ItamarMordo commited on
Commit
8dc1c1f
·
verified ·
1 Parent(s): a35a0ff

Update server.py

Browse files
Files changed (1) hide show
  1. server.py +17 -32
server.py CHANGED
@@ -14,8 +14,11 @@ from pydantic import BaseModel
14
  from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
15
 
16
  # --- Configuration ---
17
- # UPGRADE: Switching to 7B model (fits in 16GB RAM easily)
18
- MODEL_ID = "Qwen/Qwen2.5-7B-Instruct"
 
 
 
19
  PORT = int(os.environ.get("PORT", 8000))
20
  HOST = "0.0.0.0"
21
  DB_FILE = "chatbot.db"
@@ -24,7 +27,7 @@ JSON_SOURCE = "database.json"
24
  logging.basicConfig(level=logging.INFO)
25
  logger = logging.getLogger(__name__)
26
 
27
- app = FastAPI(title="Hebrew Smart Home Technician (7B)", version="14.0.0")
28
 
29
  app.add_middleware(
30
  CORSMiddleware,
@@ -97,34 +100,22 @@ def expand_query(query: str) -> List[str]:
97
  return list(search_terms)
98
 
99
  def simplify_schematic_text(text: str) -> str:
100
- # 1. ניקוי שמות רכיבים טכניים מה-DB החדש
101
  text = text.replace("ESP32-DEVKITC-32D_ESP32-DEVKITC-30GPIO", "בקר ESP32")
102
-
103
- # Specific component mappings based on your schematic
104
  text = re.sub(r"U2\b", "ESP32 (U2)", text)
105
  text = re.sub(r"U4\b", "Level Shifter (U4)", text)
106
  text = re.sub(r"U3\b", "Main Board (U3)", text)
107
  text = re.sub(r"LCD1\b", "Nextion Screen (LCD1)", text)
108
-
109
- # Clean up the raw database strings if they still exist
110
  text = text.replace("LevelConvertor (BOB_19002)", "Level Shifter (U4)")
111
  text = text.replace("BOB_19002", "")
112
  text = text.replace("StudentBoard (0201080606)", "Main Board (U3)")
113
  text = text.replace("0201080606", "")
114
  text = text.replace("040709_0804060105", "")
115
-
116
- # 2. Clean Generic names
117
  text = text.replace("(NEXTION)", "")
118
  text = text.replace("HMI", "")
119
-
120
- # 3. Clean up syntax
121
- text = text.replace("()", "") # Remove empty brackets
122
- text = text.replace(" ", " ") # Remove double spaces
123
-
124
- # 4. Translate verbs for the Hebrew sentence structure
125
  text = text.replace("connects to", "מחובר אל")
126
  text = text.replace("Pin", "פין")
127
-
128
  return text
129
 
130
  def query_knowledge_base(query: str) -> Tuple[str, List[str]]:
@@ -146,7 +137,6 @@ def query_knowledge_base(query: str) -> Tuple[str, List[str]]:
146
  score += 100
147
  elif term.lower() in topic.lower():
148
  score += 10
149
-
150
  results.append((score, content))
151
  seen_ids.add(row_id)
152
  conn.close()
@@ -155,7 +145,6 @@ def query_knowledge_base(query: str) -> Tuple[str, List[str]]:
155
 
156
  if results:
157
  raw_content = results[0][1]
158
-
159
  if "Schematic" in raw_content:
160
  lines = raw_content.split('\n')
161
  relevant_lines = []
@@ -167,7 +156,6 @@ def query_knowledge_base(query: str) -> Tuple[str, List[str]]:
167
  relevant_lines.append(f"- {clean_line}")
168
 
169
  if relevant_lines:
170
- # Limit to relevant lines only
171
  context_str = "רשימת החיבורים שנמצאה:\n" + "\n".join(relevant_lines)
172
  return context_str, search_terms
173
  else:
@@ -192,7 +180,7 @@ class Message(BaseModel):
192
 
193
  class ChatCompletionRequest(BaseModel):
194
  messages: List[Message]
195
- max_tokens: Optional[int] = 512
196
  temperature: Optional[float] = 0.2
197
  stream: Optional[bool] = False
198
 
@@ -200,11 +188,11 @@ class ChatCompletionRequest(BaseModel):
200
  async def startup_event():
201
  global model, tokenizer, device
202
  init_db()
203
- logger.info(f"Loading UPGRADED model: {MODEL_ID}...")
204
  device = "cuda" if torch.cuda.is_available() else "cpu"
205
  try:
206
  tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
207
- # Using float16 if possible, otherwise standard. 7B fits in 16GB easily in float16.
208
  model = AutoModelForCausalLM.from_pretrained(
209
  MODEL_ID,
210
  torch_dtype=torch.float16 if device != "cpu" else "auto",
@@ -239,17 +227,15 @@ async def chat_completions(request: ChatCompletionRequest):
239
  "תפקידך: אתה טכנאי מומחה המסביר על חיבורים במערכת 'בית חכם'.\n"
240
  "הנחיות:\n"
241
  "1. דבר אך ורק בשפה העברית.\n"
242
- "2. השתמש במידע המצורף כדי לענות. הוא מכיל רשימת חיבורים מדויקת.\n"
243
- "3. כאשר אתה מתאר חיבור, ציין במפורש את שמות הרכיבים ואת מספרי הפינים.\n"
244
- "4. דוגמה לתשובה טובה: 'המסך מחובר לממיר הרמות בפין 7, והממיר מחובר לבקר בפין RX'.\n"
245
- "5. אל תמציא מידע. אם חסר פין, אל תנחש אותו.\n"
246
- f"\nמידע טכני מהשרטוט:\n{context}"
247
  )
248
  else:
249
  sys_msg = (
250
  "You are a Hebrew-speaking assistant. "
251
  "If the user says 'Shalom', reply with 'שלום! איך אוכל לעזור?'. "
252
- "Otherwise, say you don't have that information."
253
  )
254
 
255
  if request.messages[0].role == "system":
@@ -264,8 +250,8 @@ async def chat_completions(request: ChatCompletionRequest):
264
 
265
  gen_kwargs = dict(
266
  inputs=prompt_ids,
267
- max_new_tokens=request.max_tokens,
268
- temperature=0.3, # Slightly higher for 7B as it's more stable
269
  top_p=0.9,
270
  do_sample=True,
271
  pad_token_id=tokenizer.eos_token_id,
@@ -282,7 +268,6 @@ async def chat_completions(request: ChatCompletionRequest):
282
  full_text = ""
283
  for text in streamer:
284
  if text:
285
- # Still keep the guard just in case, though 7B needs it less
286
  if not is_valid_text(text):
287
  logger.warning(f"Blocked: {text}")
288
  break
 
14
  from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
15
 
16
  # --- Configuration ---
17
+ # OPTION 1: Stick with 7B for smarts, but limit output length
18
+ MODEL_ID = "Qwen/Qwen2.5-7B-Instruct"
19
+ # OPTION 2: Switch back to 3B for 2x speed boost (Uncomment below line)
20
+ # MODEL_ID = "Qwen/Qwen2.5-3B-Instruct"
21
+
22
  PORT = int(os.environ.get("PORT", 8000))
23
  HOST = "0.0.0.0"
24
  DB_FILE = "chatbot.db"
 
27
  logging.basicConfig(level=logging.INFO)
28
  logger = logging.getLogger(__name__)
29
 
30
+ app = FastAPI(title="Hebrew Smart Home Technician (Optimized)", version="14.1.0")
31
 
32
  app.add_middleware(
33
  CORSMiddleware,
 
100
  return list(search_terms)
101
 
102
  def simplify_schematic_text(text: str) -> str:
 
103
  text = text.replace("ESP32-DEVKITC-32D_ESP32-DEVKITC-30GPIO", "בקר ESP32")
 
 
104
  text = re.sub(r"U2\b", "ESP32 (U2)", text)
105
  text = re.sub(r"U4\b", "Level Shifter (U4)", text)
106
  text = re.sub(r"U3\b", "Main Board (U3)", text)
107
  text = re.sub(r"LCD1\b", "Nextion Screen (LCD1)", text)
 
 
108
  text = text.replace("LevelConvertor (BOB_19002)", "Level Shifter (U4)")
109
  text = text.replace("BOB_19002", "")
110
  text = text.replace("StudentBoard (0201080606)", "Main Board (U3)")
111
  text = text.replace("0201080606", "")
112
  text = text.replace("040709_0804060105", "")
 
 
113
  text = text.replace("(NEXTION)", "")
114
  text = text.replace("HMI", "")
115
+ text = text.replace("()", "")
116
+ text = text.replace(" ", " ")
 
 
 
 
117
  text = text.replace("connects to", "מחובר אל")
118
  text = text.replace("Pin", "פין")
 
119
  return text
120
 
121
  def query_knowledge_base(query: str) -> Tuple[str, List[str]]:
 
137
  score += 100
138
  elif term.lower() in topic.lower():
139
  score += 10
 
140
  results.append((score, content))
141
  seen_ids.add(row_id)
142
  conn.close()
 
145
 
146
  if results:
147
  raw_content = results[0][1]
 
148
  if "Schematic" in raw_content:
149
  lines = raw_content.split('\n')
150
  relevant_lines = []
 
156
  relevant_lines.append(f"- {clean_line}")
157
 
158
  if relevant_lines:
 
159
  context_str = "רשימת החיבורים שנמצאה:\n" + "\n".join(relevant_lines)
160
  return context_str, search_terms
161
  else:
 
180
 
181
  class ChatCompletionRequest(BaseModel):
182
  messages: List[Message]
183
+ max_tokens: Optional[int] = 256 # Reduced for speed!
184
  temperature: Optional[float] = 0.2
185
  stream: Optional[bool] = False
186
 
 
188
  async def startup_event():
189
  global model, tokenizer, device
190
  init_db()
191
+ logger.info(f"Loading model: {MODEL_ID}...")
192
  device = "cuda" if torch.cuda.is_available() else "cpu"
193
  try:
194
  tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
195
+ # Using float16 for speed on compatible hardware, or auto
196
  model = AutoModelForCausalLM.from_pretrained(
197
  MODEL_ID,
198
  torch_dtype=torch.float16 if device != "cpu" else "auto",
 
227
  "תפקידך: אתה טכנאי מומחה המסביר על חיבורים במערכת 'בית חכם'.\n"
228
  "הנחיות:\n"
229
  "1. דבר אך ורק בשפה העברית.\n"
230
+ "2. השתמש במידע המצורף כדי לענות.\n"
231
+ "3. קצר ולעניין! כתוב משפט אחד או שניים בלבד.\n"
232
+ f"\nמידע טכני:\n{context}"
 
 
233
  )
234
  else:
235
  sys_msg = (
236
  "You are a Hebrew-speaking assistant. "
237
  "If the user says 'Shalom', reply with 'שלום! איך אוכל לעזור?'. "
238
+ "Keep answers short."
239
  )
240
 
241
  if request.messages[0].role == "system":
 
250
 
251
  gen_kwargs = dict(
252
  inputs=prompt_ids,
253
+ max_new_tokens=request.max_tokens, # Using the reduced limit
254
+ temperature=0.2,
255
  top_p=0.9,
256
  do_sample=True,
257
  pad_token_id=tokenizer.eos_token_id,
 
268
  full_text = ""
269
  for text in streamer:
270
  if text:
 
271
  if not is_valid_text(text):
272
  logger.warning(f"Blocked: {text}")
273
  break