Abbasid's picture
Update app.py
3f109c4 verified
raw
history blame
16 kB
from smolagents import CodeAgent, DuckDuckGoSearchTool, OpenAIServerModel, tool, FinalAnswerTool, VisitWebpageTool, GradioUI, LiteLLMModel
import requests
import pytz
from typing import Optional, Tuple, Union, Any # Added Any
import re
# from google.colab import userdata # Assuming Colab environment
import io
import contextlib
import sys
import traceback
import os
@tool
def parse_height_from_text(
text: str,
prefer_units: str = "cm",
max_expected: float = 1000.0
) -> Optional[float]:
"""
Extracts and converts the FIRST valid height measurement found in a given text string into centimeters.
**Usage Workflow:**
1. Use this FIRST on the initial user query to get the user's height. Store this value.
2. LATER, after getting web search results, you might use this again on individual search result snippets
if they contain height information (e.g., "Character X is 6'2\" tall").
Args:
text: Input text containing potential height measurements (can be user query or web search snippet).
prefer_units: Preferred unit system ('cm', 'm', 'ft', 'in') if units are ambiguous in the text. Default is 'cm'.
max_expected: Safety limit to ignore potentially nonsensical values during parsing (in cm).
Returns:
float | None: Height in centimeters if a valid measurement is found and parsed, otherwise None.
"""
height_pattern = r"""
(?:^|\b|(?<=\s))(\d+\.?\d*)\s*(?:(cm|centi.*)|(m|meters?|metres)|(ft|feet|')|(in|inches?|"))\b
"""
matches = re.finditer(height_pattern, text, re.IGNORECASE | re.VERBOSE | re.UNICODE)
unit_conversion = {"cm": 1.0, "m": 100.0, "ft": 30.48, "in": 2.54}
for match in matches:
try:
value = float(match.group(1))
raw_unit = next((g for g in match.groups()[1:] if g), "").lower()
if any(u in raw_unit for u in ["cm", "centi"]): unit = "cm"
elif any(u in raw_unit for u in ["m", "meter", "metre"]): unit = "m"
elif any(u in raw_unit for u in ["ft", "feet", "'"]): unit = "ft"
elif any(u in raw_unit for u in ["in", "inch", "\""]): unit = "in"
else: unit = prefer_units
converted = value * unit_conversion[unit]
if 0.1 < converted < max_expected: return round(converted, 2)
except (ValueError, KeyError, TypeError): continue
return None
@tool
def create_comparison_statement(
target: str,
user_height: float,
reference_height: float,
) -> str:
"""
Creates ONE human-readable comparison statement based on height proximity. Output format example:
"👤 You're almost the same height as Sherlock Holmes! (185.0 cm vs 183.0 cm)"
**Usage Workflow:**
1. Call this tool *AFTER* finding a target name, extracting their height, and validating it (e.g., `if 50 < reference_height < 250:`).
2. Call this for *each* validated target you want to include.
3. Collect the string outputs and combine them for the final answer.
Args:
target: The name of the character/object/person being compared against (extracted from search results).
user_height: The user's height in centimeters.
reference_height: The specific reference target's height in centimeters (parsed and VALIDATED from search results).
Returns:
str: A single formatted comparison string indicating height similarity.
"""
diff = user_height - reference_height
abs_diff = abs(diff)
comparison_phrase = ""
# Define thresholds for different phrases (adjust as needed)
exact_threshold = 1.0 # Within 1 cm difference
close_threshold = 4.0 # Within 4 cm difference
if abs_diff <= exact_threshold:
comparison_phrase = f"You're exactly the same height as {target}!"
elif abs_diff <= close_threshold:
if diff > 0:
comparison_phrase = f"You're slightly taller than {target}!"
else:
comparison_phrase = f"You're slightly shorter than {target}!"
elif diff > 0: # User is significantly taller
comparison_phrase = f"You're noticeably taller than {target}."
else: # User is significantly shorter
comparison_phrase = f"You're noticeably shorter than {target}."
# Use a simple emoji or none
emoji = "👤"
return (
f"{emoji} {comparison_phrase} "
f"({user_height:.1f} cm vs {reference_height:.1f} cm)"
)
# # --- Instantiate Model ---
# try:
# OR_API_KEY = userdata.get("OR_TOKEN")
# if not OR_API_KEY: raise ValueError("OR_TOKEN not found in Colab userdata.")
# except (ImportError, NameError):
# import os
# OR_API_KEY = os.environ.get("OR_TOKEN")
# if not OR_API_KEY: raise ValueError("API Key OR_TOKEN not found in environment variables.")
# model = OpenAIServerModel(
# model_id='qwen/qwen-2.5-coder-32b-instruct:free',
# api_base='https://openrouter.ai/api/v1',
# api_key=userdata.get("OR_TOKEN"),
# )
# Replace all calls to HfApiModel
llm_model = LiteLLMModel(
model_id="gemini/gemini-2.0-flash", # you can see other model names here: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models. It is important to prefix the name with "gemini/"
api_key=os.environ.get('GEM_TOKEN'),
max_tokens=8192
)
# --- Task Generation Function (No change needed here) ---
# It generates the *instructions* for the agent run
def create_height_comparison_task(user_query: str) -> str:
"""Combines user query with detailed instructions encouraging diverse searches and robust parsing."""
escaped_query = user_query.replace("'", "\\'") # Simple escaping
instructions = f"""
TASK: Analyze the user query '{escaped_query}' and perform the following steps to find height comparisons with **diverse figures (people, characters)**:
1. **Parse User Height:** Use `parse_height_from_text` on the user query ('{escaped_query}') to get the user's height in cm. Print and store it. If none found, use `final_answer` to ask for clarification like "Please provide your height clearly (e.g., '180 cm', '5 ft 11 in').".
2. **Web Search (Diverse Queries):** If height found, use `web_search` to find **fictional characters, historical figures, scientists, artists, athletes, and other interesting people** of similar height. Formulate 2-3 specific queries using the user's height in cm (e.g., if user height is 180cm, search for `"historical figures 180 cm tall"`, `"celebrities around 180cm height"`, `"fictional characters exactly 180 cm"`). Print the search results clearly.
3. **Extract & Validate from Search Results:** CRITICAL STEP. Read the `web_search` Observation snippets carefully.
* Identify potential (Name, Height String) pairs. Prioritize clear mentions of height linked to a name.
* For each potential pair:
* Use `parse_height_from_text` on the relevant part of the search snippet string containing the height info. Store the result in cm (e.g., `extracted_cm`).
* **Validate using Python code:** Check if `extracted_cm` is NOT `None` AND if it's within a reasonable human range (e.g., `if extracted_cm is not None and 50 < extracted_cm < 250:`).
* Collect valid (Name, Validated Height cm) pairs into a Python list. Print this list. Aim for diverse examples.
4. **Generate Multiple Comparisons:** Check the validated matches list.
* If empty after searching, use `final_answer` stating no relevant matches were found for that height.
* If matches exist, select **up to 3-4 diverse ones**.
* Create an empty list `comparison_outputs = []`.
* **Loop** through the selected matches. For each (name, ref_height_cm), call `create_comparison_statement(target=name, user_height=USER_HEIGHT_CM, reference_height=ref_height_cm)`. Append the resulting string to `comparison_outputs`.
5. **Final Answer:** Combine the generated strings from `comparison_outputs` into a single response (e.g., separated by newlines: `"\\n".join(comparison_outputs)`). Add a brief introductory sentence like "Here are some figures with similar heights:". Return the complete message using `final_answer`.
Follow Thought-Code-Observation meticulously. Handle `None` returns from `parse_height_from_text` gracefully in your Python code logic. Use the tools as described in their docstrings.
"""
return instructions
# --- Define the Subclassed Agent ---
class HeightComparisonAgent(CodeAgent):
"""
An agent that intercepts the user query in the run method,
transforms it into a detailed task using create_height_comparison_task,
and then executes the detailed task using the parent CodeAgent's run method.
This allows GradioUI to monitor the execution of the *detailed* task.
"""
def run(self, task: str, **kwargs: Any) -> str:
"""
Overrides the default run method.
'task' received here is expected to be the raw user query from GradioUI.
"""
user_query = task # Assume the input 'task' is the user query
print(f"[HeightComparisonAgent] Intercepted run call with user query: '{user_query}'")
if not user_query or not user_query.strip():
return "Please enter a valid query." # Handle empty input
# 1. Generate the detailed task description using the helper function
detailed_task = create_height_comparison_task(user_query)
print(f"[HeightComparisonAgent] Generated detailed task (first 200 chars): {detailed_task[:200]}...")
# 2. Call the *parent* class's run method with the DETAILED task
# This is the core step. super().run() executes the actual agent logic
# that GradioUI is presumably monitoring via its verbose output.
print(f"[HeightComparisonAgent] Calling super().run() with the detailed task...")
try:
# Pass the generated 'detailed_task' as the 'task' argument to the parent's run method
final_result = super().run(task=detailed_task, **kwargs)
print(f"[HeightComparisonAgent] super().run() finished.")
# GradioUI should display the final_result automatically
return final_result
except Exception as e:
print(f"[HeightComparisonAgent] Error during super().run(): {e}")
traceback.print_exc()
# Return a user-friendly error message
return f"An error occurred while processing your request: {e}"
# --- Instantiate the Subclassed Agent ---
# IMPORTANT: Use the HeightComparisonAgent class, not CodeAgent directly.
# Set verbosity_level=3 so the parent's run method (super().run) generates the verbose output.
# --- Instantiate the Agent ---
height_agent = None
initialization_error_message = None # <<< Make sure this line is BEFORE the if
if llm_model is not None:
try:
height_agent = HeightComparisonAgent(
tools=[DuckDuckGoSearchTool(), VisitWebpageTool(), parse_height_from_text, create_comparison_statement, FinalAnswerTool()],
model=llm_model,
verbosity_level=3, # <<< ESSENTIAL for capturing reasoning steps
max_steps=20,
)
print("--- HeightComparisonAgent initialized successfully. ---")
except Exception as e:
# Store the error if agent creation fails even with a model
initialization_error_message = f"ERROR: Failed to initialize HeightComparisonAgent: {e}\n{traceback.format_exc()}"
print(initialization_error_message)
height_agent = None # Ensure agent is None on error
else:
# Store the error if the LLM model itself failed to initialize
initialization_error_message = (
"ERROR: Could not initialize any Language Model backend.\n\n"
f"Please check the Space logs (check the 'Logs' tab above the app).\n"
f"Verify that at least one of these secrets is correctly set in Space Settings -> Secrets:\n"
f"Also ensure necessary libraries are in requirements.txt."
)
print(initialization_error_message)
# height_agent is already None
# --- Wrapper Function to Run Agent and Capture Output ---
def run_agent_wrapper(query: str) -> Tuple[str, str]:
"""
Runs the height_agent and captures its stdout (reasoning steps).
Returns (reasoning_log, final_answer).
"""
# Access the global variables
global height_agent, initialization_error_message
if height_agent is None:
# If agent initialization failed, return the stored error message
return (initialization_error_message or "Agent not initialized (unknown error).",
"Agent failed to initialize. See reasoning log for details.")
print(f"\n--- Running agent for query: '{query}' ---") # Log to console
log_stream = io.StringIO()
final_answer = "Agent execution did not complete." # Default message
try:
# Redirect stdout to capture prints from agent.run() (due to verbosity=3)
with contextlib.redirect_stdout(log_stream):
# Make sure to call the run method of the specific agent instance
final_answer = height_agent.run(query) # Pass the raw query
print("\n--- Agent execution finished successfully. ---") # Add marker to log
except Exception as e:
print(f"\n--- Error during agent execution wrapper: {e} ---") # Log to console
# Print exception details *into the captured log*
print("\n\n******** ERROR DURING EXECUTION ********\n", file=log_stream)
traceback.print_exc(file=log_stream)
final_answer = f"An error occurred during processing. See reasoning log. Error: {e}"
finally:
reasoning_log = log_stream.getvalue()
log_stream.close()
print("--- Finished capturing stdout. ---") # Log to console
return reasoning_log, final_answer
# --- Build Gradio Interface Manually with gr.Blocks ---
print("--- Building Gradio Interface with gr.Blocks ---")
# Make sure theme is applied correctly if desired
# theme = gr.themes.Default() # Or another theme
# with gr.Blocks(theme=theme, css="footer {visibility: hidden}") as demo:
with gr.Blocks(css="footer {visibility: hidden}") as demo: # Hides the default footer
gr.Markdown("# Height Comparison Agent")
gr.Markdown("Enter your height (e.g., '180 cm', '5ft 11in') to find characters/figures of similar height.")
with gr.Row():
with gr.Column(scale=1):
query_input = gr.Textbox(
label="Your Query (including height)",
placeholder="e.g., I am 175cm tall",
lines=2 # Allow slightly more room for input
)
submit_button = gr.Button("Compare Heights", variant="primary")
with gr.Column(scale=2):
final_answer_output = gr.Textbox(
label="Final Answer",
interactive=False,
lines=5
)
gr.Markdown("## Agent Reasoning Steps")
# Use gr.Code for better formatting of logs, especially if they contain code blocks
reasoning_output = gr.Code(
label="Reasoning Log",
language="text", # Use 'markdown' if logs might contain markdown
interactive=False,
lines=20
)
# Link components: When button is clicked, call wrapper, update outputs
submit_button.click(
fn=run_agent_wrapper, # Function to call
inputs=[query_input], # Component(s) providing input
outputs=[reasoning_output, final_answer_output] # Components to update
# Ensure the order matches the return tuple from run_agent_wrapper: (log, answer)
)
# Add an example input
gr.Examples(
examples=[
"I am 188cm tall",
"How tall is someone who is 5 foot 8 inches?",
"My height is 1.65m",
],
inputs=query_input
)
# --- Launch Gradio ---
print("--- Launching Gradio demo ---")
demo.launch(ssr=False) # ssr=False recommended, share=True not needed for Spaces