Create langgraph
Browse files
langgraph
ADDED
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
from langchain_core.messages import (
|
4 |
+
BaseMessage,
|
5 |
+
ToolMessage,
|
6 |
+
HumanMessage,
|
7 |
+
)
|
8 |
+
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
|
9 |
+
from langgraph.graph import END, StateGraph
|
10 |
+
|
11 |
+
|
12 |
+
def create_agent(llm, tools, system_message: str):
|
13 |
+
"""Create an agent."""
|
14 |
+
prompt = ChatPromptTemplate.from_messages(
|
15 |
+
[
|
16 |
+
(
|
17 |
+
"system",
|
18 |
+
"You are a helpful AI assistant, collaborating with other assistants."
|
19 |
+
" Use the provided tools to progress towards answering the question."
|
20 |
+
" If you are unable to fully answer, that's OK, another assistant with different tools "
|
21 |
+
" will help where you left off. Execute what you can to make progress."
|
22 |
+
" If you or any of the other assistants have the final answer or deliverable,"
|
23 |
+
" prefix your response with FINAL ANSWER so the team knows to stop."
|
24 |
+
" You have access to the following tools: {tool_names}.\n{system_message}",
|
25 |
+
),
|
26 |
+
MessagesPlaceholder(variable_name="messages"),
|
27 |
+
]
|
28 |
+
)
|
29 |
+
prompt = prompt.partial(system_message=system_message)
|
30 |
+
prompt = prompt.partial(tool_names=", ".join([tool.name for tool in tools]))
|
31 |
+
return prompt | llm.bind_tools(tools)
|
32 |
+
|
33 |
+
from langchain_core.tools import tool
|
34 |
+
from typing import Annotated
|
35 |
+
from langchain_experimental.utilities import PythonREPL
|
36 |
+
from langchain_community.tools.tavily_search import TavilySearchResults
|
37 |
+
|
38 |
+
tavily_tool = TavilySearchResults(max_results=5)
|
39 |
+
|
40 |
+
# Warning: This executes code locally, which can be unsafe when not sandboxed
|
41 |
+
|
42 |
+
repl = PythonREPL()
|
43 |
+
|
44 |
+
@tool
|
45 |
+
def python_repl(
|
46 |
+
code: Annotated[str, "The python code to execute to generate your chart."]
|
47 |
+
):
|
48 |
+
"""Use this to execute python code. If you want to see the output of a value,
|
49 |
+
you should print it out with `print(...)`. This is visible to the user."""
|
50 |
+
try:
|
51 |
+
result = repl.run(code)
|
52 |
+
except BaseException as e:
|
53 |
+
return f"Failed to execute. Error: {repr(e)}"
|
54 |
+
result_str = f"Successfully executed:\n```python\n{code}\n```\nStdout: {result}"
|
55 |
+
return (
|
56 |
+
result_str + "\n\nIf you have completed all tasks, respond with FINAL ANSWER."
|
57 |
+
)
|
58 |
+
|
59 |
+
import operator
|
60 |
+
from typing import Annotated, Sequence, TypedDict
|
61 |
+
|
62 |
+
from langchain_openai import ChatOpenAI
|
63 |
+
from typing_extensions import TypedDict
|
64 |
+
|
65 |
+
|
66 |
+
# This defines the object that is passed between each node
|
67 |
+
# in the graph. We will create different nodes for each agent and tool
|
68 |
+
class AgentState(TypedDict):
|
69 |
+
messages: Annotated[Sequence[BaseMessage], operator.add]
|
70 |
+
sender: str
|
71 |
+
|
72 |
+
import functools
|
73 |
+
from langchain_core.messages import AIMessage
|
74 |
+
|
75 |
+
|
76 |
+
# Helper function to create a node for a given agent
|
77 |
+
def agent_node(state, agent, name):
|
78 |
+
result = agent.invoke(state)
|
79 |
+
# We convert the agent output into a format that is suitable to append to the global state
|
80 |
+
if isinstance(result, ToolMessage):
|
81 |
+
pass
|
82 |
+
else:
|
83 |
+
result = AIMessage(**result.dict(exclude={"type", "name"}), name=name)
|
84 |
+
return {
|
85 |
+
"messages": [result],
|
86 |
+
# Since we have a strict workflow, we can
|
87 |
+
# track the sender so we know who to pass to next.
|
88 |
+
"sender": name,
|
89 |
+
}
|
90 |
+
|
91 |
+
llm = ChatOpenAI(model="gpt-4-1106-preview")
|
92 |
+
|
93 |
+
# Research agent and node
|
94 |
+
research_agent = create_agent(
|
95 |
+
llm,
|
96 |
+
[tavily_tool],
|
97 |
+
system_message="You should provide accurate data for the chart_generator to use.",
|
98 |
+
)
|
99 |
+
research_node = functools.partial(agent_node, agent=research_agent, name="Researcher")
|
100 |
+
|
101 |
+
# chart_generator
|
102 |
+
chart_agent = create_agent(
|
103 |
+
llm,
|
104 |
+
[python_repl],
|
105 |
+
system_message="Any charts you display will be visible by the user.",
|
106 |
+
)
|
107 |
+
chart_node = functools.partial(agent_node, agent=chart_agent, name="chart_generator")
|
108 |
+
|
109 |
+
from langgraph.prebuilt import ToolNode
|
110 |
+
|
111 |
+
tools = [tavily_tool, python_repl]
|
112 |
+
tool_node = ToolNode(tools)
|
113 |
+
|
114 |
+
# Either agent can decide to end
|
115 |
+
from typing import Literal
|
116 |
+
|
117 |
+
def router(state) -> Literal["call_tool", "__end__", "continue"]:
|
118 |
+
# This is the router
|
119 |
+
messages = state["messages"]
|
120 |
+
last_message = messages[-1]
|
121 |
+
if last_message.tool_calls:
|
122 |
+
# The previous agent is invoking a tool
|
123 |
+
return "call_tool"
|
124 |
+
if "FINAL ANSWER" in last_message.content:
|
125 |
+
# Any agent decided the work is done
|
126 |
+
return "__end__"
|
127 |
+
return "continue"
|
128 |
+
|
129 |
+
workflow = StateGraph(AgentState)
|
130 |
+
|
131 |
+
workflow.add_node("Researcher", research_node)
|
132 |
+
workflow.add_node("chart_generator", chart_node)
|
133 |
+
workflow.add_node("call_tool", tool_node)
|
134 |
+
|
135 |
+
workflow.add_conditional_edges(
|
136 |
+
"Researcher",
|
137 |
+
router,
|
138 |
+
{"continue": "chart_generator", "call_tool": "call_tool", "__end__": END},
|
139 |
+
)
|
140 |
+
workflow.add_conditional_edges(
|
141 |
+
"chart_generator",
|
142 |
+
router,
|
143 |
+
{"continue": "Researcher", "call_tool": "call_tool", "__end__": END},
|
144 |
+
)
|
145 |
+
|
146 |
+
workflow.add_conditional_edges(
|
147 |
+
"call_tool",
|
148 |
+
# Each agent node updates the 'sender' field
|
149 |
+
# the tool calling node does not, meaning
|
150 |
+
# this edge will route back to the original agent
|
151 |
+
# who invoked the tool
|
152 |
+
lambda x: x["sender"],
|
153 |
+
{
|
154 |
+
"Researcher": "Researcher",
|
155 |
+
"chart_generator": "chart_generator",
|
156 |
+
},
|
157 |
+
)
|
158 |
+
workflow.set_entry_point("Researcher")
|
159 |
+
graph = workflow.compile()
|
160 |
+
|
161 |
+
from IPython.display import Image, display
|
162 |
+
|
163 |
+
try:
|
164 |
+
display(Image(graph.get_graph(xray=True).draw_mermaid_png()))
|
165 |
+
except:
|
166 |
+
# This requires some extra dependencies and is optional
|
167 |
+
pass
|
168 |
+
|
169 |
+
events = graph.stream(
|
170 |
+
{
|
171 |
+
"messages": [
|
172 |
+
HumanMessage(
|
173 |
+
content="Fetch the UK's GDP over the past 5 years,"
|
174 |
+
" then draw a line graph of it."
|
175 |
+
" Once you code it up, finish."
|
176 |
+
)
|
177 |
+
],
|
178 |
+
},
|
179 |
+
# Maximum number of steps to take in the graph
|
180 |
+
{"recursion_limit": 150},
|
181 |
+
)
|
182 |
+
for s in events:
|
183 |
+
print(s)
|
184 |
+
print("----")
|