
ããŒã«ã«LLMã§Agentæ§ç¯
ããŒã«ã«LLMã䜿ã£ãŠAgentãæ§ç¯ããŠãããŒã«ã®åŒã³åºããŸã§ãã£ãŠã¿ãããšæããŸããããŒã«ã«LLMã¯å¶åŸ¡ãé£ãããAgentãšãšããŠåäœãããã®ãé£ããå°è±¡ã§ããããæ¥æ¬èªã§ãæ§èœã®é«ãã¢ãã«ãç»å ŽããŠããã®ã§ãã©ã€ããŠã¿ãŸãã
LLMã®æºå
vLLMã§ã¢ãã«ããããã€ããŸãã
from typing import List, Optional
from pydantic import BaseModel, Field
from vllm import LLM, SamplingParams
from outlines.serve.vllm import JSONLogitsProcessor
from langchain.agents import Tool
import json
model_id = 'tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.1'
vllm_model = LLM(
model=model_id, dtype="float16",
quantization="bitsandbytes", load_format="bitsandbytes",
gpu_memory_utilization=1.0, max_model_len=2048,
)ããŒã¿ã¢ãã«ãšããã³ãã
LLMã®å¶åŸ¡æ§ãé«ããããã«åœ¢åŒãæå®ããŠå¿çã§ããããã«ããŒã¿ã¢ãã«ãå®çŸ©ããŸããæ±ºããããããŒã¿ã¢ãã«ã§å¿çããããã®ããã³ãããåãããŠäœããŸãã
class ThoughtAction(BaseModel):
thought: str = Field(description="èããããšïŒäŸïŒããŒã«ã«ããæ£ç¢ºãªå€ãåŸãããã®ã§ãæçµçãªçããšããïŒ")
action: Optional[str] = Field(description="䜿çšããããŒã«å", default=None)
action_input: Optional[str] = Field(description="ããŒã«ãžã®å
¥å", default=None)
class AgentStep(BaseModel):
step: ThoughtAction
observation: Optional[str] = Field(description="ããŒã«ããã®åºåçµæ", default=None)
class AgentResponse(BaseModel):
steps: List[AgentStep]
final_answer: str
prompt_template = """
<|start_header_id|>system<|end_header_id|>
ããªãã¯ä»¥äžã®ããŒã«ã䜿çšã§ããæç€ºã«å¿ å®ã§åªç§ãªAIã¢ã·ã¹ã¿ã³ãã§ãã
å©çšå¯èœãªããŒã«:
{tools}
<|eot_id|>
<|start_header_id|>user<|end_header_id|>
ã¹ããããã€ã¹ãããã§èãã質åã«åçããŠãã ããã
- èããŠå®è¡ããçµç·¯ã¯å®è¡å±¥æŽã§äžããããŸãã
- å®è¡å±¥æŽããåçãæ¢ããŠãã ããã
- Observationããã芳å¯ããŠãã ããã
- ä»åã®æèã¯{n}åç®ã§ãã
- å¿
ãååã®å®è¡å±¥æŽãšå¥ã®è¡åããŠãã ããã
- 以äžã®åœ¢åŒã§JSONãçæããŠãã ããã
{schema}
æèã®äŸïŒ
Thought: èããããšïŒäŸïŒããŒã«ã«ããæ£ç¢ºãªå€ãåŸãããã®ã§ãæçµçãªçããšããïŒ
Action: è¡ãããšïŒäŸïŒããŒã«ã§èšç®ããïŒ
ActionInput: è¡åã«å¿
èŠãªæ
å ±(äŸ:1 + 2)
Observation: è¡åã®çµæ(äŸ:3)
{thought}
質å:
{question}
åç:
<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
"""Agentã¯ã©ã¹
Agentã¯ã©ã¹ãå®çŸ©ããŠãAgentã®ããžãã¯ãå®è£ ããŸãããã®Agentã¯ã©ã¹ããããŒã«åŒã³åºããã§ããããã«ããŠãããŸãã
class StructuredVLLMAgent:
def __init__(self, model, tools: List[Tool] = None):
self.vllm_model = model
self.tools = tools or []
self.tool_map = {tool.name: tool for tool in self.tools}
self.schema = ThoughtAction.model_json_schema()
json_logits_processor = JSONLogitsProcessor(self.schema, self.vllm_model)
self.sampling_params = SamplingParams(
temperature=0.1, top_p=0.95, max_tokens=1024,
repetition_penalty=1.2,
logits_processors=[json_logits_processor]
)
self.prompt_template = prompt_template
def _format_tools(self) -> str:
"""ããŒã«æ
å ±ã®ãã©ãŒããã"""
return "\n".join([
f"- {tool.name}: {tool.description}"
for tool in self.tools
])
def _execute_tool(self, tool_name: str, tool_input: str) -> str:
"""ããŒã«ã®å®è¡"""
if tool_name not in self.tool_map:
return f"Error: Tool '{tool_name}' not found"
try:
return str(self.tool_map[tool_name].func(tool_input))
except Exception as e:
return f"Error executing tool: {str(e)}"
def run(self, question: str, max_steps: int = 5) -> AgentResponse:
steps: List[AgentStep] = []
final_answer = None
for n in range(max_steps):
# ããã³ããã®çæ
thought = ''
if steps:
for m, x in enumerate(steps):
thought += f"{m+1}åç®ã®å®è¡å±¥æŽ:\n"
thought += f'\tThought:{x.step.thought}\n\tAction:{x.step.action}\n\tActionInput:{x.step.action_input}\n\tObservation:{x.observation}\n'
current_prompt = self.prompt_template.format(
tools=self._format_tools(),
question=question,
schema=self.schema,
thought=thought,
n=n+1,
)
outputs = self.vllm_model.generate(
[current_prompt],
self.sampling_params,
)
response_text = outputs[0].outputs[0].text
try:
# JSONå¿çã®ããŒã¹
step_data = json.loads(response_text)
current_step = ThoughtAction(**step_data)
# ããŒã«ã®å®è¡ãå¿
èŠãªå Žå
observation = None
if current_step.action and current_step.action_input:
observation = self._execute_tool(
current_step.action,
current_step.action_input
)
else:
observation = current_step.thought
steps.append(AgentStep(
step=current_step,
observation=observation
))
# æçµåçã«éããããã§ãã¯
if observation is not None:
final_answer = observation
break
except json.JSONDecodeError:
print(f"Error parsing JSON response: {response_text}")
continue
if not final_answer:
final_answer = "æå€§ã¹ãããæ°ã«éããŸããã"
return AgentResponse(steps=steps, final_answer=final_answer)Agentã®ç¹°ãè¿ãå®è¡åæ°ã¯5åãšããŠããŸãã
Agentã®äœãæ¹ãToolã¯Langchainã®æ¹åŒã«äžèŽããããã«å®è£
ããŠããŸãããAgentã¯ã©ã¹ã¯ç¬èªå®çŸ©ãšãªã£ãŠããŸããããŒã«ã«LLMã§å©çšå¯èœãªLangchainã®ãšã³ã·ã¹ãã ã䜿ãããšäŸ¿å©ãªã®ã§ããã®ãããªæ§æã«ããŠããŸãã
Toolãç¬èªå®çŸ©ã§å®è£
ããŠãæéã¯å€§ããå€ãããªãã§ãããæ¡åŒµæ§ãèãããšLangchainã«åãããæ¹ãè¯ãã§ãããã
Toolã®å®çŸ©
Toolã¯äžèšã®ããã«èšç®ããã ãã®Functionã§ãã
def calculator(expression: str) -> str:
try:
return f'{expression}ã®èšç®çµæã¯ã{str(eval(expression))}ã§ãã'
except Exception as e:
return f"èšç®ãšã©ãŒ: {str(e)}"
tools = [
Tool(
name="calculator",
func=calculator,
description="æ°åŒã®èšç®ãå®è¡ããŸããå
¥åã¯æ°åŒã®æååã§ãã"
)
]ãšãŒãžã§ã³ãã®å®è¡
ä»åã¯4ã€ã®ã¯ãšãªããã¹ãããŠã¿ãŸããã
agent = StructuredVLLMAgent(vllm_model, tools=tools)
# ãã¹ãã¯ãšãª
queries = [
"1234 * 5678 ãèšç®ããŠãã ãã",
"2024幎ã¯ä»€åäœå¹Žã§ããïŒ",
"å䟡3000åã®ååã«20%ã®å©çãã€ããŠå£²ããŸãããå®äŸ¡ã¯ãããïŒ",
"128-256ãèšç®ããŠ"
]
# ãã¹ãå®è¡
for query in queries:
print(f"\n質å: {query}")
response = agent.run(query)
print("\nåç:")
print(response.final_answer)
print("\nå±¥æŽ:")
print(json.dumps(response.model_dump(), indent=2, ensure_ascii=False))å®è¡çµæ
å®è¡ã¯ä»¥äžã®2ã€ã®LLMã䜿ã£ãŠã¿ãŸããã
Llama3.1
質å: 1234 * 5678 ãèšç®ããŠãã ãã
åç:
1234*5678ã®èšç®çµæã¯ã7006652ã§ãã
質å: 2024幎ã¯ä»€åäœå¹Žã§ããïŒ
åç:
2024幎ã¯ä»€å6幎ã§ãã
質å: å䟡3000åã®ååã«20%ã®å©çãã€ããŠå£²ããŸãããå®äŸ¡ã¯ãããïŒ
åç:
3000 * (1+0.2)ã®èšç®çµæã¯ã3600.0ã§ãã
質å: 128-256ãèšç®ããŠ
åç:
128 - 256ã®èšç®çµæã¯ã-128ã§ããLlama3
質å: 1234 * 5678 ãèšç®ããŠãã ãã
åç:
1234 * 5678ã®èšç®çµæã¯ã7006652ã§ãã
質å: 2024幎ã¯ä»€åäœå¹Žã§ããïŒ
åç:
2024 - 2019ã®èšç®çµæã¯ã5ã§ãã
質å: å䟡3000åã®ååã«20%ã®å©çãã€ããŠå£²ããŸãããå®äŸ¡ã¯ãããïŒ
åç:
3000 * (100 / 80)ã®èšç®çµæã¯ã3750.0ã§ãã
質å: 128-256ãèšç®ããŠ
åç:
128 - 256ã®èšç®çµæã¯ã-128ã§ããLlama3.1ã¯å
šåæ£è§£ããŠããããã§ãããLlama3ã®æ¹ã¯ééã£ãŠãããã®ããããŸããã
ãã¥ãŒãã³ã°ããŒã¿ãéã£ãŠãããšæãã®ã§ãæ¯èŒã«ããŸãæå³ã¯ãªããããããŸããããæ¥æ¬èªã§8Bã¢ãã«ã§Agentãåãããšã確èªã§ããŸããã