大模型 | kotaemon核心GraphRAG、Agent、多模态代码解读！

AI女王

已于 2024-09-12 10:41:12 修改

阅读量1k

点赞数 18

文章标签：人工智能 AI大模型 ai 大模型 GraphRAG Agent 多模态

于 2024-09-07 08:00:00 首次发布

本文链接：https://blog.csdn.net/m0_56255097/article/details/141953000

版权

要说最近RAG方面火热的项目当属kotaemon，短时间暴涨8k star

kotaemon的亮点是可定制化RAG UI，核心技术点是混合索引（Vector、Keyword、GraphRAG）、复杂推理Agent（ReAct、ReWOO、MemoryGIST 和 GraphReader）、多模态。

一、混合索引（GraphRAG）

混合索引主要是指：全文和矢量融合，这里还有一个选型就是集成了RAG的新范式：GraphRAG

看代码直接用的微软GraphRAG

检索后重排采用LLMReranker

RERANK_PROMPT_TEMPLATE = """Given the following question and context,
return YES if the context is relevant to the question and NO if it isn't.

> Question: {question}
> Context:
>>>
{context}
>>>
> Relevant (YES / NO):"""

二、复杂推理Agent

推理目前主要实现了react与rewoo，tools包括google搜索工具、llm工具、wikipedia工具，可以自定义扩展。

react还是经典的Thought、Action、Action Input、Observation模式

zero_shot_react_prompt = PromptTemplate(
    template="""Answer the following questions as best you can. Give answer in {lang}. You have access to the following tools:
{tool_description}
Use the following format:

Question: the input question you must answer
Thought: you should always think about what to do

Action: the action to take, should be one of [{tool_names}]

Action Input: the input to the action, should be different from the action input of the same action in previous steps.

Observation: the result of the action

... (this Thought/Action/Action Input/Observation can repeat N times)
#Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin! After each Action Input.

Question: {instruction}
Thought:{agent_scratchpad}
    """
)

rewoo（Reasoning WithOut Observation），该范式将推理过程与外部观察分离
在这里插入图片描述

planner的prompt模版

from kotaemon.llms import PromptTemplate

zero_shot_planner_prompt = PromptTemplate(
    template="""You are an AI agent who makes step-by-step plans to solve a problem under the help of external tools.
For each step, make one plan followed by one tool-call, which will be executed later to retrieve evidence for that step.
You should store each evidence into a distinct variable #E1, #E2, #E3 ... that can be referred to in later tool-call inputs.

##Available Tools##
{tool_description}

##Output Format (Replace '<...>')##
#Plan1: <describe your plan here>
#E1: <toolname>[<input here>] (eg. Search[What is Python])
#Plan2: <describe next plan>
#E2: <toolname>[<input here, you can use #E1 to represent its expected output>]
And so on...

##Your Task##
{task}

##Now Begin##
"""
)

solver的prompt模版

zero_shot_solver_prompt = PromptTemplate(
    template="""You are an AI agent who solves a problem with my assistance. I will provide step-by-step plans(#Plan) and evidences(#E) that could be helpful.
Your task is to briefly summarize each step, then make a short final conclusion for your task. Give answer in {lang}.

##My Plans and Evidences##
{plan_evidence}

##Example Output##
First, I <did something> , and I think <...>; Second, I <...>, and I think <...>; ....
So, <your conclusion>.

##Your Task##
{task}

##Now Begin##
"""
)