Spaces:
Sleeping
Sleeping
Update server.py
Browse files
server.py
CHANGED
|
@@ -13,7 +13,6 @@ from fastapi.responses import StreamingResponse
|
|
| 13 |
from pydantic import BaseModel
|
| 14 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 15 |
|
| 16 |
-
# --- Configuration ---
|
| 17 |
MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 18 |
PORT = int(os.environ.get("PORT", 8000))
|
| 19 |
HOST = "0.0.0.0"
|
|
@@ -23,7 +22,7 @@ JSON_SOURCE = "database.json"
|
|
| 23 |
logging.basicConfig(level=logging.INFO)
|
| 24 |
logger = logging.getLogger(__name__)
|
| 25 |
|
| 26 |
-
app = FastAPI(title="Hebrew RAG Chatbot", version="5.
|
| 27 |
|
| 28 |
app.add_middleware(
|
| 29 |
CORSMiddleware,
|
|
@@ -64,7 +63,7 @@ def init_db():
|
|
| 64 |
def query_knowledge_base(query: str) -> str:
|
| 65 |
conn = sqlite3.connect(DB_FILE)
|
| 66 |
cursor = conn.cursor()
|
| 67 |
-
# חיפוש פשוט
|
| 68 |
keywords = [w for w in query.split() if len(w) > 2]
|
| 69 |
results = []
|
| 70 |
|
|
@@ -76,7 +75,7 @@ def query_knowledge_base(query: str) -> str:
|
|
| 76 |
results.append(row[0])
|
| 77 |
|
| 78 |
conn.close()
|
| 79 |
-
return "\n---\n".join(results[:2])
|
| 80 |
|
| 81 |
def save_message(role: str, content: str):
|
| 82 |
try:
|
|
@@ -94,7 +93,7 @@ class Message(BaseModel):
|
|
| 94 |
class ChatCompletionRequest(BaseModel):
|
| 95 |
messages: List[Message]
|
| 96 |
max_tokens: Optional[int] = 512
|
| 97 |
-
temperature: Optional[float] = 0.
|
| 98 |
stream: Optional[bool] = False
|
| 99 |
|
| 100 |
@app.on_event("startup")
|
|
@@ -131,30 +130,29 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 131 |
|
| 132 |
last_user_msg = next((m for m in reversed(request.messages) if m.role == "user"), None)
|
| 133 |
|
| 134 |
-
# 1. חיפוש מידע
|
| 135 |
context = ""
|
| 136 |
if last_user_msg:
|
| 137 |
save_message("user", last_user_msg.content)
|
| 138 |
context = query_knowledge_base(last_user_msg.content)
|
| 139 |
|
| 140 |
-
# 2. בניית ההנחיה (פשוטה וישירה בעברית)
|
| 141 |
if context:
|
|
|
|
| 142 |
sys_msg = (
|
| 143 |
-
"
|
| 144 |
-
"
|
| 145 |
-
|
| 146 |
-
"
|
|
|
|
|
|
|
| 147 |
)
|
| 148 |
else:
|
| 149 |
-
sys_msg = "
|
| 150 |
|
| 151 |
-
# עדכון או הוספת הודעת מערכת
|
| 152 |
if request.messages[0].role == "system":
|
| 153 |
request.messages[0].content = sys_msg
|
| 154 |
else:
|
| 155 |
request.messages.insert(0, Message(role="system", content=sys_msg))
|
| 156 |
|
| 157 |
-
# 3. יצירת תשובה
|
| 158 |
try:
|
| 159 |
conversation = [msg.dict() for msg in request.messages]
|
| 160 |
prompt_ids = tokenizer.apply_chat_template(conversation, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(device)
|
|
@@ -163,11 +161,11 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 163 |
gen_kwargs = dict(
|
| 164 |
inputs=prompt_ids,
|
| 165 |
max_new_tokens=request.max_tokens,
|
| 166 |
-
temperature=0.
|
| 167 |
top_p=0.9,
|
| 168 |
do_sample=True,
|
| 169 |
pad_token_id=tokenizer.eos_token_id,
|
| 170 |
-
repetition_penalty=1.0 # ל
|
| 171 |
)
|
| 172 |
|
| 173 |
if request.stream:
|
|
@@ -187,7 +185,7 @@ async def chat_completions(request: ChatCompletionRequest):
|
|
| 187 |
|
| 188 |
return StreamingResponse(stream_generator(), media_type="text/event-stream; charset=utf-8")
|
| 189 |
|
| 190 |
-
return {}
|
| 191 |
|
| 192 |
if __name__ == "__main__":
|
| 193 |
uvicorn.run("server:app", host=HOST, port=PORT, reload=False)
|
|
|
|
| 13 |
from pydantic import BaseModel
|
| 14 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 15 |
|
|
|
|
| 16 |
MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 17 |
PORT = int(os.environ.get("PORT", 8000))
|
| 18 |
HOST = "0.0.0.0"
|
|
|
|
| 22 |
logging.basicConfig(level=logging.INFO)
|
| 23 |
logger = logging.getLogger(__name__)
|
| 24 |
|
| 25 |
+
app = FastAPI(title="Hebrew RAG Chatbot", version="5.1.0")
|
| 26 |
|
| 27 |
app.add_middleware(
|
| 28 |
CORSMiddleware,
|
|
|
|
| 63 |
def query_knowledge_base(query: str) -> str:
|
| 64 |
conn = sqlite3.connect(DB_FILE)
|
| 65 |
cursor = conn.cursor()
|
| 66 |
+
# חיפוש מילים פשוט
|
| 67 |
keywords = [w for w in query.split() if len(w) > 2]
|
| 68 |
results = []
|
| 69 |
|
|
|
|
| 75 |
results.append(row[0])
|
| 76 |
|
| 77 |
conn.close()
|
| 78 |
+
return "\n---\n".join(results[:2])
|
| 79 |
|
| 80 |
def save_message(role: str, content: str):
|
| 81 |
try:
|
|
|
|
| 93 |
class ChatCompletionRequest(BaseModel):
|
| 94 |
messages: List[Message]
|
| 95 |
max_tokens: Optional[int] = 512
|
| 96 |
+
temperature: Optional[float] = 0.2
|
| 97 |
stream: Optional[bool] = False
|
| 98 |
|
| 99 |
@app.on_event("startup")
|
|
|
|
| 130 |
|
| 131 |
last_user_msg = next((m for m in reversed(request.messages) if m.role == "user"), None)
|
| 132 |
|
|
|
|
| 133 |
context = ""
|
| 134 |
if last_user_msg:
|
| 135 |
save_message("user", last_user_msg.content)
|
| 136 |
context = query_knowledge_base(last_user_msg.content)
|
| 137 |
|
|
|
|
| 138 |
if context:
|
| 139 |
+
# הנחיה מדויקת להשתמש במידע הקיים
|
| 140 |
sys_msg = (
|
| 141 |
+
"System: You are a helpful assistant for the 'Telem Smart Home' project. "
|
| 142 |
+
"Instructions:\n"
|
| 143 |
+
"1. Answer ONLY in Hebrew.\n"
|
| 144 |
+
"2. Use the context provided below to answer.\n"
|
| 145 |
+
"3. Do not invent new information.\n\n"
|
| 146 |
+
f"Context:\n{context}"
|
| 147 |
)
|
| 148 |
else:
|
| 149 |
+
sys_msg = "System: You are a helpful assistant. Answer strictly in Hebrew. If you don't know, say so politely."
|
| 150 |
|
|
|
|
| 151 |
if request.messages[0].role == "system":
|
| 152 |
request.messages[0].content = sys_msg
|
| 153 |
else:
|
| 154 |
request.messages.insert(0, Message(role="system", content=sys_msg))
|
| 155 |
|
|
|
|
| 156 |
try:
|
| 157 |
conversation = [msg.dict() for msg in request.messages]
|
| 158 |
prompt_ids = tokenizer.apply_chat_template(conversation, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(device)
|
|
|
|
| 161 |
gen_kwargs = dict(
|
| 162 |
inputs=prompt_ids,
|
| 163 |
max_new_tokens=request.max_tokens,
|
| 164 |
+
temperature=0.2, # נמוך = יציב
|
| 165 |
top_p=0.9,
|
| 166 |
do_sample=True,
|
| 167 |
pad_token_id=tokenizer.eos_token_id,
|
| 168 |
+
repetition_penalty=1.0 # מבוטל לחלוטין למניעת שיבושי שפה
|
| 169 |
)
|
| 170 |
|
| 171 |
if request.stream:
|
|
|
|
| 185 |
|
| 186 |
return StreamingResponse(stream_generator(), media_type="text/event-stream; charset=utf-8")
|
| 187 |
|
| 188 |
+
return {}
|
| 189 |
|
| 190 |
if __name__ == "__main__":
|
| 191 |
uvicorn.run("server:app", host=HOST, port=PORT, reload=False)
|