Spaces:
Running
Running
File size: 19,113 Bytes
eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 1d5e56d eb10851 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 |
"""
Agent Service - Central Brain for Sales & Feedback Agents
Manages LLM conversation loop with tool calling
"""
from typing import Dict, Any, List, Optional
import os
from tools_service import ToolsService
class AgentService:
"""
Manages the conversation loop between User -> LLM -> Tools -> Response
"""
def __init__(
self,
tools_service: ToolsService,
embedding_service,
qdrant_service,
advanced_rag,
hf_token: str,
feedback_tracking=None # NEW: Optional feedback tracking
):
self.tools_service = tools_service
self.embedding_service = embedding_service
self.qdrant_service = qdrant_service
self.advanced_rag = advanced_rag
self.hf_token = hf_token
self.feedback_tracking = feedback_tracking
# Load system prompts
self.prompts = self._load_prompts()
def _load_prompts(self) -> Dict[str, str]:
"""Load system prompts from files"""
prompts = {}
prompts_dir = "prompts"
for mode in ["sales_agent", "feedback_agent"]:
filepath = os.path.join(prompts_dir, f"{mode}.txt")
try:
with open(filepath, 'r', encoding='utf-8') as f:
prompts[mode] = f.read()
print(f"✓ Loaded prompt: {mode}")
except Exception as e:
print(f"⚠️ Error loading {mode} prompt: {e}")
prompts[mode] = ""
return prompts
async def chat(
self,
user_message: str,
conversation_history: List[Dict],
mode: str = "sales", # "sales" or "feedback"
user_id: Optional[str] = None,
access_token: Optional[str] = None, # NEW: For authenticated API calls
max_iterations: int = 3
) -> Dict[str, Any]:
"""
Main conversation loop
Args:
user_message: User's input
conversation_history: Previous messages [{"role": "user", "content": ...}, ...]
mode: "sales" or "feedback"
user_id: User ID (for feedback mode to check purchase history)
access_token: JWT token for authenticated API calls
max_iterations: Maximum tool call iterations to prevent infinite loops
Returns:
{
"message": "Bot response",
"tool_calls": [...], # List of tools called (for debugging)
"mode": mode
}
"""
print(f"\n🤖 Agent Mode: {mode}")
print(f"👤 User Message: {user_message}")
print(f"🔑 Auth Info:")
print(f" - User ID: {user_id}")
print(f" - Access Token: {'✅ Received' if access_token else '❌ None'}")
# Store user_id and access_token for tool calls
self.current_user_id = user_id
self.current_access_token = access_token
if access_token:
print(f" - Stored access_token for tools: {access_token[:20]}...")
if user_id:
print(f" - Stored user_id for tools: {user_id}")
# Select system prompt
system_prompt = self._get_system_prompt(mode)
# Build conversation context
messages = self._build_messages(system_prompt, conversation_history, user_message)
# Agentic loop: LLM may call tools multiple times
tool_calls_made = []
current_response = None
for iteration in range(max_iterations):
print(f"\n🔄 Iteration {iteration + 1}")
# Call LLM
llm_response = await self._call_llm(messages)
print(f"🧠 LLM Response: {llm_response[:200]}...")
# Check if LLM wants to call a tool
tool_call = self._parse_tool_call(llm_response)
if not tool_call:
# No tool call -> This is the final response
current_response = llm_response
break
# Execute tool
print(f"🔧 Tool Called: {tool_call['tool_name']}")
# Auto-inject real user_id for get_purchased_events
if tool_call['tool_name'] == 'get_purchased_events' and self.current_user_id:
print(f"🔄 Auto-injecting real user_id: {self.current_user_id}")
tool_call['arguments']['user_id'] = self.current_user_id
tool_result = await self.tools_service.execute_tool(
tool_call['tool_name'],
tool_call['arguments'],
access_token=self.current_access_token # Pass access_token
)
# Record tool call
tool_calls_made.append({
"function": tool_call['tool_name'],
"arguments": tool_call['arguments'],
"result": tool_result
})
# Add tool result to conversation
messages.append({
"role": "assistant",
"content": llm_response
})
messages.append({
"role": "system",
"content": f"Tool Result:\n{self._format_tool_result({'result': tool_result})}"
})
# If tool returns "run_rag_search", handle it specially
if isinstance(tool_result, dict) and tool_result.get("action") == "run_rag_search":
rag_results = await self._execute_rag_search(tool_result["query"])
messages[-1]["content"] = f"RAG Search Results:\n{rag_results}"
# Clean up response
final_response = current_response or llm_response
final_response = self._clean_response(final_response)
return {
"message": final_response,
"tool_calls": tool_calls_made,
"mode": mode
}
def _get_system_prompt(self, mode: str) -> str:
"""Get system prompt for selected mode with tools definition"""
prompt_key = f"{mode}_agent" if mode in ["sales", "feedback"] else "sales_agent"
base_prompt = self.prompts.get(prompt_key, "")
# Add tools definition (filtered by mode)
tools_definition = self._get_tools_definition(mode)
return f"{base_prompt}\n\n{tools_definition}"
def _get_tools_definition(self, mode: str = "sales") -> str:
"""Get tools definition in text format for prompt, filtered by mode"""
# Base header
header = """
# AVAILABLE TOOLS
You can call the following tools when needed. To call a tool, output a JSON block like this:
```json
{
"tool_call": "tool_name",
"arguments": {
"arg1": "value1",
"arg2": "value2"
}
}
```
## Tools List:
"""
# Tools available for ALL modes
common_tools = """
### search_events
Search for events matching user criteria.
Arguments:
- query (string): Search keywords
- vibe (string, optional): Mood/vibe (e.g., "chill", "sôi động")
- time (string, optional): Time period (e.g., "cuối tuần này")
Example:
```json
{"tool_call": "search_events", "arguments": {"query": "nhạc rock", "vibe": "sôi động"}}
```
### get_event_details
Get detailed information about a specific event.
Arguments:
- event_id (string): Event ID from search results
Example:
```json
{"tool_call": "get_event_details", "arguments": {"event_id": "6900ae38eb03f29702c7fd1d"}}
```
"""
# Tools ONLY for sales mode
sales_only_tools = """
### save_lead
Save customer contact information.
Arguments:
- email (string, optional): Email address
- phone (string, optional): Phone number
- interest (string, optional): What they're interested in
Example:
```json
{"tool_call": "save_lead", "arguments": {"email": "user@example.com", "interest": "Rock show"}}
```
"""
# Tools ONLY for feedback mode
feedback_only_tools = """
### get_purchased_events
Check which events the user has attended. MUST call this tool to get REAL data from API.
Arguments:
- user_id (string): User ID
Example:
```json
{"tool_call": "get_purchased_events", "arguments": {"user_id": "user_123"}}
```
### save_feedback
Save user's feedback/review for an event.
Arguments:
- event_id (string): Event ID
- rating (integer): 1-5 stars
- comment (string, optional): User's comment
Example:
```json
{"tool_call": "save_feedback", "arguments": {"event_id": "abc123", "rating": 5, "comment": "Tuyệt vời!"}}
```
"""
# Footer with important notes
footer = """
**IMPORTANT:**
- Call tools ONLY when you need real-time data
- After receiving tool results, respond naturally to the user
- Don't expose raw JSON to users - always format nicely
- NEVER invent or fabricate data - always use real results from tools
"""
# Build tools definition based on mode
if mode == "feedback":
return header + common_tools + feedback_only_tools + footer
else: # sales mode (default)
return header + common_tools + sales_only_tools + footer
def _build_messages(
self,
system_prompt: str,
history: List[Dict],
user_message: str
) -> List[Dict]:
"""Build messages array for LLM"""
messages = [{"role": "system", "content": system_prompt}]
# Add conversation history
messages.extend(history)
# Add current user message
messages.append({"role": "user", "content": user_message})
return messages
async def _call_llm(self, messages: List[Dict]) -> str:
"""
Call HuggingFace LLM directly using chat_completion (conversational)
"""
try:
from huggingface_hub import AsyncInferenceClient
# Create async client
client = AsyncInferenceClient(token=self.hf_token)
# Call HF API with chat completion (conversational)
response_text = ""
async for message in await client.chat_completion(
messages=messages, # Use messages directly
model="openai/gpt-oss-20b", # GPT-OSS 20B
max_tokens=512,
temperature=0.7,
stream=True
):
if message.choices and message.choices[0].delta.content:
response_text += message.choices[0].delta.content
return response_text
except Exception as e:
print(f"⚠️ LLM Call Error: {e}")
return "Xin lỗi, tôi đang gặp chút vấn đề kỹ thuật. Bạn thử lại sau nhé!"
def _messages_to_prompt(self, messages: List[Dict]) -> str:
"""Convert messages array to single prompt string"""
prompt_parts = []
for msg in messages:
role = msg["role"]
content = msg["content"]
if role == "system":
prompt_parts.append(f"[SYSTEM]\n{content}\n")
elif role == "user":
prompt_parts.append(f"[USER]\n{content}\n")
elif role == "assistant":
prompt_parts.append(f"[ASSISTANT]\n{content}\n")
return "\n".join(prompt_parts)
def _format_tool_result(self, tool_result: Dict) -> str:
"""Format tool result for feeding back to LLM"""
result = tool_result.get("result", {})
# Special handling for purchased events list
if isinstance(result, list):
print(f"\n🔍 Formatting {len(result)} purchased events for LLM")
if not result:
return "User has not purchased any events yet."
# Format each event clearly
formatted_events = []
for i, event in enumerate(result, 1):
event_info = []
event_info.append(f"Event {i}:")
# Extract key fields
if 'eventName' in event:
event_info.append(f" Name: {event['eventName']}")
if 'eventCode' in event:
event_info.append(f" Code: {event['eventCode']}")
if '_id' in event:
event_info.append(f" ID: {event['_id']}")
if 'startTimeEventTime' in event:
event_info.append(f" Date: {event['startTimeEventTime']}")
formatted_events.append("\n".join(event_info))
formatted = "User's Purchased Events:\n\n" + "\n\n".join(formatted_events)
print(f"📤 Sending to LLM:\n{formatted}")
return formatted
# Default formatting for other results
if isinstance(result, dict):
# Pretty print key info
formatted = []
for key, value in result.items():
if key not in ["success", "error"]:
formatted.append(f"{key}: {value}")
return "\n".join(formatted)
return str(result)
async def _execute_rag_search(self, query_params: Dict) -> str:
"""
Execute RAG search for event discovery
Called when LLM wants to search_events
"""
query = query_params.get("query", "")
vibe = query_params.get("vibe", "")
# Build search query
search_text = f"{query} {vibe}".strip()
print(f"🔍 RAG Search: {search_text}")
# Use embedding + qdrant
embedding = self.embedding_service.encode_text(search_text)
results = self.qdrant_service.search(
query_embedding=embedding,
limit=5
)
# Format results
formatted = []
for i, result in enumerate(results, 1):
# Result is a dict with keys: id, score, payload
payload = result.get("payload", {})
texts = payload.get("texts", [])
text = texts[0] if texts else ""
event_id = payload.get("id_use", "")
formatted.append(f"{i}. {text[:100]}... (ID: {event_id})")
return "\n".join(formatted) if formatted else "Không tìm thấy sự kiện phù hợp."
def _parse_tool_call(self, llm_response: str) -> Optional[Dict]:
"""
Parse LLM response to detect tool calls using structured JSON
Returns:
{"tool_name": "...", "arguments": {...}} or None
"""
import json
import re
# Method 1: Look for JSON code block
json_match = re.search(r'```json\s*(\{.*?\})\s*```', llm_response, re.DOTALL)
if json_match:
try:
data = json.loads(json_match.group(1))
return self._extract_tool_from_json(data)
except json.JSONDecodeError:
pass
# Method 2: Look for inline JSON object
# Find all potential JSON objects
json_objects = re.findall(r'\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}', llm_response)
for json_str in json_objects:
try:
data = json.loads(json_str)
tool_call = self._extract_tool_from_json(data)
if tool_call:
return tool_call
except json.JSONDecodeError:
continue
# Method 3: Nested JSON (for complex structures)
try:
# Find outermost curly braces
if '{' in llm_response and '}' in llm_response:
start = llm_response.find('{')
# Find matching closing brace
count = 0
for i, char in enumerate(llm_response[start:], start):
if char == '{':
count += 1
elif char == '}':
count -= 1
if count == 0:
json_str = llm_response[start:i+1]
data = json.loads(json_str)
return self._extract_tool_from_json(data)
except (json.JSONDecodeError, ValueError):
pass
return None
def _extract_tool_from_json(self, data: dict) -> Optional[Dict]:
"""
Extract tool call information from parsed JSON
Supports multiple formats:
- {"tool_call": "search_events", "arguments": {...}}
- {"function": "search_events", "parameters": {...}}
- {"name": "search_events", "args": {...}}
"""
# Format 1: tool_call + arguments
if "tool_call" in data and isinstance(data["tool_call"], str):
return {
"tool_name": data["tool_call"],
"arguments": data.get("arguments", {})
}
# Format 2: function + parameters
if "function" in data:
return {
"tool_name": data["function"],
"arguments": data.get("parameters", data.get("arguments", {}))
}
# Format 3: name + args
if "name" in data:
return {
"tool_name": data["name"],
"arguments": data.get("args", data.get("arguments", {}))
}
# Format 4: Direct tool name as key
valid_tools = ["search_events", "get_event_details", "get_purchased_events", "save_feedback", "save_lead"]
for tool in valid_tools:
if tool in data:
return {
"tool_name": tool,
"arguments": data[tool] if isinstance(data[tool], dict) else {}
}
return None
def _clean_response(self, response: str) -> str:
"""Remove JSON artifacts from final response"""
# Remove JSON blocks
if "```json" in response:
response = response.split("```json")[0]
if "```" in response:
response = response.split("```")[0]
# Remove tool call markers
if "{" in response and "tool_call" in response:
# Find the last natural sentence before JSON
lines = response.split("\n")
cleaned = []
for line in lines:
if "{" in line and "tool_call" in line:
break
cleaned.append(line)
response = "\n".join(cleaned)
return response.strip()
|