added langgraph implementation
Browse files
app.py
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
import os
|
|
|
2 |
import gradio as gr
|
3 |
import asyncio
|
4 |
import requests
|
@@ -11,21 +12,175 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
11 |
|
12 |
# --- Basic Agent Definition ---
|
13 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
14 |
-
|
15 |
-
from
|
16 |
-
from
|
17 |
-
from
|
18 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
# Initialize the Hugging Face model
|
21 |
-
llm = HuggingFaceInferenceAPI(model_name="Qwen/Qwen2.5-Coder-32B-Instruct")
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
class BasicAgent:
|
31 |
def __init__(self):
|
@@ -35,13 +190,19 @@ class BasicAgent:
|
|
35 |
# fixed_answer = "This is a default answer."
|
36 |
# print(f"Agent returning fixed answer: {fixed_answer}")
|
37 |
# Create agent with all the tools
|
38 |
-
|
39 |
-
wiki_tools,
|
40 |
-
llm=llm
|
41 |
-
)
|
42 |
# Example query agent might receive
|
43 |
-
fixed_answer = await agent.run(question)
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
47 |
"""
|
|
|
1 |
import os
|
2 |
+
import time
|
3 |
import gradio as gr
|
4 |
import asyncio
|
5 |
import requests
|
|
|
12 |
|
13 |
# --- Basic Agent Definition ---
|
14 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
15 |
+
|
16 |
+
from typing import TypedDict, List, Dict, Any, Optional, Annotated
|
17 |
+
from langgraph.graph import StateGraph, START, END
|
18 |
+
from langchain_openai import ChatOpenAI
|
19 |
+
# from langchain_huggingface.llms import HuggingFaceEndpoint
|
20 |
+
from langchain_core.messages import AnyMessage, SystemMessage, HumanMessage
|
21 |
+
from langgraph.graph.message import add_messages
|
22 |
+
from langchain_community.utilities import WikipediaAPIWrapper
|
23 |
+
from langchain_community.tools import WikipediaQueryRun
|
24 |
+
from langchain_community.document_loaders import YoutubeLoader, WebBaseLoader
|
25 |
+
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
|
26 |
+
from langgraph.prebuilt import ToolNode, tools_condition
|
27 |
+
from langchain.tools import tool
|
28 |
|
29 |
# Initialize the Hugging Face model
|
|
|
30 |
|
31 |
+
llm = HuggingFaceEndpoint(
|
32 |
+
repo_id="Qwen/Qwen3-30B-A3B", # "Qwen/Qwen2.5-72B-Instruct",
|
33 |
+
provider="nebius", # "hf-inference",
|
34 |
+
max_new_tokens=8192,
|
35 |
+
do_sample=False,
|
36 |
+
# temperature=0.,
|
37 |
+
)
|
38 |
+
|
39 |
+
chat_model = ChatHuggingFace(llm=llm)
|
40 |
+
|
41 |
+
# Define tools
|
42 |
+
|
43 |
+
|
44 |
+
@tool
|
45 |
+
def fetch_website(url:str) -> str:
|
46 |
+
"""Fetch the content of a website.
|
47 |
+
Args:
|
48 |
+
url: The URL of the website to fetch.
|
49 |
+
Returns:
|
50 |
+
The title and content of the website.
|
51 |
+
"""
|
52 |
+
loader = WebBaseLoader(url)
|
53 |
+
docs = loader.load()
|
54 |
+
return docs[0].page_content
|
55 |
+
|
56 |
+
@tool
|
57 |
+
def ask_wiki(query: str) -> str:
|
58 |
+
"""Retrieve information from Wikipedia based on a user query.
|
59 |
+
Args:
|
60 |
+
query: A user query.
|
61 |
+
Returns:
|
62 |
+
A single string containing the retrieved article from Wikipedia.
|
63 |
+
"""
|
64 |
+
if not query.strip():
|
65 |
+
return "Please provide a valid query."
|
66 |
+
try:
|
67 |
+
wiki_toolapi_wrapper = WikipediaAPIWrapper(top_k_results=1, doc_content_chars_max=1000)
|
68 |
+
wiki_tool = WikipediaQueryRun(api_wrapper=wiki_toolapi_wrapper)
|
69 |
+
result = wiki_tool.run(query)
|
70 |
+
return result
|
71 |
+
except Exception as e:
|
72 |
+
return f"Error retrieving information: {str(e)}"
|
73 |
+
|
74 |
+
@tool
|
75 |
+
def youtube_transcript(url: str) -> str:
|
76 |
+
"""Retrieve transcript from Youtube based url.
|
77 |
+
Args:
|
78 |
+
url: input youtube url.
|
79 |
+
Returns:
|
80 |
+
A single string containing the transcript of the youtube videos.
|
81 |
+
"""
|
82 |
+
max_attempts = 5 # Set a maximum number of attempts
|
83 |
+
attempts = 0
|
84 |
+
loader = YoutubeLoader.from_youtube_url(url, add_video_info=False)
|
85 |
+
while attempts < max_attempts:
|
86 |
+
try:
|
87 |
+
docs = loader.load()
|
88 |
+
return docs[0].page_content
|
89 |
+
except Exception as e:
|
90 |
+
attempts += 1
|
91 |
+
print(f"Attempt {attempts} failed: {e}")
|
92 |
+
# Optionally add a delay before retrying
|
93 |
+
time.sleep(1) # Import the time module
|
94 |
+
return "Failed to retrieve transcript after multiple attempts."
|
95 |
+
|
96 |
+
# @tool
|
97 |
+
# def divide(a: int, b: int) -> float:
|
98 |
+
# """Divide a and b for occasional calculations.
|
99 |
+
# Args:
|
100 |
+
# a: integer
|
101 |
+
# b: integer
|
102 |
+
# Returns:
|
103 |
+
# A single float containing the result of the division.
|
104 |
+
# """
|
105 |
+
# return a / b
|
106 |
+
|
107 |
+
# Equip llm with tools
|
108 |
+
tools_list = [
|
109 |
+
fetch_website,
|
110 |
+
ask_wiki,
|
111 |
+
youtube_transcript,
|
112 |
+
]
|
113 |
+
|
114 |
+
llm_with_tools = chat_model.bind_tools(
|
115 |
+
tools_list,
|
116 |
+
# parallel_tool_calls=False
|
117 |
+
)
|
118 |
+
|
119 |
+
# Define Agent Workflow
|
120 |
+
|
121 |
+
class AgentState(TypedDict):
|
122 |
+
messages: Annotated[list[AnyMessage], add_messages]
|
123 |
+
|
124 |
+
|
125 |
+
def assistant(state: AgentState):
|
126 |
+
# System message
|
127 |
+
textual_description_of_tool="""
|
128 |
+
fetch_website(url: str) -> str:
|
129 |
+
Fetch the content of a website.
|
130 |
+
Args:
|
131 |
+
url: The URL of the website to fetch.
|
132 |
+
Returns:
|
133 |
+
The title and content of the website.
|
134 |
+
|
135 |
+
ask_wiki(query: str) -> str:
|
136 |
+
Retreive information from Wikipedia based on a user query.
|
137 |
+
Args:
|
138 |
+
query: A user query.
|
139 |
+
Returns:
|
140 |
+
A single string containing the retrieved article from Wikipedia.
|
141 |
+
|
142 |
+
youtube_transcript(url: str) -> str:
|
143 |
+
Args:
|
144 |
+
url: input youtube url.
|
145 |
+
Returns:
|
146 |
+
A single string containing the transcript of the youtube videos.
|
147 |
+
"""
|
148 |
+
sys_msg = SystemMessage(
|
149 |
+
content=f"You are a helpful assistant at answering user questions. You can access provided tools:\n{textual_description_of_tool}\n"
|
150 |
+
)
|
151 |
+
|
152 |
+
return {
|
153 |
+
"messages": [llm_with_tools.invoke([sys_msg] + state["messages"])],
|
154 |
+
}
|
155 |
+
|
156 |
+
# Build the StateGraph for the agent
|
157 |
+
# The graph
|
158 |
+
builder = StateGraph(AgentState)
|
159 |
+
|
160 |
+
# Define nodes: these do the work
|
161 |
+
builder.add_node("assistant", assistant)
|
162 |
+
builder.add_node("tools", ToolNode(tools_list))
|
163 |
+
|
164 |
+
# Define edges: these determine how the control flow moves
|
165 |
+
builder.add_edge(START, "assistant")
|
166 |
+
builder.add_conditional_edges(
|
167 |
+
"assistant",
|
168 |
+
# If the latest message requires a tool, route to tools
|
169 |
+
# Otherwise, provide a direct response
|
170 |
+
tools_condition,
|
171 |
+
)
|
172 |
+
builder.add_edge("tools", "assistant")
|
173 |
+
agent_graph = builder.compile()
|
174 |
+
|
175 |
+
messages = [
|
176 |
+
HumanMessage(
|
177 |
+
# content="Who is Barack Obama?"
|
178 |
+
# content="Divide 6790 by 5"
|
179 |
+
content=user_prompt
|
180 |
+
)
|
181 |
+
]
|
182 |
+
messages = agent_graph.invoke({"messages": messages}, config={"callbacks": [langfuse_handler]})
|
183 |
+
|
184 |
|
185 |
class BasicAgent:
|
186 |
def __init__(self):
|
|
|
190 |
# fixed_answer = "This is a default answer."
|
191 |
# print(f"Agent returning fixed answer: {fixed_answer}")
|
192 |
# Create agent with all the tools
|
193 |
+
|
|
|
|
|
|
|
194 |
# Example query agent might receive
|
195 |
+
# fixed_answer = await agent.run(question)
|
196 |
+
messages = [
|
197 |
+
HumanMessage(
|
198 |
+
# content="Who is Barack Obama?"
|
199 |
+
# content="Divide 6790 by 5"
|
200 |
+
content=question
|
201 |
+
)
|
202 |
+
]
|
203 |
+
response_text = messages['messages'][-1].content
|
204 |
+
return response_text.split('</think>')[-1]
|
205 |
+
|
206 |
|
207 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
208 |
"""
|