File size: 16,639 Bytes
8f308fb
80a97c8
a5d9083
207f9f7
80a97c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f308fb
 
80a97c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f308fb
80a97c8
8f308fb
a5d9083
8f308fb
 
80a97c8
a5d9083
 
 
 
 
 
80a97c8
 
a5d9083
80a97c8
 
 
8f308fb
80a97c8
 
 
 
e440f24
 
 
 
8f308fb
 
e440f24
 
 
 
 
 
 
 
 
80a97c8
8f308fb
 
 
80a97c8
 
8f308fb
80a97c8
 
 
8f308fb
80a97c8
 
8f308fb
 
 
80a97c8
 
8f308fb
 
 
 
e440f24
a5d9083
 
 
 
 
 
 
 
 
80a97c8
a5d9083
 
80a97c8
a5d9083
8f308fb
a5d9083
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80a97c8
 
 
8f308fb
 
80a97c8
 
 
8f308fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
207f9f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# llm_router.py - FIXED VERSION
import logging
import asyncio
from typing import Dict
from .models_config import LLM_CONFIG

logger = logging.getLogger(__name__)

class LLMRouter:
    def __init__(self, hf_token):
        self.hf_token = hf_token
        self.health_status = {}
        logger.info("LLMRouter initialized")
        if hf_token:
            logger.info("HF token available")
        else:
            logger.warning("No HF token provided")
        
    async def route_inference(self, task_type: str, prompt: str, **kwargs):
        """
        Smart routing based on task specialization
        """
        logger.info(f"Routing inference for task: {task_type}")
        model_config = self._select_model(task_type)
        logger.info(f"Selected model: {model_config['model_id']}")
        
        # Health check and fallback logic
        if not await self._is_model_healthy(model_config["model_id"]):
            logger.warning(f"Model unhealthy, using fallback")
            model_config = self._get_fallback_model(task_type)
            logger.info(f"Fallback model: {model_config['model_id']}")
            
        # FIXED: Ensure task_type is passed to the _call_hf_endpoint method
        result = await self._call_hf_endpoint(model_config, prompt, task_type, **kwargs)
        logger.info(f"Inference complete for {task_type}")
        return result
    
    def _select_model(self, task_type: str) -> dict:
        model_map = {
            "intent_classification": LLM_CONFIG["models"]["classification_specialist"],
            "embedding_generation": LLM_CONFIG["models"]["embedding_specialist"],
            "safety_check": LLM_CONFIG["models"]["safety_checker"],
            "general_reasoning": LLM_CONFIG["models"]["reasoning_primary"],
            "response_synthesis": LLM_CONFIG["models"]["reasoning_primary"]
        }
        return model_map.get(task_type, LLM_CONFIG["models"]["reasoning_primary"])
    
    async def _is_model_healthy(self, model_id: str) -> bool:
        """
        Check if the model is healthy and available
        Mark models as healthy by default - actual availability checked at API call time
        """
        # Check cached health status
        if model_id in self.health_status:
            return self.health_status[model_id]
        
        # All models marked healthy initially - real check happens during API call
        self.health_status[model_id] = True
        return True
    
    def _get_fallback_model(self, task_type: str) -> dict:
        """
        Get fallback model configuration for the task type
        """
        # Fallback mapping
        fallback_map = {
            "intent_classification": LLM_CONFIG["models"]["reasoning_primary"],
            "embedding_generation": LLM_CONFIG["models"]["embedding_specialist"],
            "safety_check": LLM_CONFIG["models"]["reasoning_primary"],
            "general_reasoning": LLM_CONFIG["models"]["reasoning_primary"],
            "response_synthesis": LLM_CONFIG["models"]["reasoning_primary"]
        }
        return fallback_map.get(task_type, LLM_CONFIG["models"]["reasoning_primary"])
    
    async def _call_hf_endpoint(self, model_config: dict, prompt: str, task_type: str, **kwargs):
        """
        FIXED: Make actual call to Hugging Face Chat Completions API
        Uses the correct chat completions protocol with retry logic and exponential backoff
        
        IMPORTANT: task_type parameter is now properly included in the method signature
        """
        # Retry configuration
        max_retries = kwargs.get('max_retries', 3)
        initial_delay = kwargs.get('initial_delay', 1.0)  # Start with 1 second
        max_delay = kwargs.get('max_delay', 16.0)  # Cap at 16 seconds
        timeout = kwargs.get('timeout', 30)
        
        try:
            import requests
            from requests.exceptions import Timeout, RequestException, ConnectionError as RequestsConnectionError
            
            model_id = model_config["model_id"]
            
            # Use the chat completions endpoint
            api_url = "https://router.huggingface.co/v1/chat/completions"
            
            logger.info(f"Calling HF Chat Completions API for model: {model_id}")
            logger.debug(f"Prompt length: {len(prompt)}")
            logger.info("=" * 80)
            logger.info("LLM API REQUEST - COMPLETE PROMPT:")
            logger.info("=" * 80)
            logger.info(f"Model: {model_id}")
            
            # FIXED: task_type is now properly available as a parameter
            logger.info(f"Task Type: {task_type}")
            logger.info(f"Prompt Length: {len(prompt)} characters")
            logger.info("-" * 40)
            logger.info("FULL PROMPT CONTENT:")
            logger.info("-" * 40)
            logger.info(prompt)
            logger.info("-" * 40)
            logger.info("END OF PROMPT")
            logger.info("=" * 80)
            
            # Prepare the request payload
            max_tokens = kwargs.get('max_tokens', 512)
            temperature = kwargs.get('temperature', 0.7)
            
            payload = {
                "model": model_id,
                "messages": [
                    {
                        "role": "user",
                        "content": prompt
                    }
                ],
                "max_tokens": max_tokens,
                "temperature": temperature,
                "stream": False
            }
            
            headers = {
                "Authorization": f"Bearer {self.hf_token}",
                "Content-Type": "application/json"
            }
            
            # Retry logic with exponential backoff
            last_exception = None
            for attempt in range(max_retries + 1):
                try:
                    if attempt > 0:
                        # Calculate exponential backoff delay
                        delay = min(initial_delay * (2 ** (attempt - 1)), max_delay)
                        logger.warning(f"Retry attempt {attempt}/{max_retries} after {delay:.1f}s delay (exponential backoff)")
                        await asyncio.sleep(delay)
                    
                    logger.info(f"Sending request to: {api_url} (attempt {attempt + 1}/{max_retries + 1})")
                    logger.debug(f"Payload: {payload}")
                    
                    response = requests.post(api_url, json=payload, headers=headers, timeout=timeout)
                    
                    if response.status_code == 200:
                        result = response.json()
                        logger.debug(f"Raw response: {result}")
                        
                        if 'choices' in result and len(result['choices']) > 0:
                            generated_text = result['choices'][0]['message']['content']
                            
                            if not generated_text or generated_text.strip() == "":
                                logger.warning(f"Empty or invalid response, using fallback")
                                return None
                            
                            if attempt > 0:
                                logger.info(f"Successfully retrieved response after {attempt} retry attempts")
                            
                            logger.info(f"HF API returned response (length: {len(generated_text)})")
                            logger.info("=" * 80)
                            logger.info("COMPLETE LLM API RESPONSE:")
                            logger.info("=" * 80)
                            logger.info(f"Model: {model_id}")
                            
                            # FIXED: task_type is now properly available
                            logger.info(f"Task Type: {task_type}")
                            logger.info(f"Response Length: {len(generated_text)} characters")
                            logger.info("-" * 40)
                            logger.info("FULL RESPONSE CONTENT:")
                            logger.info("-" * 40)
                            logger.info(generated_text)
                            logger.info("-" * 40)
                            logger.info("END OF LLM RESPONSE")
                            logger.info("=" * 80)
                            return generated_text
                        else:
                            logger.error(f"Unexpected response format: {result}")
                            return None
                    elif response.status_code == 503:
                        # Model is loading - this is retryable
                        if attempt < max_retries:
                            logger.warning(f"Model loading (503), will retry (attempt {attempt + 1}/{max_retries + 1})")
                            last_exception = Exception(f"Model loading (503)")
                            continue
                        else:
                            # After max retries, try fallback model
                            logger.warning(f"Model loading (503) after {max_retries} retries, trying fallback model")
                            fallback_config = self._get_fallback_model(task_type)
                            
                            # FIXED: Ensure task_type is passed in recursive call
                            return await self._call_hf_endpoint(fallback_config, prompt, task_type, **kwargs)
                    else:
                        # Non-retryable HTTP errors
                        logger.error(f"HF API error: {response.status_code} - {response.text}")
                        return None
                        
                except Timeout as e:
                    last_exception = e
                    if attempt < max_retries:
                        logger.warning(f"Request timeout (attempt {attempt + 1}/{max_retries + 1}): {str(e)}")
                        continue
                    else:
                        logger.error(f"Request timeout after {max_retries} retries: {str(e)}")
                        # Try fallback model on final timeout
                        logger.warning("Attempting fallback model due to persistent timeout")
                        fallback_config = self._get_fallback_model(task_type)
                        return await self._call_hf_endpoint(fallback_config, prompt, task_type, **kwargs)
                        
                except (RequestsConnectionError, RequestException) as e:
                    last_exception = e
                    if attempt < max_retries:
                        logger.warning(f"Connection error (attempt {attempt + 1}/{max_retries + 1}): {str(e)}")
                        continue
                    else:
                        logger.error(f"Connection error after {max_retries} retries: {str(e)}")
                        # Try fallback model on final connection error
                        logger.warning("Attempting fallback model due to persistent connection error")
                        fallback_config = self._get_fallback_model(task_type)
                        return await self._call_hf_endpoint(fallback_config, prompt, task_type, **kwargs)
            
            # If we exhausted all retries and didn't return
            if last_exception:
                logger.error(f"Failed after {max_retries} retries. Last error: {last_exception}")
                return None
                
        except ImportError:
            logger.warning("requests library not available, using mock response")
            return f"[Mock] Response to: {prompt[:100]}..."
        except Exception as e:
            logger.error(f"Error calling HF endpoint: {e}", exc_info=True)
            return None
    
    async def get_available_models(self):
        """
        Get list of available models for testing
        """
        return list(LLM_CONFIG["models"].keys())
    
    async def health_check(self):
        """
        Perform health check on all models
        """
        health_status = {}
        for model_name, model_config in LLM_CONFIG["models"].items():
            model_id = model_config["model_id"]
            is_healthy = await self._is_model_healthy(model_id)
            health_status[model_name] = {
                "model_id": model_id,
                "healthy": is_healthy
            }
        
        return health_status
    
    def prepare_context_for_llm(self, raw_context: Dict, max_tokens: int = 4000) -> str:
        """Smart context windowing for LLM calls"""
        
        try:
            from transformers import AutoTokenizer
            
            # Initialize tokenizer lazily
            if not hasattr(self, 'tokenizer'):
                try:
                    self.tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
                except Exception as e:
                    logger.warning(f"Could not load tokenizer: {e}, using character count estimation")
                    self.tokenizer = None
        except ImportError:
            logger.warning("transformers library not available, using character count estimation")
            self.tokenizer = None
        
        # Priority order for context elements
        priority_elements = [
            ('current_query', 1.0),
            ('recent_interactions', 0.8),
            ('user_preferences', 0.6),
            ('session_summary', 0.4),
            ('historical_context', 0.2)
        ]
        
        formatted_context = []
        total_tokens = 0
        
        for element, priority in priority_elements:
            # Map element names to context keys
            element_key_map = {
                'current_query': raw_context.get('user_input', ''),
                'recent_interactions': raw_context.get('interaction_contexts', []),
                'user_preferences': raw_context.get('preferences', {}),
                'session_summary': raw_context.get('session_context', {}),
                'historical_context': raw_context.get('user_context', '')
            }
            
            content = element_key_map.get(element, '')
            
            # Convert to string if needed
            if isinstance(content, dict):
                content = str(content)
            elif isinstance(content, list):
                content = "\n".join([str(item) for item in content[:10]])  # Limit to 10 items
            
            if not content:
                continue
            
            # Estimate tokens
            if self.tokenizer:
                try:
                    tokens = len(self.tokenizer.encode(content))
                except:
                    # Fallback to character-based estimation (rough: 1 token ≈ 4 chars)
                    tokens = len(content) // 4
            else:
                # Character-based estimation (rough: 1 token ≈ 4 chars)
                tokens = len(content) // 4
            
            if total_tokens + tokens <= max_tokens:
                formatted_context.append(f"=== {element.upper()} ===\n{content}")
                total_tokens += tokens
            elif priority > 0.5:  # Critical elements - truncate if needed
                available = max_tokens - total_tokens
                if available > 100:  # Only truncate if we have meaningful space
                    truncated = self._truncate_to_tokens(content, available)
                    formatted_context.append(f"=== {element.upper()} (TRUNCATED) ===\n{truncated}")
                break
        
        return "\n\n".join(formatted_context)
    
    def _truncate_to_tokens(self, content: str, max_tokens: int) -> str:
        """Truncate content to fit within token limit"""
        if not self.tokenizer:
            # Simple character-based truncation
            max_chars = max_tokens * 4
            if len(content) <= max_chars:
                return content
            return content[:max_chars-3] + "..."
        
        try:
            # Tokenize and truncate
            tokens = self.tokenizer.encode(content)
            if len(tokens) <= max_tokens:
                return content
            
            truncated_tokens = tokens[:max_tokens-3]  # Leave room for "..."
            truncated_text = self.tokenizer.decode(truncated_tokens)
            return truncated_text + "..."
        except Exception as e:
            logger.warning(f"Error truncating with tokenizer: {e}, using character truncation")
            max_chars = max_tokens * 4
            if len(content) <= max_chars:
                return content
            return content[:max_chars-3] + "..."