Update agent.py
Browse files
agent.py
CHANGED
@@ -9,7 +9,6 @@ from llama_index.core.query_engine import RetrieverQueryEngine
|
|
9 |
from llama_index.readers.file import PDFReader, DocxReader, CSVReader, ImageReader
|
10 |
import os
|
11 |
from typing import List, Dict, Any
|
12 |
-
from llama_index.readers.web import SimpleWebPageReader
|
13 |
from llama_index.core.tools.ondemand_loader_tool import OnDemandLoaderTool
|
14 |
from llama_index.tools.arxiv import ArxivToolSpec
|
15 |
import duckduckgo_search as ddg
|
@@ -20,6 +19,7 @@ from llama_index.callbacks.wandb import WandbCallbackHandler
|
|
20 |
from llama_index.core.callbacks.base import CallbackManager
|
21 |
from llama_index.core.callbacks.llama_debug import LlamaDebugHandler
|
22 |
from llama_index.core import Settings
|
|
|
23 |
|
24 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
25 |
from llama_index.llms.huggingface import HuggingFaceLLM
|
@@ -240,170 +240,197 @@ analysis_agent = FunctionAgent(
|
|
240 |
)
|
241 |
|
242 |
|
|
|
|
|
|
|
243 |
class IntelligentSourceRouter:
|
244 |
def __init__(self):
|
245 |
-
# Initialize
|
246 |
-
self.
|
247 |
-
|
248 |
-
|
249 |
-
self.web_reader = SimpleWebPageReader()
|
250 |
-
|
251 |
-
# Create OnDemandLoaderTool for web content
|
252 |
-
self.web_loader_tool = OnDemandLoaderTool.from_defaults(
|
253 |
-
self.web_reader,
|
254 |
-
name="Web Content Loader",
|
255 |
-
description="Load and analyze web page content with intelligent chunking and search"
|
256 |
-
)
|
257 |
-
|
258 |
-
def web_search_fallback(self, query: str, max_results: int = 5) -> str:
|
259 |
-
try:
|
260 |
-
results = ddg.DDGS().text(query, max_results=max_results)
|
261 |
-
return "\n".join([f"{i}. **{r['title']}**\n URL: {r['href']}\n {r['body']}" for i, r in enumerate(results, 1)])
|
262 |
-
except Exception as e:
|
263 |
-
return f"Search failed: {str(e)}"
|
264 |
-
|
265 |
-
def extract_web_content(self, urls: List[str], query: str) -> str:
|
266 |
-
"""Extract and analyze content from web URLs"""
|
267 |
-
try:
|
268 |
-
content_results = []
|
269 |
-
for url in urls[:3]: # Limit to top 3 URLs
|
270 |
-
try:
|
271 |
-
result = self.web_loader_tool.call(
|
272 |
-
urls=[url],
|
273 |
-
query=f"Extract information relevant to: {query}"
|
274 |
-
)
|
275 |
-
content_results.append(f"**Content from {url}:**\n{result}")
|
276 |
-
except Exception as e:
|
277 |
-
content_results.append(f"**Failed to load {url}**: {str(e)}")
|
278 |
-
|
279 |
-
return "\n\n".join(content_results)
|
280 |
-
except Exception as e:
|
281 |
-
return f"Content extraction failed: {str(e)}"
|
282 |
-
|
283 |
def detect_intent_and_route(self, query: str) -> str:
|
284 |
-
#
|
285 |
intent_prompt = f"""
|
286 |
Analyze this query and determine if it's scientific research or general information:
|
287 |
Query: "{query}"
|
288 |
-
|
289 |
Choose ONE source:
|
290 |
- arxiv: For scientific research, academic papers, technical studies, algorithms, experiments
|
291 |
- web_search: For all other information (current events, general facts, weather, how-to guides, etc.)
|
292 |
-
|
293 |
Respond with ONLY "arxiv" or "web_search".
|
294 |
"""
|
295 |
-
|
296 |
-
response = proj_llm.complete(intent_prompt)
|
297 |
selected_source = response.text.strip().lower()
|
298 |
-
|
299 |
-
# Execute search and extract content
|
300 |
results = [f"**Query**: {query}", f"**Selected Source**: {selected_source}", "="*50]
|
301 |
-
|
302 |
try:
|
303 |
if selected_source == 'arxiv':
|
304 |
-
result = self.
|
305 |
results.append(f"**ArXiv Research:**\n{result}")
|
306 |
-
|
307 |
-
|
308 |
-
#
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
except Exception as e:
|
319 |
results.append(f"**Search failed**: {str(e)}")
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
return "\n\n".join(results)
|
322 |
|
323 |
# Initialize router
|
324 |
intelligent_router = IntelligentSourceRouter()
|
325 |
|
326 |
# Create enhanced research tool
|
327 |
-
def enhanced_smart_research_tool(query: str, task_context: str = "", max_results: int =
|
328 |
full_query = f"{query} {task_context}".strip()
|
329 |
-
return intelligent_router.
|
330 |
|
331 |
research_tool = FunctionTool.from_defaults(
|
332 |
fn=enhanced_smart_research_tool,
|
333 |
-
name="
|
334 |
-
description="Intelligent research
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
return "Code executed successfully"
|
357 |
-
|
358 |
-
except Exception as e:
|
359 |
-
return f"Code execution failed: {str(e)}"
|
360 |
-
|
361 |
-
code_execution_tool = FunctionTool.from_defaults(
|
362 |
-
fn=execute_python_code,
|
363 |
-
name="Python Code Execution",
|
364 |
-
description="Execute Python code safely for calculations and data processing"
|
365 |
)
|
366 |
|
367 |
-
|
368 |
-
code_agent = ReActAgent(
|
369 |
name="CodeAgent",
|
370 |
description="Advanced calculations, data processing, and final answer synthesis using ReAct reasoning",
|
371 |
system_prompt="""
|
372 |
You are a coding and reasoning specialist using ReAct methodology.
|
373 |
-
|
374 |
For each task:
|
375 |
1. THINK: Analyze what needs to be calculated or processed
|
376 |
2. ACT: Execute appropriate code or calculations
|
377 |
3. OBSERVE: Review results and determine if more work is needed
|
378 |
4. REPEAT: Continue until you have the final answer
|
379 |
-
|
380 |
Always show your reasoning process clearly and provide exact answers as required by GAIA.
|
381 |
""",
|
382 |
-
llm=proj_llm,
|
383 |
-
|
384 |
-
max_steps = 5
|
385 |
)
|
386 |
|
387 |
-
# Créer des outils à partir des agents
|
388 |
-
def analysis_function(query: str, files=None):
|
389 |
-
ctx = Context(analysis_agent)
|
390 |
-
return analysis_agent.run(query, ctx=ctx)
|
391 |
-
|
392 |
-
|
393 |
-
def code_function(query: str):
|
394 |
-
ctx = Context(code_agent)
|
395 |
-
return code_agent.run(query, ctx=ctx)
|
396 |
-
|
397 |
analysis_tool = FunctionTool.from_defaults(
|
398 |
fn=analysis_function,
|
399 |
name="AnalysisAgent",
|
400 |
-
description="Advanced multimodal analysis
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
401 |
)
|
402 |
|
|
|
403 |
code_tool = FunctionTool.from_defaults(
|
404 |
fn=code_function,
|
405 |
name="CodeAgent",
|
406 |
-
description="Advanced
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
)
|
408 |
|
409 |
class EnhancedGAIAAgent:
|
@@ -428,7 +455,6 @@ class EnhancedGAIAAgent:
|
|
428 |
3. OBSERVE: Review results from specialist tools
|
429 |
4. REPEAT: Continue until you have the final answer. If you give a final answer, FORMAT: Ensure answer is EXACT GAIA format (number only, word only, etc.)
|
430 |
|
431 |
-
|
432 |
IMPORTANT: Use tools strategically - only when their specific expertise is needed.
|
433 |
For simple questions, you can answer directly without using any tools.
|
434 |
|
|
|
9 |
from llama_index.readers.file import PDFReader, DocxReader, CSVReader, ImageReader
|
10 |
import os
|
11 |
from typing import List, Dict, Any
|
|
|
12 |
from llama_index.core.tools.ondemand_loader_tool import OnDemandLoaderTool
|
13 |
from llama_index.tools.arxiv import ArxivToolSpec
|
14 |
import duckduckgo_search as ddg
|
|
|
19 |
from llama_index.core.callbacks.base import CallbackManager
|
20 |
from llama_index.core.callbacks.llama_debug import LlamaDebugHandler
|
21 |
from llama_index.core import Settings
|
22 |
+
from llama_index.core.agent.workflow import CodeActAgent
|
23 |
|
24 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
25 |
from llama_index.llms.huggingface import HuggingFaceLLM
|
|
|
240 |
)
|
241 |
|
242 |
|
243 |
+
from llama_index.tools.arxiv import ArxivToolSpec
|
244 |
+
from llama_index.tools.duckduckgo import DuckDuckGoSearchToolSpec
|
245 |
+
|
246 |
class IntelligentSourceRouter:
|
247 |
def __init__(self):
|
248 |
+
# Initialize ArXiv and DuckDuckGo as LlamaIndex tools
|
249 |
+
self.arxiv_tool = ArxivToolSpec().to_tool_list()[0]
|
250 |
+
self.duckduckgo_tool = DuckDuckGoSearchToolSpec().to_tool_list()[0]
|
251 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
def detect_intent_and_route(self, query: str) -> str:
|
253 |
+
# Use your LLM to decide between arxiv and web_search
|
254 |
intent_prompt = f"""
|
255 |
Analyze this query and determine if it's scientific research or general information:
|
256 |
Query: "{query}"
|
|
|
257 |
Choose ONE source:
|
258 |
- arxiv: For scientific research, academic papers, technical studies, algorithms, experiments
|
259 |
- web_search: For all other information (current events, general facts, weather, how-to guides, etc.)
|
|
|
260 |
Respond with ONLY "arxiv" or "web_search".
|
261 |
"""
|
262 |
+
response = text_llm.complete(intent_prompt)
|
|
|
263 |
selected_source = response.text.strip().lower()
|
264 |
+
|
|
|
265 |
results = [f"**Query**: {query}", f"**Selected Source**: {selected_source}", "="*50]
|
|
|
266 |
try:
|
267 |
if selected_source == 'arxiv':
|
268 |
+
result = self.arxiv_tool.call(query=query, max_results=3)
|
269 |
results.append(f"**ArXiv Research:**\n{result}")
|
270 |
+
else:
|
271 |
+
result = self.duckduckgo_tool.call(query=query, max_results=5)
|
272 |
+
# Format results if needed
|
273 |
+
if isinstance(result, list):
|
274 |
+
formatted = []
|
275 |
+
for i, r in enumerate(result, 1):
|
276 |
+
formatted.append(
|
277 |
+
f"{i}. **{r.get('title', '')}**\n URL: {r.get('href', '')}\n {r.get('body', '')}"
|
278 |
+
)
|
279 |
+
result = "\n".join(formatted)
|
280 |
+
results.append(f"**Web Search Results:**\n{result}")
|
|
|
281 |
except Exception as e:
|
282 |
results.append(f"**Search failed**: {str(e)}")
|
283 |
+
return "\n\n".join(results)
|
284 |
+
|
285 |
+
class IntelligentSourceRouter:
|
286 |
+
def __init__(self):
|
287 |
+
# Initialize Arxiv and DuckDuckGo tools
|
288 |
+
self.arxiv_tool = ArxivToolSpec().to_tool_list()[0]
|
289 |
+
self.duckduckgo_tool = DuckDuckGoSearchToolSpec().to_tool_list()[0]
|
290 |
+
|
291 |
+
def detect_intent_and_extract_content(self, query: str, max_results: int = 3) -> str:
|
292 |
+
# Use your LLM to decide between arxiv and web_search
|
293 |
+
intent_prompt = f"""
|
294 |
+
Analyze this query and determine if it's scientific research or general information:
|
295 |
+
Query: "{query}"
|
296 |
+
Choose ONE source:
|
297 |
+
- arxiv: For scientific research, academic papers, technical studies, algorithms, experiments
|
298 |
+
- web_search: For all other information (current events, general facts, weather, how-to guides, etc.)
|
299 |
+
Respond with ONLY "arxiv" or "web_search".
|
300 |
+
"""
|
301 |
+
response = text_llm.complete(intent_prompt)
|
302 |
+
selected_source = response.text.strip().lower()
|
303 |
+
|
304 |
+
results = [f"**Query**: {query}", f"**Selected Source**: {selected_source}", "="*50]
|
305 |
+
try:
|
306 |
+
if selected_source == 'arxiv':
|
307 |
+
# Extract abstracts and paper summaries (deep content)
|
308 |
+
arxiv_results = self.arxiv_tool.call(query=query, max_results=max_results)
|
309 |
+
results.append(f"**Extracted ArXiv Content:**\n{arxiv_results}")
|
310 |
+
else:
|
311 |
+
# DuckDuckGo returns a list of dicts with 'href', 'title', 'body'
|
312 |
+
web_results = self.duckduckgo_tool.call(query=query, max_results=max_results)
|
313 |
+
if isinstance(web_results, list):
|
314 |
+
formatted = []
|
315 |
+
for i, r in enumerate(web_results, 1):
|
316 |
+
formatted.append(
|
317 |
+
f"{i}. **{r.get('title', '')}**\n URL: {r.get('href', '')}\n {r.get('body', '')}"
|
318 |
+
)
|
319 |
+
web_content = "\n".join(formatted)
|
320 |
+
else:
|
321 |
+
web_content = str(web_results)
|
322 |
+
results.append(f"**Extracted Web Content:**\n{web_content}")
|
323 |
+
except Exception as e:
|
324 |
+
results.append(f"**Extraction failed**: {str(e)}")
|
325 |
return "\n\n".join(results)
|
326 |
|
327 |
# Initialize router
|
328 |
intelligent_router = IntelligentSourceRouter()
|
329 |
|
330 |
# Create enhanced research tool
|
331 |
+
def enhanced_smart_research_tool(query: str, task_context: str = "", max_results: int = 3) -> str:
|
332 |
full_query = f"{query} {task_context}".strip()
|
333 |
+
return intelligent_router.detect_intent_and_extract_content(full_query, max_results=max_results)
|
334 |
|
335 |
research_tool = FunctionTool.from_defaults(
|
336 |
fn=enhanced_smart_research_tool,
|
337 |
+
name="Research Tool",
|
338 |
+
description="""Intelligent research specialist that automatically routes between scientific and general sources. Use this tool when you need:
|
339 |
+
|
340 |
+
**Scientific Research (ArXiv):**
|
341 |
+
- Academic papers, research studies, technical algorithms
|
342 |
+
- Scientific experiments, theories, mathematical concepts
|
343 |
+
- Recent developments in AI, ML, physics, chemistry, etc.
|
344 |
+
|
345 |
+
**General Research (Web + Content Extraction):**
|
346 |
+
- Current events, news, real-time information
|
347 |
+
- Biographical information, company details, locations
|
348 |
+
- How-to guides, technical documentation
|
349 |
+
- Weather data, sports results, cultural information
|
350 |
+
- Product specifications, reviews, comparisons
|
351 |
+
|
352 |
+
**Automatic Features:**
|
353 |
+
- Intelligently selects between ArXiv and web search
|
354 |
+
- Extracts full content from web pages (not just snippets)
|
355 |
+
- Provides source attribution and detailed information
|
356 |
+
|
357 |
+
**When to use:** Questions requiring external knowledge not in your training data, current events, scientific research, or factual verification.
|
358 |
+
|
359 |
+
**Input format:** Provide the research query with any relevant context."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
360 |
)
|
361 |
|
362 |
+
code_agent = CodeActAgent(
|
|
|
363 |
name="CodeAgent",
|
364 |
description="Advanced calculations, data processing, and final answer synthesis using ReAct reasoning",
|
365 |
system_prompt="""
|
366 |
You are a coding and reasoning specialist using ReAct methodology.
|
367 |
+
|
368 |
For each task:
|
369 |
1. THINK: Analyze what needs to be calculated or processed
|
370 |
2. ACT: Execute appropriate code or calculations
|
371 |
3. OBSERVE: Review results and determine if more work is needed
|
372 |
4. REPEAT: Continue until you have the final answer
|
373 |
+
|
374 |
Always show your reasoning process clearly and provide exact answers as required by GAIA.
|
375 |
""",
|
376 |
+
llm=proj_llm, # Your language model instance
|
377 |
+
max_steps=5 # Optional: limit the number of reasoning steps
|
|
|
378 |
)
|
379 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
380 |
analysis_tool = FunctionTool.from_defaults(
|
381 |
fn=analysis_function,
|
382 |
name="AnalysisAgent",
|
383 |
+
description="""Advanced multimodal document analysis specialist. Use this tool when you need to:
|
384 |
+
|
385 |
+
**Document Processing:**
|
386 |
+
- Analyze PDF, Word, CSV, or image files provided with the question
|
387 |
+
- Extract specific information from tables, charts, or structured documents
|
388 |
+
- Cross-reference information across multiple documents
|
389 |
+
- Perform semantic search within document collections
|
390 |
+
|
391 |
+
**Content Analysis:**
|
392 |
+
- Summarize long documents or extract key facts
|
393 |
+
- Find specific data points, numbers, or text within files
|
394 |
+
- Analyze visual content in images (charts, graphs, diagrams)
|
395 |
+
- Compare information between different document sources
|
396 |
+
|
397 |
+
**When to use:** Questions involving file attachments, document analysis, data extraction from PDFs/images, or when you need to process structured/unstructured content.
|
398 |
+
|
399 |
+
**Input format:** Provide the query and mention any relevant files or context."""
|
400 |
)
|
401 |
|
402 |
+
|
403 |
code_tool = FunctionTool.from_defaults(
|
404 |
fn=code_function,
|
405 |
name="CodeAgent",
|
406 |
+
description="""Advanced computational specialist using ReAct reasoning. Use this tool when you need:
|
407 |
+
|
408 |
+
**Mathematical Calculations:**
|
409 |
+
- Complex arithmetic, algebra, statistics, probability
|
410 |
+
- Unit conversions, percentage calculations
|
411 |
+
- Financial calculations (interest, loans, investments)
|
412 |
+
- Scientific calculations (physics, chemistry formulas)
|
413 |
+
|
414 |
+
**Data Processing:**
|
415 |
+
- Parsing and analyzing numerical data
|
416 |
+
- String manipulation and text processing
|
417 |
+
- Date/time calculations and conversions
|
418 |
+
- List operations, sorting, filtering
|
419 |
+
|
420 |
+
**Logical Operations:**
|
421 |
+
- Step-by-step problem solving with code
|
422 |
+
- Verification of calculations or logic
|
423 |
+
- Pattern analysis and data validation
|
424 |
+
- Algorithm implementation for specific problems
|
425 |
+
|
426 |
+
**Programming Tasks:**
|
427 |
+
- Code generation for specific computational needs
|
428 |
+
- Data structure manipulation
|
429 |
+
- Regular expression operations
|
430 |
+
|
431 |
+
**When to use:** Questions requiring precise calculations, data manipulation, logical reasoning with code, or when you need to verify numerical results.
|
432 |
+
|
433 |
+
**Input format:** Describe the calculation or processing task clearly, including any specific requirements or constraints."""
|
434 |
)
|
435 |
|
436 |
class EnhancedGAIAAgent:
|
|
|
455 |
3. OBSERVE: Review results from specialist tools
|
456 |
4. REPEAT: Continue until you have the final answer. If you give a final answer, FORMAT: Ensure answer is EXACT GAIA format (number only, word only, etc.)
|
457 |
|
|
|
458 |
IMPORTANT: Use tools strategically - only when their specific expertise is needed.
|
459 |
For simple questions, you can answer directly without using any tools.
|
460 |
|