str:
|
|
|
|
| 89 |
text = text.replace("ESP32-DEVKITC-32D_ESP32-DEVKITC-30GPIO", "בקר ESP32")
|
| 90 |
text = text.replace("BOB_19002", "ממיר רמות")
|
| 91 |
text = text.replace("TLM-SH_board", "הלוח הראשי")
|
|
@@ -132,13 +134,11 @@ def query_knowledge_base(query: str) -> str:
|
|
| 132 |
relevant_lines = []
|
| 133 |
target_terms = [t.lower() for t in search_terms if len(t) > 1]
|
| 134 |
for line in lines:
|
| 135 |
-
|
|
|
|
|
|
|
| 136 |
relevant_lines.append(simplify_schematic_text(line.strip()))
|
| 137 |
|
| 138 |
-
# Optimization: Limit lines
|
| 139 |
-
if len(relevant_lines) > 10:
|
| 140 |
-
relevant_lines = relevant_lines[:10]
|
| 141 |
-
|
| 142 |
if relevant_lines:
|
| 143 |
return "הנה החיבורים שנמצאו:\n" + "\n".join(relevant_lines)
|
| 144 |
|
|
@@ -161,25 +161,18 @@ class Message(BaseModel):
|
|
| 161 |
class ChatCompletionRequest(BaseModel):
|
| 162 |
messages: List[Message]
|
| 163 |
max_tokens: Optional[int] = 512
|
| 164 |
-
temperature: Optional[float] = 0.
|
| 165 |
stream: Optional[bool] = False
|
| 166 |
|
| 167 |
-
# --- פילטר שפה (מונע סינית/ג'יבריש) ---
|
| 168 |
-
def is_valid_text(text: str) -> bool:
|
| 169 |
-
# בודק אם הטקסט מכיל תווים בטווח של סינית/יפנית/קוריאנית
|
| 170 |
-
for char in text:
|
| 171 |
-
if 0x4E00 <= ord(char) <= 0x9FFF: # CJK Unified Ideographs
|
| 172 |
-
return False
|
| 173 |
-
if 0x3040 <= ord(char) <= 0x309F: # Hiragana
|
| 174 |
-
return False
|
| 175 |
-
return True
|
| 176 |
-
|
| 177 |
@app.on_event("startup")
|
| 178 |
async def startup_event():
|
| 179 |
global model, tokenizer, device
|
| 180 |
init_db()
|
| 181 |
-
|
|
|
|
|
|
|
| 182 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
|
| 183 |
try:
|
| 184 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 185 |
model = AutoModelForCausalLM.from_pretrained(
|
|
@@ -189,10 +182,10 @@ async def startup_event():
|
|
| 189 |
low_cpu_mem_usage=True,
|
| 190 |
trust_remote_code=True
|
| 191 |
)
|
| 192 |
-
logger.info("Model loaded!")
|
| 193 |
except Exception as e:
|
| 194 |
logger.error(f"Load failed: {e}")
|
| 195 |
-
raise RuntimeError("Model failed")
|
| 196 |
|
| 197 |
def run_generation(model, tokenizer, params):
|
| 198 |
try:
|
|
@@ -213,11 +206,11 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 213 |
|
| 214 |
if context:
|
| 215 |
sys_msg = (
|
| 216 |
-
"אתה
|
| 217 |
"הנחיות:\n"
|
| 218 |
-
"1. השתמש במידע המצורף.\n"
|
| 219 |
-
"2. אם
|
| 220 |
-
"3.
|
| 221 |
f"\nמידע טכני:\n{context}"
|
| 222 |
)
|
| 223 |
else:
|
|
@@ -240,7 +233,7 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 240 |
top_p=0.9,
|
| 241 |
do_sample=True,
|
| 242 |
pad_token_id=tokenizer.eos_token_id,
|
| 243 |
-
repetition_penalty=1.05
|
| 244 |
)
|
| 245 |
|
| 246 |
if request.stream:
|
|
@@ -252,7 +245,7 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 252 |
async def stream_generator():
|
| 253 |
full_text = ""
|
| 254 |
for text in streamer:
|
| 255 |
-
if text
|
| 256 |
full_text += text
|
| 257 |
yield f"data: {json.dumps({'choices': [{'delta': {'content': text}}]}, ensure_ascii=False)}\n\n"
|
| 258 |
save_message("assistant", full_text)
|
|
|
|
| 14 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 15 |
|
| 16 |
# --- Configuration ---
|
| 17 |
+
# UPGRADE: Switching to the 3B model for better intelligence and Hebrew support
|
| 18 |
+
MODEL_ID = "Qwen/Qwen2.5-3B-Instruct"
|
| 19 |
PORT = int(os.environ.get("PORT", 8000))
|
| 20 |
HOST = "0.0.0.0"
|
| 21 |
DB_FILE = "chatbot.db"
|
|
|
|
| 24 |
logging.basicConfig(level=logging.INFO)
|
| 25 |
logger = logging.getLogger(__name__)
|
| 26 |
|
| 27 |
+
app = FastAPI(title="Hebrew Smart Home Technician (3B)", version="12.0.0")
|
| 28 |
|
| 29 |
app.add_middleware(
|
| 30 |
CORSMiddleware,
|
|
|
|
| 87 |
return list(search_terms)
|
| 88 |
|
| 89 |
def simplify_schematic_text(text: str) -> str:
|
| 90 |
+
# ניקוי אגרסיבי והפיכה לעברית
|
| 91 |
text = text.replace("ESP32-DEVKITC-32D_ESP32-DEVKITC-30GPIO", "בקר ESP32")
|
| 92 |
text = text.replace("BOB_19002", "ממיר רמות")
|
| 93 |
text = text.replace("TLM-SH_board", "הלוח הראשי")
|
|
|
|
| 134 |
relevant_lines = []
|
| 135 |
target_terms = [t.lower() for t in search_terms if len(t) > 1]
|
| 136 |
for line in lines:
|
| 137 |
+
# לוגיקה משופרת: אם חיפשנו רכיב (כמו מסך), נביא את כל החיבורים שלו
|
| 138 |
+
line_lower = line.lower()
|
| 139 |
+
if "connect" in line_lower and any(t in line_lower for t in target_terms):
|
| 140 |
relevant_lines.append(simplify_schematic_text(line.strip()))
|
| 141 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 142 |
if relevant_lines:
|
| 143 |
return "הנה החיבורים שנמצאו:\n" + "\n".join(relevant_lines)
|
| 144 |
|
|
|
|
| 161 |
class ChatCompletionRequest(BaseModel):
|
| 162 |
messages: List[Message]
|
| 163 |
max_tokens: Optional[int] = 512
|
| 164 |
+
temperature: Optional[float] = 0.3 # 3B model is smarter, can handle slightly higher temp
|
| 165 |
stream: Optional[bool] = False
|
| 166 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
@app.on_event("startup")
|
| 168 |
async def startup_event():
|
| 169 |
global model, tokenizer, device
|
| 170 |
init_db()
|
| 171 |
+
|
| 172 |
+
# UPGRADE LOGIC
|
| 173 |
+
logger.info(f"Loading UPGRADED model: {MODEL_ID}...")
|
| 174 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 175 |
+
|
| 176 |
try:
|
| 177 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 178 |
model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
| 182 |
low_cpu_mem_usage=True,
|
| 183 |
trust_remote_code=True
|
| 184 |
)
|
| 185 |
+
logger.info("Model loaded successfully!")
|
| 186 |
except Exception as e:
|
| 187 |
logger.error(f"Load failed: {e}")
|
| 188 |
+
raise RuntimeError("Model failed to load. Check internet or RAM.")
|
| 189 |
|
| 190 |
def run_generation(model, tokenizer, params):
|
| 191 |
try:
|
|
|
|
| 206 |
|
| 207 |
if context:
|
| 208 |
sys_msg = (
|
| 209 |
+
"תפקידך: אתה המדריך הטכני של פרויקט 'בית חכם תלם'. עליך לענות בעברית טבעית וברורה.\n"
|
| 210 |
"הנחיות:\n"
|
| 211 |
+
"1. השתמש במידע המצורף כדי לענות.\n"
|
| 212 |
+
"2. אם המידע הוא רשימת חיבורים, אל תקריא אותה כמו רובוט. הסבר אותה במשפט פשוט.\n"
|
| 213 |
+
"3. דוגמה: במקום 'פין 1 מחובר לפין 2', תגיד: 'המסך מתחבר לבקר בפיני התקשורת וכן למתח'.\n"
|
| 214 |
f"\nמידע טכני:\n{context}"
|
| 215 |
)
|
| 216 |
else:
|
|
|
|
| 233 |
top_p=0.9,
|
| 234 |
do_sample=True,
|
| 235 |
pad_token_id=tokenizer.eos_token_id,
|
| 236 |
+
repetition_penalty=1.05
|
| 237 |
)
|
| 238 |
|
| 239 |
if request.stream:
|
|
|
|
| 245 |
async def stream_generator():
|
| 246 |
full_text = ""
|
| 247 |
for text in streamer:
|
| 248 |
+
if text:
|
| 249 |
full_text += text
|
| 250 |
yield f"data: {json.dumps({'choices': [{'delta': {'content': text}}]}, ensure_ascii=False)}\n\n"
|
| 251 |
save_message("assistant", full_text)
|