Update app.py
Browse files
app.py
CHANGED
|
@@ -9,7 +9,6 @@ import ast
|
|
| 9 |
import operator as op
|
| 10 |
import wikipedia
|
| 11 |
|
| 12 |
-
# Tool implementations
|
| 13 |
class Tool:
|
| 14 |
def __init__(self, name: str, description: str, func):
|
| 15 |
self.name = name
|
|
@@ -32,7 +31,6 @@ def duckduckgo_search(query: str) -> str:
|
|
| 32 |
response = requests.get(url, params=params, timeout=10)
|
| 33 |
data = response.json()
|
| 34 |
|
| 35 |
-
# Get abstract or first related topic
|
| 36 |
if data.get('Abstract'):
|
| 37 |
return f"Search result: {data['Abstract']}"
|
| 38 |
elif data.get('RelatedTopics') and len(data['RelatedTopics']) > 0:
|
|
@@ -50,7 +48,6 @@ def wikipedia_search(query: str) -> str:
|
|
| 50 |
"""Search Wikipedia for information."""
|
| 51 |
try:
|
| 52 |
wikipedia.set_lang("en")
|
| 53 |
-
# Get summary
|
| 54 |
summary = wikipedia.summary(query, sentences=3, auto_suggest=True)
|
| 55 |
return f"Wikipedia: {summary}"
|
| 56 |
except wikipedia.exceptions.DisambiguationError as e:
|
|
@@ -80,7 +77,6 @@ def get_weather(location: str) -> str:
|
|
| 80 |
|
| 81 |
def calculate(expression: str) -> str:
|
| 82 |
"""Safely evaluate mathematical expressions."""
|
| 83 |
-
# Supported operators
|
| 84 |
operators = {
|
| 85 |
ast.Add: op.add,
|
| 86 |
ast.Sub: op.sub,
|
|
@@ -99,7 +95,6 @@ def calculate(expression: str) -> str:
|
|
| 99 |
elif isinstance(node, ast.UnaryOp):
|
| 100 |
return operators[type(node.op)](eval_expr(node.operand))
|
| 101 |
elif isinstance(node, ast.Call):
|
| 102 |
-
# Support basic math functions
|
| 103 |
if node.func.id == 'abs':
|
| 104 |
return abs(eval_expr(node.args[0]))
|
| 105 |
elif node.func.id == 'round':
|
|
@@ -108,9 +103,7 @@ def calculate(expression: str) -> str:
|
|
| 108 |
raise TypeError(node)
|
| 109 |
|
| 110 |
try:
|
| 111 |
-
# Clean the expression
|
| 112 |
expression = expression.strip()
|
| 113 |
-
# Parse and evaluate
|
| 114 |
node = ast.parse(expression, mode='eval')
|
| 115 |
result = eval_expr(node.body)
|
| 116 |
return f"Result: {result}"
|
|
@@ -120,7 +113,6 @@ def calculate(expression: str) -> str:
|
|
| 120 |
def python_repl(code: str) -> str:
|
| 121 |
"""Execute safe Python code (limited to basic operations)."""
|
| 122 |
try:
|
| 123 |
-
# Whitelist of safe builtins
|
| 124 |
safe_builtins = {
|
| 125 |
'abs': abs, 'round': round, 'min': min, 'max': max,
|
| 126 |
'sum': sum, 'len': len, 'range': range, 'list': list,
|
|
@@ -129,23 +121,18 @@ def python_repl(code: str) -> str:
|
|
| 129 |
'sorted': sorted, 'reversed': reversed,
|
| 130 |
}
|
| 131 |
|
| 132 |
-
# Create restricted namespace
|
| 133 |
namespace = {'__builtins__': safe_builtins}
|
| 134 |
|
| 135 |
-
# Capture output
|
| 136 |
from io import StringIO
|
| 137 |
import sys
|
| 138 |
old_stdout = sys.stdout
|
| 139 |
sys.stdout = StringIO()
|
| 140 |
|
| 141 |
-
# Execute code
|
| 142 |
exec(code, namespace)
|
| 143 |
|
| 144 |
-
# Get output
|
| 145 |
output = sys.stdout.getvalue()
|
| 146 |
sys.stdout = old_stdout
|
| 147 |
|
| 148 |
-
# Also get any variables that were set
|
| 149 |
result_vars = {k: v for k, v in namespace.items() if k != '__builtins__' and not k.startswith('_')}
|
| 150 |
|
| 151 |
result = output if output else str(result_vars) if result_vars else "Code executed successfully (no output)"
|
|
@@ -153,7 +140,6 @@ def python_repl(code: str) -> str:
|
|
| 153 |
except Exception as e:
|
| 154 |
return f"Python error: {str(e)}"
|
| 155 |
|
| 156 |
-
# Define tools
|
| 157 |
TOOLS = [
|
| 158 |
Tool(
|
| 159 |
name="duckduckgo_search",
|
|
@@ -182,14 +168,12 @@ TOOLS = [
|
|
| 182 |
),
|
| 183 |
]
|
| 184 |
|
| 185 |
-
# Create tool descriptions for prompt
|
| 186 |
def get_tool_descriptions() -> str:
|
| 187 |
descriptions = []
|
| 188 |
for tool in TOOLS:
|
| 189 |
descriptions.append(f"- {tool.name}: {tool.description}")
|
| 190 |
return "\n".join(descriptions)
|
| 191 |
|
| 192 |
-
# Agent prompts
|
| 193 |
THINK_ONLY_PROMPT = """You are a helpful AI assistant. You solve problems by thinking through them step-by-step.
|
| 194 |
|
| 195 |
For each question:
|
|
@@ -297,11 +281,10 @@ def think_only_mode(question: str) -> Generator[str, None, None]:
|
|
| 297 |
messages = [{"role": "user", "content": prompt}]
|
| 298 |
|
| 299 |
yield "**Mode: Think-Only (Chain-of-Thought)**\n\n"
|
| 300 |
-
yield "
|
| 301 |
|
| 302 |
response = call_llm(messages, temperature=0.7, max_tokens=800)
|
| 303 |
|
| 304 |
-
# Parse and format the response
|
| 305 |
lines = response.split('\n')
|
| 306 |
for line in lines:
|
| 307 |
if line.strip():
|
|
@@ -329,33 +312,29 @@ def act_only_mode(question: str, max_iterations: int = 5) -> Generator[str, None
|
|
| 329 |
|
| 330 |
response = call_llm(messages, temperature=0.5, max_tokens=300)
|
| 331 |
|
| 332 |
-
# Check for final answer
|
| 333 |
if 'Answer:' in response:
|
| 334 |
answer_match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
|
| 335 |
if answer_match:
|
| 336 |
-
yield f"
|
| 337 |
break
|
| 338 |
|
| 339 |
-
# Parse action
|
| 340 |
action_name, action_input = parse_action(response)
|
| 341 |
|
| 342 |
if action_name and action_input:
|
| 343 |
-
yield f"
|
| 344 |
-
yield f"
|
| 345 |
|
| 346 |
-
# Execute tool
|
| 347 |
observation = call_tool(action_name, action_input)
|
| 348 |
-
yield f"
|
| 349 |
|
| 350 |
-
# Add to conversation
|
| 351 |
messages.append({"role": "assistant", "content": response})
|
| 352 |
messages.append({"role": "user", "content": f"Observation: {observation}\n\nContinue with another action or provide the final answer."})
|
| 353 |
else:
|
| 354 |
-
yield f"
|
| 355 |
break
|
| 356 |
|
| 357 |
if iteration >= max_iterations:
|
| 358 |
-
yield "
|
| 359 |
|
| 360 |
yield "\n---\n**Mode completed**\n"
|
| 361 |
|
|
@@ -374,44 +353,37 @@ def react_mode(question: str, max_iterations: int = 5) -> Generator[str, None, N
|
|
| 374 |
|
| 375 |
response = call_llm(messages, temperature=0.7, max_tokens=400)
|
| 376 |
|
| 377 |
-
# Parse thoughts
|
| 378 |
thought_matches = re.findall(r'Thought:\s*(.+?)(?=\n(?:Action:|Answer:|$))', response, re.IGNORECASE | re.DOTALL)
|
| 379 |
for thought in thought_matches:
|
| 380 |
yield f"**Thought:** {thought.strip()}\n\n"
|
| 381 |
|
| 382 |
-
# Check for final answer
|
| 383 |
if 'Answer:' in response:
|
| 384 |
answer_match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
|
| 385 |
if answer_match:
|
| 386 |
yield f"**Answer:** {answer_match.group(1).strip()}\n\n"
|
| 387 |
break
|
| 388 |
|
| 389 |
-
# Parse action
|
| 390 |
action_name, action_input = parse_action(response)
|
| 391 |
|
| 392 |
if action_name and action_input:
|
| 393 |
-
yield f"
|
| 394 |
yield f"**Action Input:** {action_input}\n\n"
|
| 395 |
|
| 396 |
-
# Execute tool
|
| 397 |
observation = call_tool(action_name, action_input)
|
| 398 |
yield f"**Observation:** {observation}\n\n"
|
| 399 |
|
| 400 |
-
# Add to conversation
|
| 401 |
messages.append({"role": "assistant", "content": response})
|
| 402 |
messages.append({"role": "user", "content": f"Observation: {observation}\n\nThought:"})
|
| 403 |
else:
|
| 404 |
-
# If no action but also no answer, there might be an issue
|
| 405 |
if 'Answer:' not in response:
|
| 406 |
-
yield f"
|
| 407 |
break
|
| 408 |
|
| 409 |
if iteration >= max_iterations:
|
| 410 |
-
yield "
|
| 411 |
|
| 412 |
yield "\n---\n**Mode completed**\n"
|
| 413 |
|
| 414 |
-
# Example questions
|
| 415 |
EXAMPLES = [
|
| 416 |
"What is the capital of France and what's the current weather there?",
|
| 417 |
"Who wrote 'To Kill a Mockingbird' and when was it published?",
|
|
@@ -434,19 +406,17 @@ def run_comparison(question: str, mode: str):
|
|
| 434 |
else:
|
| 435 |
return "Invalid mode selected.", "", ""
|
| 436 |
|
| 437 |
-
# Gradio Interface
|
| 438 |
with gr.Blocks(title="LLM Reasoning Modes Comparison") as demo:
|
| 439 |
gr.Markdown("""
|
| 440 |
-
#
|
| 441 |
|
| 442 |
Compare three reasoning approaches using **openai/gpt-oss-20b**:
|
| 443 |
|
| 444 |
- **Think-Only**: Chain-of-Thought reasoning only (no tools)
|
| 445 |
- **Act-Only**: Tool use only (no explicit reasoning)
|
| 446 |
-
- **ReAct**: Interleaved Thought
|
| 447 |
|
| 448 |
-
|
| 449 |
-
🔍 DuckDuckGo Search | 📚 Wikipedia | 🌤️ Weather API | 🧮 Calculator | 🐍 Python REPL
|
| 450 |
""")
|
| 451 |
|
| 452 |
with gr.Row():
|
|
@@ -461,12 +431,12 @@ with gr.Blocks(title="LLM Reasoning Modes Comparison") as demo:
|
|
| 461 |
value="All (Compare)",
|
| 462 |
label="Select Mode"
|
| 463 |
)
|
| 464 |
-
submit_btn = gr.Button("
|
| 465 |
|
| 466 |
with gr.Column(scale=1):
|
| 467 |
-
gr.Markdown("
|
| 468 |
for idx, example in enumerate(EXAMPLES):
|
| 469 |
-
gr.Button(f"
|
| 470 |
fn=lambda ex=example: ex,
|
| 471 |
outputs=question_input
|
| 472 |
)
|
|
@@ -486,7 +456,6 @@ with gr.Blocks(title="LLM Reasoning Modes Comparison") as demo:
|
|
| 486 |
inputs=[question_input, mode_dropdown],
|
| 487 |
outputs=[think_output, act_output, react_output]
|
| 488 |
)
|
| 489 |
-
|
| 490 |
-
|
| 491 |
if __name__ == "__main__":
|
| 492 |
-
demo.launch()
|
|
|
|
| 9 |
import operator as op
|
| 10 |
import wikipedia
|
| 11 |
|
|
|
|
| 12 |
class Tool:
|
| 13 |
def __init__(self, name: str, description: str, func):
|
| 14 |
self.name = name
|
|
|
|
| 31 |
response = requests.get(url, params=params, timeout=10)
|
| 32 |
data = response.json()
|
| 33 |
|
|
|
|
| 34 |
if data.get('Abstract'):
|
| 35 |
return f"Search result: {data['Abstract']}"
|
| 36 |
elif data.get('RelatedTopics') and len(data['RelatedTopics']) > 0:
|
|
|
|
| 48 |
"""Search Wikipedia for information."""
|
| 49 |
try:
|
| 50 |
wikipedia.set_lang("en")
|
|
|
|
| 51 |
summary = wikipedia.summary(query, sentences=3, auto_suggest=True)
|
| 52 |
return f"Wikipedia: {summary}"
|
| 53 |
except wikipedia.exceptions.DisambiguationError as e:
|
|
|
|
| 77 |
|
| 78 |
def calculate(expression: str) -> str:
|
| 79 |
"""Safely evaluate mathematical expressions."""
|
|
|
|
| 80 |
operators = {
|
| 81 |
ast.Add: op.add,
|
| 82 |
ast.Sub: op.sub,
|
|
|
|
| 95 |
elif isinstance(node, ast.UnaryOp):
|
| 96 |
return operators[type(node.op)](eval_expr(node.operand))
|
| 97 |
elif isinstance(node, ast.Call):
|
|
|
|
| 98 |
if node.func.id == 'abs':
|
| 99 |
return abs(eval_expr(node.args[0]))
|
| 100 |
elif node.func.id == 'round':
|
|
|
|
| 103 |
raise TypeError(node)
|
| 104 |
|
| 105 |
try:
|
|
|
|
| 106 |
expression = expression.strip()
|
|
|
|
| 107 |
node = ast.parse(expression, mode='eval')
|
| 108 |
result = eval_expr(node.body)
|
| 109 |
return f"Result: {result}"
|
|
|
|
| 113 |
def python_repl(code: str) -> str:
|
| 114 |
"""Execute safe Python code (limited to basic operations)."""
|
| 115 |
try:
|
|
|
|
| 116 |
safe_builtins = {
|
| 117 |
'abs': abs, 'round': round, 'min': min, 'max': max,
|
| 118 |
'sum': sum, 'len': len, 'range': range, 'list': list,
|
|
|
|
| 121 |
'sorted': sorted, 'reversed': reversed,
|
| 122 |
}
|
| 123 |
|
|
|
|
| 124 |
namespace = {'__builtins__': safe_builtins}
|
| 125 |
|
|
|
|
| 126 |
from io import StringIO
|
| 127 |
import sys
|
| 128 |
old_stdout = sys.stdout
|
| 129 |
sys.stdout = StringIO()
|
| 130 |
|
|
|
|
| 131 |
exec(code, namespace)
|
| 132 |
|
|
|
|
| 133 |
output = sys.stdout.getvalue()
|
| 134 |
sys.stdout = old_stdout
|
| 135 |
|
|
|
|
| 136 |
result_vars = {k: v for k, v in namespace.items() if k != '__builtins__' and not k.startswith('_')}
|
| 137 |
|
| 138 |
result = output if output else str(result_vars) if result_vars else "Code executed successfully (no output)"
|
|
|
|
| 140 |
except Exception as e:
|
| 141 |
return f"Python error: {str(e)}"
|
| 142 |
|
|
|
|
| 143 |
TOOLS = [
|
| 144 |
Tool(
|
| 145 |
name="duckduckgo_search",
|
|
|
|
| 168 |
),
|
| 169 |
]
|
| 170 |
|
|
|
|
| 171 |
def get_tool_descriptions() -> str:
|
| 172 |
descriptions = []
|
| 173 |
for tool in TOOLS:
|
| 174 |
descriptions.append(f"- {tool.name}: {tool.description}")
|
| 175 |
return "\n".join(descriptions)
|
| 176 |
|
|
|
|
| 177 |
THINK_ONLY_PROMPT = """You are a helpful AI assistant. You solve problems by thinking through them step-by-step.
|
| 178 |
|
| 179 |
For each question:
|
|
|
|
| 281 |
messages = [{"role": "user", "content": prompt}]
|
| 282 |
|
| 283 |
yield "**Mode: Think-Only (Chain-of-Thought)**\n\n"
|
| 284 |
+
yield "Generating thoughts...\n\n"
|
| 285 |
|
| 286 |
response = call_llm(messages, temperature=0.7, max_tokens=800)
|
| 287 |
|
|
|
|
| 288 |
lines = response.split('\n')
|
| 289 |
for line in lines:
|
| 290 |
if line.strip():
|
|
|
|
| 312 |
|
| 313 |
response = call_llm(messages, temperature=0.5, max_tokens=300)
|
| 314 |
|
|
|
|
| 315 |
if 'Answer:' in response:
|
| 316 |
answer_match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
|
| 317 |
if answer_match:
|
| 318 |
+
yield f"**Answer:** {answer_match.group(1).strip()}\n\n"
|
| 319 |
break
|
| 320 |
|
|
|
|
| 321 |
action_name, action_input = parse_action(response)
|
| 322 |
|
| 323 |
if action_name and action_input:
|
| 324 |
+
yield f"**Action:** {action_name}\n"
|
| 325 |
+
yield f"**Action Input:** {action_input}\n\n"
|
| 326 |
|
|
|
|
| 327 |
observation = call_tool(action_name, action_input)
|
| 328 |
+
yield f"**Observation:** {observation}\n\n"
|
| 329 |
|
|
|
|
| 330 |
messages.append({"role": "assistant", "content": response})
|
| 331 |
messages.append({"role": "user", "content": f"Observation: {observation}\n\nContinue with another action or provide the final answer."})
|
| 332 |
else:
|
| 333 |
+
yield f"Could not parse action from response. Response: {response}\n\n"
|
| 334 |
break
|
| 335 |
|
| 336 |
if iteration >= max_iterations:
|
| 337 |
+
yield "**Reached maximum iterations.**\n\n"
|
| 338 |
|
| 339 |
yield "\n---\n**Mode completed**\n"
|
| 340 |
|
|
|
|
| 353 |
|
| 354 |
response = call_llm(messages, temperature=0.7, max_tokens=400)
|
| 355 |
|
|
|
|
| 356 |
thought_matches = re.findall(r'Thought:\s*(.+?)(?=\n(?:Action:|Answer:|$))', response, re.IGNORECASE | re.DOTALL)
|
| 357 |
for thought in thought_matches:
|
| 358 |
yield f"**Thought:** {thought.strip()}\n\n"
|
| 359 |
|
|
|
|
| 360 |
if 'Answer:' in response:
|
| 361 |
answer_match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
|
| 362 |
if answer_match:
|
| 363 |
yield f"**Answer:** {answer_match.group(1).strip()}\n\n"
|
| 364 |
break
|
| 365 |
|
|
|
|
| 366 |
action_name, action_input = parse_action(response)
|
| 367 |
|
| 368 |
if action_name and action_input:
|
| 369 |
+
yield f"**Action:** {action_name}\n"
|
| 370 |
yield f"**Action Input:** {action_input}\n\n"
|
| 371 |
|
|
|
|
| 372 |
observation = call_tool(action_name, action_input)
|
| 373 |
yield f"**Observation:** {observation}\n\n"
|
| 374 |
|
|
|
|
| 375 |
messages.append({"role": "assistant", "content": response})
|
| 376 |
messages.append({"role": "user", "content": f"Observation: {observation}\n\nThought:"})
|
| 377 |
else:
|
|
|
|
| 378 |
if 'Answer:' not in response:
|
| 379 |
+
yield f"No action found. Response: {response}\n\n"
|
| 380 |
break
|
| 381 |
|
| 382 |
if iteration >= max_iterations:
|
| 383 |
+
yield "**Reached maximum iterations.**\n\n"
|
| 384 |
|
| 385 |
yield "\n---\n**Mode completed**\n"
|
| 386 |
|
|
|
|
| 387 |
EXAMPLES = [
|
| 388 |
"What is the capital of France and what's the current weather there?",
|
| 389 |
"Who wrote 'To Kill a Mockingbird' and when was it published?",
|
|
|
|
| 406 |
else:
|
| 407 |
return "Invalid mode selected.", "", ""
|
| 408 |
|
|
|
|
| 409 |
with gr.Blocks(title="LLM Reasoning Modes Comparison") as demo:
|
| 410 |
gr.Markdown("""
|
| 411 |
+
# LLM Reasoning Modes Comparison
|
| 412 |
|
| 413 |
Compare three reasoning approaches using **openai/gpt-oss-20b**:
|
| 414 |
|
| 415 |
- **Think-Only**: Chain-of-Thought reasoning only (no tools)
|
| 416 |
- **Act-Only**: Tool use only (no explicit reasoning)
|
| 417 |
+
- **ReAct**: Interleaved Thought, Action, Observation
|
| 418 |
|
| 419 |
+
Available Tools: DuckDuckGo Search | Wikipedia | Weather API | Calculator | Python REPL
|
|
|
|
| 420 |
""")
|
| 421 |
|
| 422 |
with gr.Row():
|
|
|
|
| 431 |
value="All (Compare)",
|
| 432 |
label="Select Mode"
|
| 433 |
)
|
| 434 |
+
submit_btn = gr.Button("Run", variant="primary", size="lg")
|
| 435 |
|
| 436 |
with gr.Column(scale=1):
|
| 437 |
+
gr.Markdown("**Example Questions**")
|
| 438 |
for idx, example in enumerate(EXAMPLES):
|
| 439 |
+
gr.Button(f"Example {idx+1}", size="sm").click(
|
| 440 |
fn=lambda ex=example: ex,
|
| 441 |
outputs=question_input
|
| 442 |
)
|
|
|
|
| 456 |
inputs=[question_input, mode_dropdown],
|
| 457 |
outputs=[think_output, act_output, react_output]
|
| 458 |
)
|
| 459 |
+
|
|
|
|
| 460 |
if __name__ == "__main__":
|
| 461 |
+
demo.launch()
|