agent_builder / app.py
burtenshaw
first commit
fb09fd5
raw
history blame
17.5 kB
import os
from typing import Optional
import gradio as gr
import requests
from smolagents import CodeAgent, Tool
from smolagents.models import HfApiModel
from smolagents.monitoring import LogLevel
from gradio import ChatMessage
DEFAULT_MODEL = "Qwen/Qwen2.5-Coder-32B-Instruct"
HF_API_TOKEN = os.getenv("HF_TOKEN")
# Tool descriptions for the UI
TOOL_DESCRIPTIONS = {
"Hub Collections": "Add tool collections from Hugging Face Hub.",
"Spaces": "Add tools from Hugging Face Spaces.",
}
def search_spaces(query, limit=1):
"""
Search for Hugging Face Spaces using the API.
Returns the first result or None if no results.
"""
try:
url = f"https://huggingface.co/api/spaces?search={query}&limit={limit}"
response = requests.get(
url, headers={"Authorization": f"Bearer {HF_API_TOKEN}"}
)
response.raise_for_status()
spaces = response.json()
if not spaces:
return None
# Get the first space
space = spaces[0]
space_id = space["id"]
# Extract title and description
title = space_id.split("/")[-1] # Default to the last part of the ID
description = f"Tool from {space_id}"
# Try to get title from different possible locations
if "title" in space:
title = space["title"]
elif "cardData" in space and "title" in space["cardData"]:
title = space["cardData"]["title"]
# Try to get description from different possible locations
if "description" in space:
description = space["description"]
elif "cardData" in space and "description" in space["cardData"]:
description = space["cardData"]["description"]
return {
"id": space_id,
"title": title,
"description": description,
}
except Exception as e:
print(f"Error searching spaces: {e}")
return None
def get_space_metadata(space_id):
"""
Get metadata for a specific Hugging Face Space.
"""
try:
url = f"https://huggingface.co/api/spaces/{space_id}"
response = requests.get(
url, headers={"Authorization": f"Bearer {HF_API_TOKEN}"}
)
response.raise_for_status()
space = response.json()
# Extract title and description from the space data
# The structure can vary, so we need to handle different cases
title = space_id
description = f"Tool from {space_id}"
# Try to get title from different possible locations
if "title" in space:
title = space["title"]
elif "cardData" in space and "title" in space["cardData"]:
title = space["cardData"]["title"]
else:
# Use the last part of the space_id as a fallback title
title = space_id.split("/")[-1]
# Try to get description from different possible locations
if "description" in space:
description = space["description"]
elif "cardData" in space and "description" in space["cardData"]:
description = space["cardData"]["description"]
return {
"id": space_id,
"title": title,
"description": description,
}
except Exception as e:
print(f"Error getting space metadata: {e}")
return None
def create_agent(model_name, space_tools=None):
"""
Create a CodeAgent with the specified model and tools.
"""
if not space_tools:
space_tools = []
try:
# Convert space tools to Tool objects
tools = []
for tool_info in space_tools:
space_id = tool_info["id"]
tool = Tool.from_space(
space_id,
name=tool_info.get("name", space_id),
description=tool_info.get("description", ""),
)
tools.append(tool)
# Initialize the HfApiModel with the model name
model = HfApiModel(model_id=model_name, token=HF_API_TOKEN)
# Create the agent with the tools and additional imports
agent = CodeAgent(
tools=tools,
model=model,
additional_authorized_imports=["PIL", "requests"],
verbosity_level=LogLevel.DEBUG, # Set higher verbosity for detailed logs
)
print(f"Agent created successfully with {len(tools)} tools")
return agent
except Exception as e:
print(f"Error creating agent: {e}")
# Try with a fallback model if the specified one fails
try:
print("Trying fallback model...")
fallback_model = HfApiModel(
model_id="Qwen/Qwen2.5-Coder-7B-Instruct", token=HF_API_TOKEN
)
agent = CodeAgent(
tools=tools,
model=fallback_model,
additional_authorized_imports=["PIL", "requests"],
verbosity_level=LogLevel.DEBUG, # Set higher verbosity for detailed logs
)
print("Agent created successfully with fallback model")
return agent
except Exception as e:
print(f"Error creating agent: {e}")
return None
# Event handler functions
def on_search_spaces(query):
if not query:
return "Please enter a search term.", "", "", ""
try:
space_info = search_spaces(query)
if space_info is None:
return "No spaces found.", "", "", ""
# Format the results as markdown
results_md = "### Search Results:\n"
results_md += f"- ID: `{space_info['id']}`\n"
results_md += f"- Title: {space_info['title']}\n"
results_md += f"- Description: {space_info['description']}\n"
# Return values to update the UI
return (
results_md,
space_info["id"],
space_info["title"],
space_info["description"],
)
except Exception as e:
print(f"Error in search: {e}")
return f"Error: {str(e)}", "", "", ""
def on_validate_space(space_id):
if not space_id:
return "Please enter a space ID or search term.", "", ""
try:
# First try to get metadata directly if it's a valid space ID
space_info = get_space_metadata(space_id)
# If not found, try to search for it
if space_info is None:
# Try to search for the space using the ID as a search term
space_info = search_spaces(space_id)
if space_info is None:
return f"No spaces found for '{space_id}'.", "", ""
# Format search result as markdown
result_md = f"### Found Space via Search:\n"
result_md += f"- ID: `{space_info['id']}`\n"
result_md += f"- Title: {space_info['title']}\n"
result_md += f"- Description: {space_info['description']}\n"
return (
result_md,
space_info["title"],
space_info["description"],
)
# Format direct match as markdown
result_md = f"### Space Validated Successfully:\n"
result_md += f"- ID: `{space_info['id']}`\n"
result_md += f"- Title: {space_info['title']}\n"
result_md += f"- Description: {space_info['description']}\n"
return (
result_md,
space_info["title"],
space_info["description"],
)
except Exception as e:
print(f"Error validating space: {e}")
return f"Error: {str(e)}", "", ""
def on_add_tool(space_id, space_name, space_description, current_tools):
if not space_id:
return (
current_tools,
"Please enter a space ID.",
)
# Check if this tool is already added
for tool in current_tools:
if tool["id"] == space_id:
return (
current_tools,
f"Tool '{space_id}' is already added.",
)
# Add the new tool
new_tool = {
"id": space_id,
"name": space_name if space_name else space_id,
"description": space_description if space_description else "No description",
}
updated_tools = current_tools + [new_tool]
# Format the tools as markdown
tools_md = "### Added Tools:\n"
for i, tool in enumerate(updated_tools, 1):
tools_md += f"{i}. **{tool['name']}** (`{tool['id']}`)\n"
tools_md += f" {tool['description']}\n\n"
return updated_tools, tools_md
def on_create_agent(model, space_tools):
if not space_tools:
return (
None,
[],
"",
"Please add at least one tool before creating an agent.",
"No agent created yet.",
)
try:
# Create the agent
agent = create_agent(model, space_tools)
if agent is None:
return (
None,
[],
"",
"Failed to create agent. Please try again with different tools or model.",
"No agent created yet.",
)
# Format the tools for display
tools_str = ", ".join(
[f"{tool['name']} ({tool['id']})" for tool in space_tools]
)
# Generate agent status
agent_status = update_agent_status(agent)
return (
agent,
[],
"",
f"✅ Agent created successfully with {model}!\nTools: {tools_str}",
agent_status,
)
except Exception as e:
print(f"Error creating agent: {e}")
return None, [], "", f"Error creating agent: {str(e)}", "No agent created yet."
def add_user_message(message, chat_history):
"""Add the user message to the chat history."""
# For Gradio chatbot with type="messages", we need to use ChatMessage objects
if not message:
return "", chat_history
# Add user message to chat history
chat_history = chat_history + [ChatMessage(role="user", content=message)]
return message, chat_history
def stream_to_gradio(
agent,
task: str,
reset_agent_memory: bool = False,
additional_args: Optional[dict] = None,
):
"""Runs an agent with the given task and streams the messages from the agent as gradio ChatMessages."""
from smolagents.gradio_ui import pull_messages_from_step, handle_agent_output_types
from smolagents.agent_types import AgentAudio, AgentImage, AgentText
for step_log in agent.run(
task, stream=True, reset=reset_agent_memory, additional_args=additional_args
):
for message in pull_messages_from_step(
step_log,
):
yield message
final_answer = step_log # Last log is the run's final_answer
final_answer = handle_agent_output_types(final_answer)
if isinstance(final_answer, AgentImage):
yield gr.ChatMessage(
role="assistant",
content={"path": final_answer.to_string(), "mime_type": "image/png"},
)
elif isinstance(final_answer, AgentText) and os.path.exists(
final_answer.to_string()
):
yield gr.ChatMessage(
role="assistant",
content=gr.Image(final_answer.to_string()),
)
elif isinstance(final_answer, AgentAudio):
yield gr.ChatMessage(
role="assistant",
content={"path": final_answer.to_string(), "mime_type": "audio/wav"},
)
else:
yield gr.ChatMessage(
role="assistant", content=f"**Final answer:** {str(final_answer)}"
)
def stream_agent_response(agent, message, chat_history):
"""Stream the agent's response to the chat history."""
if not message or agent is None:
return chat_history
# First yield the current chat history
yield chat_history
try:
# Stream the agent's response
for msg in stream_to_gradio(agent, message):
# Add the message to chat history
chat_history = chat_history + [msg]
# Yield updated chat history
yield chat_history
except Exception as e:
# Handle errors
error_msg = f"Error: {str(e)}"
chat_history = chat_history + [ChatMessage(role="assistant", content=error_msg)]
yield chat_history
def on_clear(agent=None):
"""Clear the chat and reset the agent."""
return (
agent,
[],
"",
"Agent cleared. Create a new one to continue.",
"",
gr.update(interactive=False),
)
def update_agent_status(agent):
"""Update the agent status display with current information."""
if agent is None:
return "No agent created yet. Add a Space tool to get started."
# Get agent information
tools = agent.tools if hasattr(agent, "tools") else []
tool_count = len(tools)
# Create status message
status = f"Agent ready with {tool_count} tools"
return status
# Create the Gradio app
with gr.Blocks(title="AI Agent Builder") as app:
gr.Markdown("# AI Agent Builder with SmolaGents")
gr.Markdown("Build your own AI agent by selecting tools from Hugging Face Spaces.")
# Agent state
agent_state = gr.State(None)
last_message = gr.State("")
space_tools_state = gr.State([])
# Message store for preserving user message
msg_store = gr.State("")
with gr.Row():
# Left sidebar for tool configuration
with gr.Column(scale=1):
gr.Markdown("## Tool Configuration")
gr.Markdown("Add multiple Hugging Face Spaces as tools for your agent:")
# Hidden model input with default value
model_input = gr.Textbox(
value=DEFAULT_MODEL,
label="Model",
visible=False,
)
# Space tool input
with gr.Group():
gr.Markdown("### Add Space as Tool")
space_tool_input = gr.Textbox(
label="Space ID or Search Term",
placeholder=("Enter a Space ID or search term"),
info="Enter a Space ID (username/space-name) or search term",
)
space_name_input = gr.Textbox(
label="Tool Name (optional)",
placeholder="Enter a name for this tool",
)
space_description_input = gr.Textbox(
label="Tool Description (optional)",
placeholder="Enter a description for this tool",
lines=2,
)
add_tool_button = gr.Button("Add Tool", variant="primary")
# Display added tools
gr.Markdown("### Added Tools")
tools_display = gr.Markdown(
"No tools added yet. Add at least one tool before creating an agent."
)
# Create agent button
create_button = gr.Button(
"Create Agent with Selected Tools", variant="secondary", size="lg"
)
# Status message
status_msg = gr.Markdown("")
# Agent status display
agent_status = gr.Markdown("No agent created yet.")
# Main content area
with gr.Column(scale=2):
# Chat interface for the agent
chatbot = gr.Chatbot(
label="Agent Chat",
height=600,
show_copy_button=True,
avatar_images=("👤", "🤖"),
type="messages", # Use messages type for ChatMessage objects
)
msg = gr.Textbox(
label="Your message",
placeholder="Type a message to your agent...",
interactive=True,
)
with gr.Row():
with gr.Column(scale=1, min_width=60):
clear = gr.Button("🗑️", scale=1)
with gr.Column(scale=8):
# Empty column for spacing
pass
# Connect event handlers
# Connect the space_tool_input submit event to the validation handler
space_tool_input.submit(
on_validate_space,
inputs=[space_tool_input],
outputs=[status_msg, space_name_input, space_description_input],
)
# Connect the add tool button
add_tool_button.click(
on_add_tool,
inputs=[
space_tool_input,
space_name_input,
space_description_input,
space_tools_state,
],
outputs=[space_tools_state, tools_display],
)
# Connect the create button to the handler
create_button.click(
on_create_agent,
inputs=[model_input, space_tools_state],
outputs=[agent_state, chatbot, msg, status_msg, agent_status],
)
# Connect the message input to the chain of handlers
msg.submit(
lambda message: (message, message, ""), # Store message and clear input
inputs=[msg],
outputs=[msg_store, msg, msg],
queue=False,
).then(
add_user_message, # Add user message to chat
inputs=[msg_store, chatbot],
outputs=[msg_store, chatbot],
queue=False,
).then(
stream_agent_response, # Generate and stream response
inputs=[agent_state, msg_store, chatbot],
outputs=chatbot,
queue=True,
)
if __name__ == "__main__":
app.queue().launch()