File size: 21,780 Bytes
e61eb95
 
4c20a23
e61eb95
 
 
 
 
 
 
 
 
 
 
 
 
 
6e0b138
e61eb95
4c20a23
 
 
 
 
 
 
 
 
 
e61eb95
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6eb012b
 
 
f39446f
6eb012b
 
 
f39446f
6eb012b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f39446f
6eb012b
 
 
 
 
 
 
 
 
 
 
 
 
 
f39446f
6eb012b
 
 
 
 
 
 
 
 
 
 
 
 
 
f39446f
6eb012b
 
 
f39446f
6eb012b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2d8a87a
6eb012b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e61eb95
 
 
 
4c20a23
e61eb95
 
4c20a23
 
e61eb95
4c20a23
af77ca7
e61eb95
af77ca7
e61eb95
4c20a23
 
 
e61eb95
 
 
 
 
 
 
4c20a23
 
 
 
e61eb95
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4c20a23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e61eb95
 
f39446f
 
 
 
 
 
 
 
e61eb95
 
6e0b138
e61eb95
4c20a23
 
e61eb95
 
9aa3841
 
e61eb95
 
 
 
 
 
 
 
 
 
 
f39446f
 
e61eb95
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4c20a23
e61eb95
4c20a23
e61eb95
 
 
4c20a23
 
 
 
 
e61eb95
4c20a23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e61eb95
 
4c20a23
e61eb95
 
4c20a23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e61eb95
4c20a23
 
 
 
e61eb95
4c20a23
 
 
e61eb95
4c20a23
 
 
 
e61eb95
4c20a23
 
e61eb95
4c20a23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e61eb95
4c20a23
 
 
 
e61eb95
4c20a23
 
 
 
 
 
 
 
 
 
e61eb95
 
 
4c20a23
e61eb95
 
 
4c20a23
 
 
 
 
 
 
 
 
 
 
 
 
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
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
import os
import requests
import base64
from typing import Dict, Any, List
from langchain.docstore.document import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.retrievers import BM25Retriever
from smolagents import CodeAgent, OpenAIServerModel, tool, Tool
from smolagents.vision_web_browser import initialize_driver, save_screenshot, helium_instructions
from smolagents.agents import ActionStep
from selenium import webdriver
from selenium.webdriver.common.by import By
from selenium.webdriver.common.keys import Keys
import helium
from PIL import Image
from io import BytesIO
from time import sleep
from smolagents import PythonInterpreterTool, SpeechToTextTool

# Langfuse observability imports
from opentelemetry.sdk.trace import TracerProvider
from openinference.instrumentation.smolagents import SmolagentsInstrumentor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry import trace
from opentelemetry.trace import format_trace_id
from langfuse import Langfuse


class BM25RetrieverTool(Tool):
    """
    BM25 retriever tool for document search when text documents are available
    """
    name = "bm25_retriever"
    description = "Uses BM25 search to retrieve relevant parts of uploaded documents. Use this when the question references an attached file or document."
    inputs = {
        "query": {
            "type": "string",
            "description": "The search query to find relevant document sections.",
        }
    }
    output_type = "string"

    def __init__(self, docs=None, **kwargs):
        super().__init__(**kwargs)
        self.docs = docs or []
        self.retriever = None
        if self.docs:
            self.retriever = BM25Retriever.from_documents(self.docs, k=5)

    def forward(self, query: str) -> str:
        if not self.retriever:
            return "No documents loaded for retrieval."

        assert isinstance(query, str), "Your search query must be a string"

        docs = self.retriever.invoke(query)
        return "\nRetrieved documents:\n" + "".join([
            f"\n\n===== Document {str(i)} =====\n" + doc.page_content
            for i, doc in enumerate(docs)
        ])


@tool
def search_item_ctrl_f(text: str, nth_result: int = 1) -> str:
    """Search for text on the current page via Ctrl + F and jump to the nth occurrence.
    
    Args:
        text: The text string to search for on the webpage
        nth_result: Which occurrence to jump to (default is 1 for first occurrence)
        
    Returns:
        str: Result of the search operation with match count and navigation status
    """
    try:
        driver = helium.get_driver()
        elements = driver.find_elements(By.XPATH, f"//*[contains(text(), '{text}')]")
        if nth_result > len(elements):
            return f"Match n°{nth_result} not found (only {len(elements)} matches found)"
        result = f"Found {len(elements)} matches for '{text}'."
        elem = elements[nth_result - 1]
        driver.execute_script("arguments[0].scrollIntoView(true);", elem)
        result += f"Focused on element {nth_result} of {len(elements)}"
        return result
    except Exception as e:
        return f"Error searching for text: {e}"


@tool
def go_back() -> str:
    """Navigate back to the previous page in browser history.
    
    Returns:
        str: Confirmation message or error description
    """
    try:
        driver = helium.get_driver()
        driver.back()
        return "Navigated back to previous page"
    except Exception as e:
        return f"Error going back: {e}"


@tool
def close_popups() -> str:
    """Close any visible modal or pop-up on the page by sending ESC key.
    
    Returns:
        str: Confirmation message or error description
    """
    try:
        driver = helium.get_driver()
        webdriver.ActionChains(driver).send_keys(Keys.ESCAPE).perform()
        return "Attempted to close popups"
    except Exception as e:
        return f"Error closing popups: {e}"


@tool
def scroll_page(direction: str = "down", amount: int = 3) -> str:
    """Scroll the webpage in the specified direction.
    
    Args:
        direction: Direction to scroll, either 'up' or 'down'
        amount: Number of scroll actions to perform
        
    Returns:
        str: Confirmation message or error description
    """
    try:
        driver = helium.get_driver()
        for _ in range(amount):
            if direction.lower() == "down":
                driver.execute_script("window.scrollBy(0, 300);")
            elif direction.lower() == "up":
                driver.execute_script("window.scrollBy(0, -300);")
            sleep(0.5)
        return f"Scrolled {direction} {amount} times"
    except Exception as e:
        return f"Error scrolling: {e}"


@tool
def get_page_text() -> str:
    """Extract all visible text from the current webpage.
    
    Returns:
        str: The visible text content of the page
    """
    try:
        driver = helium.get_driver()
        text = driver.find_element(By.TAG_NAME, "body").text
        return f"Page text (first 2000 chars): {text[:2000]}"
    except Exception as e:
        return f"Error getting page text: {e}"


def save_screenshot_callback(memory_step: ActionStep, agent: CodeAgent) -> None:
    """Save screenshots for web browser automation"""
    try:
        sleep(1.0)
        driver = helium.get_driver()
        if driver is not None:
            # Clean up old screenshots
            for previous_memory_step in agent.memory.steps:
                if isinstance(previous_memory_step, ActionStep) and previous_memory_step.step_number <= memory_step.step_number - 2:
                    previous_memory_step.observations_images = None

            png_bytes = driver.get_screenshot_as_png()
            image = Image.open(BytesIO(png_bytes))
            memory_step.observations_images = [image.copy()]

            # Update observations with current URL
            url_info = f"Current url: {driver.current_url}"
            memory_step.observations = (
                url_info if memory_step.observations is None 
                else memory_step.observations + "\n" + url_info
            )
    except Exception as e:
        print(f"Error in screenshot callback: {e}")


class GAIAAgent:
    """
    GAIA agent using smolagents with Gemini 2.0 Flash and Langfuse observability
    """

    def __init__(self, user_id: str = None, session_id: str = None):
        """Initialize the agent with Gemini 2.0 Flash, tools, and Langfuse observability"""

        # Get API keys
        gemini_api_key = os.environ.get("GOOGLE_API_KEY")
        if not gemini_api_key:
            raise ValueError("GOOGLE_API_KEY environment variable not found")

        # Initialize Langfuse observability
        self._setup_langfuse_observability()

        # Initialize Gemini 2.0 Flash model
        self.model = OpenAIServerModel(
            model_id="gemini-2.0-flash",
            api_base="https://generativelanguage.googleapis.com/v1beta/openai/",
            api_key=gemini_api_key,
        )

        # Store user and session IDs for tracking
        self.user_id = user_id or "gaia-user"
        self.session_id = session_id or "gaia-session"

        # GAIA system prompt from the leaderboard
        self.system_prompt = """You are a general AI assistant. I will ask you a question. Report your thoughts and reasoning process clearly. You should use the available tools to gather information and solve problems step by step.

When using web browser automation:
- Use helium commands like go_to(), click(), scroll_down()
- Take screenshots to see what's happening
- Handle popups and forms appropriately
- Be patient with page loading

For document retrieval:
- Use the BM25 retriever when there are text documents attached
- Search with relevant keywords from the question

Your final answer should be as few words as possible, a number, or a comma-separated list. Don't use articles, abbreviations, or units unless specified."""

        # Initialize retriever tool (will be updated when documents are loaded)
        self.retriever_tool = BM25RetrieverTool()

        # Initialize web driver for browser automation
        self.driver = None

        # Create the agent
        self.agent = None
        self._create_agent()

        # Initialize Langfuse client
        self.langfuse = Langfuse()

    def _setup_langfuse_observability(self):
        """Set up Langfuse observability with OpenTelemetry"""
        # Get Langfuse keys from environment variables
        langfuse_public_key = os.environ.get("LANGFUSE_PUBLIC_KEY")
        langfuse_secret_key = os.environ.get("LANGFUSE_SECRET_KEY")
        
        if not langfuse_public_key or not langfuse_secret_key:
            print("Warning: LANGFUSE_PUBLIC_KEY or LANGFUSE_SECRET_KEY not found. Observability will be limited.")
            return

        # Set up Langfuse environment variables
        os.environ["LANGFUSE_HOST"] = os.environ.get("LANGFUSE_HOST", "https://cloud.langfuse.com")
        
        langfuse_auth = base64.b64encode(
            f"{langfuse_public_key}:{langfuse_secret_key}".encode()
        ).decode()
        
        os.environ["OTEL_EXPORTER_OTLP_ENDPOINT"] = os.environ.get("LANGFUSE_HOST") + "/api/public/otel"
        os.environ["OTEL_EXPORTER_OTLP_HEADERS"] = f"Authorization=Basic {langfuse_auth}"

        # Create a TracerProvider for OpenTelemetry
        trace_provider = TracerProvider()
        
        # Add a SimpleSpanProcessor with the OTLPSpanExporter to send traces
        trace_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter()))
        
        # Set the global default tracer provider
        trace.set_tracer_provider(trace_provider)
        self.tracer = trace.get_tracer(__name__)
        
        # Instrument smolagents with the configured provider
        SmolagentsInstrumentor().instrument(tracer_provider=trace_provider)

    def _create_agent(self):
        """Create the CodeAgent with tools"""
        base_tools = [
            self.retriever_tool, 
            search_item_ctrl_f, 
            go_back, 
            close_popups,
            scroll_page,
            get_page_text
        ]

        self.agent = CodeAgent(
            tools=base_tools + [PythonInterpreterTool(), SpeechToTextTool()],
            model=self.model,
            add_base_tools=False,
            planning_interval=2,
            additional_authorized_imports=["helium", "requests", "BeautifulSoup", "json"],
            step_callbacks=[save_screenshot_callback] if self.driver else [],
            max_steps=10,
            description=self.system_prompt,
            verbosity_level=2,
        )

    def initialize_browser(self):
        """Initialize browser for web automation tasks"""
        try:
            chrome_options = webdriver.ChromeOptions()
            chrome_options.add_argument("--force-device-scale-factor=1")
            chrome_options.add_argument("--window-size=1000,1350")
            chrome_options.add_argument("--disable-pdf-viewer")
            chrome_options.add_argument("--window-position=0,0")
            chrome_options.add_argument("--no-sandbox")
            chrome_options.add_argument("--disable-dev-shm-usage")

            self.driver = helium.start_chrome(headless=False, options=chrome_options)

            # Recreate agent with browser tools
            self._create_agent()

            # Import helium for the agent
            self.agent.python_executor("from helium import *")

            return True
        except Exception as e:
            print(f"Failed to initialize browser: {e}")
            return False

    def load_documents_from_file(self, file_path: str):
        """Load and process documents from a file for BM25 retrieval"""
        try:
            # Read file content
            with open(file_path, 'r', encoding='utf-8') as f:
                content = f.read()

            # Split into chunks
            text_splitter = RecursiveCharacterTextSplitter(
                chunk_size=1000,
                chunk_overlap=200,
                separators=["\n\n", "\n", ".", " ", ""]
            )

            # Create documents
            chunks = text_splitter.split_text(content)
            docs = [Document(page_content=chunk, metadata={"source": file_path}) 
                    for chunk in chunks]

            # Update retriever tool
            self.retriever_tool = BM25RetrieverTool(docs)

            # Recreate agent with updated retriever
            self._create_agent()

            print(f"Loaded {len(docs)} document chunks from {file_path}")
            return True

        except Exception as e:
            print(f"Error loading documents from {file_path}: {e}")
            return False

    def download_gaia_file(self, task_id: str, api_url: str = "https://agents-course-unit4-scoring.hf.space") -> str:
        """Download file associated with GAIA task_id"""
        try:
            response = requests.get(f"{api_url}/files/{task_id}", timeout=30)
            response.raise_for_status()

            filename = f"task_{task_id}_file.txt"
            with open(filename, 'wb') as f:
                f.write(response.content)

            return filename
        except Exception as e:
            print(f"Failed to download file for task {task_id}: {e}")
            return None

    def solve_gaia_question(self, question_data: Dict[str, Any], tags: List[str] = None) -> str:
        """
        Solve a GAIA question with full Langfuse observability
        """
        question = question_data.get("Question", "")
        task_id = question_data.get("task_id", "")
        
        # Prepare tags for observability
        trace_tags = ["gaia-agent", "question-solving"]
        if tags:
            trace_tags.extend(tags)
        if task_id:
            trace_tags.append(f"task-{task_id}")

        # Start Langfuse trace with OpenTelemetry
        with self.tracer.start_as_current_span("GAIA-Question-Solving") as span:
            try:
                # Set span attributes for tracking
                span.set_attribute("langfuse.user.id", self.user_id)
                span.set_attribute("langfuse.session.id", self.session_id)
                span.set_attribute("langfuse.tags", trace_tags)
                span.set_attribute("gaia.task_id", task_id)
                span.set_attribute("gaia.question_length", len(question))

                # Get trace ID for Langfuse linking
                current_span = trace.get_current_span()
                span_context = current_span.get_span_context()
                trace_id = span_context.trace_id
                formatted_trace_id = format_trace_id(trace_id)

                # Create Langfuse trace
                langfuse_trace = self.langfuse.trace(
                    id=formatted_trace_id,
                    name="GAIA Question Solving",
                    input={"question": question, "task_id": task_id},
                    user_id=self.user_id,
                    session_id=self.session_id,
                    tags=trace_tags,
                    metadata={
                        "model": self.model.model_id,
                        "question_length": len(question),
                        "has_file": bool(task_id)
                    }
                )

                # Download and load file if task_id provided
                file_loaded = False
                if task_id:
                    file_path = self.download_gaia_file(task_id)
                    if file_path:
                        file_loaded = self.load_documents_from_file(file_path)
                        span.set_attribute("gaia.file_loaded", file_loaded)
                        print(f"Loaded file for task {task_id}")

                # Check if this requires web browsing
                web_indicators = ["navigate", "browser", "website", "webpage", "url", "click", "search on"]
                needs_browser = any(indicator in question.lower() for indicator in web_indicators)
                span.set_attribute("gaia.needs_browser", needs_browser)

                if needs_browser and not self.driver:
                    print("Initializing browser for web automation...")
                    browser_initialized = self.initialize_browser()
                    span.set_attribute("gaia.browser_initialized", browser_initialized)

                # Prepare the prompt
                prompt = f"""
Question: {question}
{f'Task ID: {task_id}' if task_id else ''}
{f'File loaded: Yes' if file_loaded else 'File loaded: No'}

Solve this step by step. Use the available tools to gather information and provide a precise answer.
                """

                if needs_browser:
                    prompt += "\n" + helium_instructions

                print("=== AGENT REASONING ===")
                result = self.agent.run(prompt)
                print("=== END REASONING ===")

                # Update Langfuse trace with result
                langfuse_trace.update(
                    output={"answer": str(result)},
                    end_time=None  # Will be set automatically
                )

                # Add success attributes
                span.set_attribute("gaia.success", True)
                span.set_attribute("gaia.answer_length", len(str(result)))

                # Flush Langfuse data
                self.langfuse.flush()

                return str(result)

            except Exception as e:
                error_msg = f"Error processing question: {str(e)}"
                print(error_msg)
                
                # Log error to span and Langfuse
                span.set_attribute("gaia.success", False)
                span.set_attribute("gaia.error", str(e))
                
                if 'langfuse_trace' in locals():
                    langfuse_trace.update(
                        output={"error": error_msg},
                        level="ERROR"
                    )
                
                self.langfuse.flush()
                return error_msg
                
            finally:
                # Clean up browser if initialized
                if self.driver:
                    try:
                        helium.kill_browser()
                    except:
                        pass

    def evaluate_answer(self, question: str, answer: str, expected_answer: str = None) -> Dict[str, Any]:
        """
        Evaluate the agent's answer using LLM-as-a-Judge and optionally compare with expected answer
        """
        evaluation_prompt = f"""
Please evaluate the following answer to a question on a scale of 1-5:

Question: {question}
Answer: {answer}
{f'Expected Answer: {expected_answer}' if expected_answer else ''}

Rate the answer on:
1. Accuracy (1-5)
2. Completeness (1-5) 
3. Clarity (1-5)

Provide your rating as JSON: {{"accuracy": X, "completeness": Y, "clarity": Z, "overall": W, "reasoning": "explanation"}}
        """

        try:
            # Use the same model to evaluate
            evaluation_result = self.agent.run(evaluation_prompt)
            
            # Try to parse JSON response
            import json
            try:
                scores = json.loads(evaluation_result)
                return scores
            except:
                # Fallback if JSON parsing fails
                return {
                    "accuracy": 3,
                    "completeness": 3,
                    "clarity": 3,
                    "overall": 3,
                    "reasoning": "Could not parse evaluation response",
                    "raw_evaluation": evaluation_result
                }
                
        except Exception as e:
            return {
                "accuracy": 1,
                "completeness": 1,
                "clarity": 1,
                "overall": 1,
                "reasoning": f"Evaluation failed: {str(e)}"
            }

    def add_user_feedback(self, trace_id: str, feedback_score: int, comment: str = None):
        """
        Add user feedback to a specific trace
        
        Args:
            trace_id: The trace ID to add feedback to
            feedback_score: Score from 0-5 (0=very bad, 5=excellent)
            comment: Optional comment from user
        """
        try:
            self.langfuse.score(
                trace_id=trace_id,
                name="user-feedback",
                value=feedback_score,
                comment=comment
            )
            self.langfuse.flush()
            print(f"User feedback added: {feedback_score}/5")
        except Exception as e:
            print(f"Error adding user feedback: {e}")


# Example usage with observability
if __name__ == "__main__":
    # Set up environment variables (you need to set these)
    # os.environ["GOOGLE_API_KEY"] = "your-gemini-api-key"
    # os.environ["LANGFUSE_PUBLIC_KEY"] = "pk-lf-..."
    # os.environ["LANGFUSE_SECRET_KEY"] = "sk-lf-..."
    
    # Test the agent with observability
    agent = GAIAAgent(
        user_id="test-user-123",
        session_id="test-session-456"
    )

    # Example question
    question_data = {
        "Question": "How many studio albums Mercedes Sosa has published between 2000-2009?",
        "task_id": ""
    }

    # Solve with full observability
    answer = agent.solve_gaia_question(
        question_data, 
        tags=["music-question", "discography"]
    )
    print(f"Answer: {answer}")

    # Evaluate the answer
    evaluation = agent.evaluate_answer(
        question_data["Question"], 
        answer
    )
    print(f"Evaluation: {evaluation}")