Update agent.py
Browse files
agent.py
CHANGED
@@ -1,100 +1,116 @@
|
|
1 |
-
from typing import
|
2 |
-
|
3 |
-
from langgraph.
|
4 |
-
from
|
5 |
-
from ai_tools import
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
workflow.add_node("process", self._process_result)
|
20 |
-
|
21 |
-
# 设置入口点
|
22 |
-
workflow.set_entry_point("tools")
|
23 |
-
|
24 |
-
# 添加条件边
|
25 |
-
workflow.add_conditional_edges(
|
26 |
-
"tools",
|
27 |
-
self._decide_next_step,
|
28 |
-
{
|
29 |
-
"continue": "process",
|
30 |
-
"end": END
|
31 |
-
}
|
32 |
-
)
|
33 |
-
|
34 |
-
workflow.add_edge("process", END)
|
35 |
-
|
36 |
-
# 添加持久化
|
37 |
-
workflow.checkpointer = SqliteSaver.from_conn_string(":memory:")
|
38 |
-
|
39 |
-
return workflow.compile()
|
40 |
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
file_name = state.get("file_name", "")
|
45 |
-
|
46 |
-
# 处理反转文本问题
|
47 |
-
if "rewsna" in question or "dnatsrednu" in question:
|
48 |
-
return {"result": self.tools.reverse_text(question.split('"')[1])}
|
49 |
-
|
50 |
-
# 处理蔬菜分类问题
|
51 |
-
if "grocery list" in question.lower() or "vegetables" in question.lower():
|
52 |
-
items = re.findall(r"[a-zA-Z]+(?=\W|\Z)", question)
|
53 |
-
return {"result": ", ".join(self.tools.categorize_vegetables(items))}
|
54 |
-
|
55 |
-
# 处理棋局问题
|
56 |
-
if "chess position" in question.lower() and file_name.endswith(".png"):
|
57 |
-
return {"result": self.tools.analyze_chess_position(file_name)}
|
58 |
-
|
59 |
-
# 处理音频文件问题
|
60 |
-
if file_name.endswith(".mp3"):
|
61 |
-
transcript = self.tools.extract_audio_transcript(file_name)
|
62 |
-
if "page numbers" in question.lower():
|
63 |
-
return {"result": transcript}
|
64 |
-
else:
|
65 |
-
return {"result": ", ".join(sorted(transcript.split(", ")))}
|
66 |
-
|
67 |
-
# 处理表格运算问题
|
68 |
-
if "* on the set S" in question:
|
69 |
-
table_data = {"operation": "*", "set": ["a", "b", "c", "d", "e"]}
|
70 |
-
return {"result": self.tools.process_table_operation(table_data)}
|
71 |
-
|
72 |
-
# 处理Python代码问题
|
73 |
-
if file_name.endswith(".py"):
|
74 |
-
return {"result": self.tools.analyze_python_code(file_name)}
|
75 |
-
|
76 |
-
# 处理Excel文件问题
|
77 |
-
if file_name.endswith(".xlsx"):
|
78 |
-
return {"result": self.tools.process_excel_file(file_name)}
|
79 |
-
|
80 |
-
return {"result": "I don't have a tool to answer this question."}
|
81 |
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
return {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
|
94 |
-
|
95 |
-
"""执行Agent"""
|
96 |
-
state = {"question": question, "file_name": file_name}
|
97 |
-
for step in self.workflow.stream(state):
|
98 |
-
if "__end__" in step:
|
99 |
-
return step["__end__"]["answer"]
|
100 |
-
return "No answer generated."
|
|
|
1 |
+
from typing import TypedDict, Annotated, Sequence
|
2 |
+
import operator
|
3 |
+
from langgraph.graph import StateGraph, END
|
4 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
5 |
+
from ai_tools import Calculator, DocRetriever, WebSearcher
|
6 |
|
7 |
+
# Configuration
|
8 |
+
MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
|
9 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
10 |
+
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
|
11 |
+
llm_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
12 |
+
|
13 |
+
# Define tools
|
14 |
+
tools = [Calculator(), WebSearcher()]
|
15 |
+
doc_retriever = DocRetriever()
|
16 |
+
tool_map = {tool.name: tool for tool in tools}
|
17 |
+
tool_map["DocRetriever"] = doc_retriever
|
18 |
+
|
19 |
+
# Agent State
|
20 |
+
class AgentState(TypedDict):
|
21 |
+
input: str
|
22 |
+
context: Annotated[Sequence[str], operator.add]
|
23 |
+
last_tool: str
|
24 |
+
|
25 |
+
# Tool calling prompt template
|
26 |
+
TOOL_PROMPT = """<|system|>
|
27 |
+
You're an expert problem solver. Use these tools:
|
28 |
+
{tool_descriptions}
|
29 |
+
|
30 |
+
Respond ONLY in this format:
|
31 |
+
Thought: <strategy>
|
32 |
+
Action: <tool_name>
|
33 |
+
Action Input: <input>
|
34 |
+
</s>
|
35 |
+
<|user|>
|
36 |
+
{input}
|
37 |
+
Context: {context}
|
38 |
+
</s>
|
39 |
+
<|assistant|>
|
40 |
+
"""
|
41 |
+
|
42 |
+
# Initialize graph
|
43 |
+
graph = StateGraph(AgentState)
|
44 |
+
|
45 |
+
# Node: Generate tool calls
|
46 |
+
def agent_node(state):
|
47 |
+
tool_list = "\n".join([f"- {t.name}: {t.description}" for t in tools])
|
48 |
+
prompt = TOOL_PROMPT.format(
|
49 |
+
tool_descriptions=tool_list,
|
50 |
+
input=state["input"],
|
51 |
+
context=state["context"]
|
52 |
+
)
|
53 |
|
54 |
+
response = llm_pipeline(
|
55 |
+
prompt,
|
56 |
+
max_new_tokens=150,
|
57 |
+
do_sample=True,
|
58 |
+
temperature=0.2,
|
59 |
+
pad_token_id=tokenizer.eos_token_id
|
60 |
+
)[0]['generated_text']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
+
# Extract tool call
|
63 |
+
action_match = re.search(r"Action: (\w+)", response)
|
64 |
+
action_input_match = re.search(r"Action Input: (.+?)\n", response)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
+
if action_match and action_input_match:
|
67 |
+
tool_name = action_match.group(1)
|
68 |
+
tool_input = action_input_match.group(1).strip()
|
69 |
+
return {
|
70 |
+
"last_tool": tool_name,
|
71 |
+
"tool_input": tool_input,
|
72 |
+
"thought": response
|
73 |
+
}
|
74 |
+
else:
|
75 |
+
return {"last_tool": "FINISH", "output": response}
|
76 |
+
|
77 |
+
# Node: Execute tools
|
78 |
+
def tool_node(state):
|
79 |
+
tool = tool_map.get(state["last_tool"])
|
80 |
+
if not tool:
|
81 |
+
return {"context": f"Error: Unknown tool {state['last_tool']}"}
|
82 |
+
|
83 |
+
result = tool.run(state["tool_input"])
|
84 |
+
return {"context": f"Tool {tool.name} returned: {result}"}
|
85 |
+
|
86 |
+
# Define graph structure
|
87 |
+
graph.add_node("agent", agent_node)
|
88 |
+
graph.add_node("tool", tool_node)
|
89 |
+
graph.set_entry_point("agent")
|
90 |
+
|
91 |
+
# Conditional edges
|
92 |
+
def route_action(state):
|
93 |
+
if state["last_tool"] == "FINISH":
|
94 |
+
return END
|
95 |
+
return "tool"
|
96 |
+
|
97 |
+
graph.add_edge("agent", "tool")
|
98 |
+
graph.add_conditional_edges("tool", route_action, {"agent": "agent", END: END})
|
99 |
+
graph.add_edge("tool", "agent") # Loop back after tool use
|
100 |
+
|
101 |
+
# Compile the agent
|
102 |
+
agent = graph.compile()
|
103 |
+
|
104 |
+
# Interface function
|
105 |
+
def run_agent(query: str, document: str = ""):
|
106 |
+
doc_retriever.document = document # Load document
|
107 |
+
state = {"input": query, "context": [], "last_tool": ""}
|
108 |
|
109 |
+
for step in agent.stream(state):
|
110 |
+
for node, value in step.items():
|
111 |
+
if node == "agent":
|
112 |
+
print(f"THOUGHT: {value['thought']}")
|
113 |
+
if node == "tool":
|
114 |
+
print(f"TOOL RESULT: {value['context']}")
|
115 |
|
116 |
+
return state["context"][-1] if state["context"] else "No output"
|
|
|
|
|
|
|
|
|
|
|
|