RAG 技术趋势与展望

RAG 技术在过去两年经历了爆发式发展。从简单的"检索+生成"到 Graph RAG、Self-RAG、Agentic RAG，架构不断演进。本章梳理当前趋势和未来方向。

RAG 演进路线

Agentic RAG

Agentic RAG 是当前最重要的趋势：让 AI Agent 来决定何时检索、检索什么、如何验证。

"""
Agentic RAG 框架
"""
from dataclasses import dataclass, field
from enum import Enum
from abc import ABC, abstractmethod
class ToolType(Enum):
VECTOR_SEARCH = "vector_search"
WEB_SEARCH = "web_search"
SQL_QUERY = "sql_query"
API_CALL = "api_call"
CALCULATOR = "calculator"
@dataclass
class AgentStep:
"""Agent 推理步骤"""
thought: str
tool: ToolType | None = None
tool_input: str = ""
tool_output: str = ""
is_final: bool = False
class RAGTool(ABC):
"""RAG 工具基类"""
name: str
description: str
@abstractmethod
def execute(self, query: str) -> str:
...
class AgenticRAG:
"""Agentic RAG 引擎"""
AGENT_PROMPT = """你是一个 RAG Agent。根据用户问题，决定使用哪个工具获取信息。
可用工具：
{tools}
当前对话：
{history}
用户问题：{query}
请按以下格式思考：
Thought: 我需要...
Tool: 工具名称
Input: 工具输入
如果已有足够信息，回答：
Thought: 我已有足够信息
Answer: 最终回答"""
def __init__(self, tools: list[RAGTool], llm_client, max_steps: int = 5):
self.tools = {t.name: t for t in tools}
self.llm = llm_client
self.max_steps = max_steps
def answer(self, query: str) -> dict:
"""Agent 驱动的 RAG 问答"""
steps: list[AgentStep] = []
history = ""
for i in range(self.max_steps):
tools_desc = "\n".join(
f"- {name}: {tool.description}" for name, tool in self.tools.items()
)
response = self.llm.generate(
self.AGENT_PROMPT.format(
tools=tools_desc, history=history, query=query
)
)
step = self._parse_response(response)
steps.append(step)
if step.is_final:
return {
"answer": step.thought,
"steps": len(steps),
"tools_used": [s.tool.value for s in steps if s.tool],
}
if step.tool and step.tool.value in self.tools:
tool = self.tools[step.tool.value]
step.tool_output = tool.execute(step.tool_input)
history += f"\nStep {i+1}: Used {step.tool.value}, got: {step.tool_output[:200]}"
return {"answer": "达到最大步数限制", "steps": len(steps), "tools_used": []}
def _parse_response(self, response: str) -> AgentStep:
"""解析 Agent 响应"""
if "Answer:" in response:
answer = response.split("Answer:")[-1].strip()
return AgentStep(thought=answer, is_final=True)
thought = ""
tool = None
tool_input = ""
for line in response.split("\n"):
if line.startswith("Thought:"):
thought = line[8:].strip()
elif line.startswith("Tool:"):
tool_name = line[5:].strip()
try:
tool = ToolType(tool_name)
except ValueError:
pass
elif line.startswith("Input:"):
tool_input = line[6:].strip()
return AgentStep(thought=thought, tool=tool, tool_input=tool_input)

长上下文 vs RAG

随着 LLM 上下文窗口扩大（Gemini 2M tokens, Claude 200K tokens），一个争论兴起：还需要 RAG 吗？

结论：长上下文和 RAG 不是替代关系，而是互补。最佳实践是 RAG 负责精准检索，长上下文负责深度理解。

关键趋势预测

RAG 系统选型决策树

本章小结

恭喜你完成了 RAG 检索增强生成实战指南的全部学习！