。

它在用图结构来完成RAG时#xff0c;使用社区这个概念并基于社区摘要来回答一些概括性的问题。

Graph

RAG流程如论文图1所…GraphRAG出自2024年4月的论文《From

Local

。

它在用图结构来完成RAG时使用社区这个概念并基于社区摘要来回答一些概括性的问题。

Graph

将文档分块论文也做了试验来分析chunk大小与后续步骤提取到的实体个数的关系如论文图2示意(gleaning是指在前面实体提取基础上提取漏掉的实体)。

虽然通常而言提取的实体提取越多越好但是还是平衡召回(recall)和精度(precision)。

让LLM从chunk中提取实体和关系在prompt中让LLM先识别实体、再识别这些实体之间的关系再按指定格式输出使用few-shot

prompt让LLM的实体和关系提取更准确(prompt在graphrag/index/graph/extractors/graph/prompts.py里)。

-Goal-

(entity{tuple_delimiter}entity_name{tuple_delimiter}entity_type{tuple_delimiter}entity_description2.

From

(relationship{tuple_delimiter}source_entity{tuple_delimiter}target_entity{tuple_delimiter}relationship_description{tuple_delimiter}relationship_strength)3.

Return

{completion_delimiter}######################

-Examples-

(entity{tuple_delimiter}Alex{tuple_delimiter}person{tuple_delimiter}Alex

character

(entity{tuple_delimiter}Taylor{tuple_delimiter}person{tuple_delimiter}Taylor

portrayed

perspective.){record_delimiter}

(entity{tuple_delimiter}Jordan{tuple_delimiter}person{tuple_delimiter}Jordan

shares

(entity{tuple_delimiter}Cruz{tuple_delimiter}person{tuple_delimiter}Cruz

associated

Device{tuple_delimiter}technology{tuple_delimiter}The

Device

(relationship{tuple_delimiter}Alex{tuple_delimiter}Taylor{tuple_delimiter}Alex

affected

device.{tuple_delimiter}7){record_delimiter}

(relationship{tuple_delimiter}Alex{tuple_delimiter}Jordan{tuple_delimiter}Alex

and

vision.{tuple_delimiter}6){record_delimiter}

(relationship{tuple_delimiter}Taylor{tuple_delimiter}Jordan{tuple_delimiter}Taylor

and

truce.{tuple_delimiter}8){record_delimiter}

(relationship{tuple_delimiter}Jordan{tuple_delimiter}Cruz{tuple_delimiter}Jordans

commitment

order.{tuple_delimiter}5){record_delimiter}

(relationship{tuple_delimiter}Taylor{tuple_delimiter}The

shows

impact.{tuple_delimiter}9){completion_delimiter}

Example

(entity{tuple_delimiter}Washington{tuple_delimiter}location{tuple_delimiter}Washington

location

(entity{tuple_delimiter}Operation:

Dulce{tuple_delimiter}mission{tuple_delimiter}Operation:

Dulce

team{tuple_delimiter}organization{tuple_delimiter}The

team

(relationship{tuple_delimiter}The

team{tuple_delimiter}Washington{tuple_delimiter}The

team

process.{tuple_delimiter}7){record_delimiter}

(relationship{tuple_delimiter}The

team{tuple_delimiter}Operation:

team

activities.{tuple_delimiter}9){completion_delimiter}

Example

Rivera{tuple_delimiter}person{tuple_delimiter}Sam

Rivera

(entity{tuple_delimiter}Alex{tuple_delimiter}person{tuple_delimiter}Alex

***

(entity{tuple_delimiter}Control{tuple_delimiter}concept{tuple_delimiter}Control

refers

(entity{tuple_delimiter}Intelligence{tuple_delimiter}concept{tuple_delimiter}Intelligence

here

communicate.){record_delimiter}

Contact{tuple_delimiter}event{tuple_delimiter}First

Contact

intelligence.){record_delimiter}

(entity{tuple_delimiter}Humanitys

Response{tuple_delimiter}event{tuple_delimiter}Humanitys

Response

intelligence.){record_delimiter}

(relationship{tuple_delimiter}Sam

Rivera{tuple_delimiter}Intelligence{tuple_delimiter}Sam

Rivera

intelligence.{tuple_delimiter}9){record_delimiter}

(relationship{tuple_delimiter}Alex{tuple_delimiter}First

leads

intelligence.{tuple_delimiter}10){record_delimiter}

(relationship{tuple_delimiter}Alex{tuple_delimiter}Humanitys

and

intelligence.{tuple_delimiter}8){record_delimiter}

(relationship{tuple_delimiter}Control{tuple_delimiter}Intelligence{tuple_delimiter}The

concept

rules.{tuple_delimiter}7){completion_delimiter}

-Real

让LLM针对提取的实体来进一步生成被称为covariate的claim包括subject,

object,

dates属性(prompt在graphrag/index/graph/extractors/claims/prompts.py里)

-Target

(subject_entity{tuple_delimiter}object_entity{tuple_delimiter}claim_type{tuple_delimiter}claim_status{tuple_delimiter}claim_start_date{tuple_delimiter}claim_end_date{tuple_delimiter}claim_description{tuple_delimiter}claim_source)3.

Return

{completion_delimiter}-Examples-

Example

B{tuple_delimiter}ANTI-COMPETITIVE

PRACTICES{tuple_delimiter}TRUE{tuple_delimiter}2022-01-10T00:00:00{tuple_delimiter}2022-01-10T00:00:00{tuple_delimiter}Company

was

2022/01/10{tuple_delimiter}According

article

B{tuple_delimiter}ANTI-COMPETITIVE

PRACTICES{tuple_delimiter}TRUE{tuple_delimiter}2022-01-10T00:00:00{tuple_delimiter}2022-01-10T00:00:00{tuple_delimiter}Company

was

2022/01/10{tuple_delimiter}According

article

C{tuple_delimiter}NONE{tuple_delimiter}CORRUPTION{tuple_delimiter}SUSPECTED{tuple_delimiter}2015-01-01T00:00:00{tuple_delimiter}2015-12-30T00:00:00{tuple_delimiter}Person

was

将同名的实体和关系合并并让LLM对实体和关系的描述生成summary这里提到即使LLM对同一个实体不能保证每次都生成一样描述但也不影响整体方案的效果。

(prompt在graphrag/index/graph/extractors/summarize/prompts.py里)

SUMMARIZE_PROMPT

Output:将前面提取的实体和关系构建成同构无向加权图实体作为图的节点关系作为图的边边的权重是关系的归一化计数。

在图上应用层次化社区发现算法Leiden得到的层次结构的每个级别包含一个社区分区每个分区是互斥的但是整体构成一个图使得可以实现分而治之的全局摘要。

对生成的社区生成摘要按如下方式来生成摘要(prompt在graphrag/index/graph/extractors/community_reports/prompts.py)

communities)按照一定的优先级将节点、关系、covariate加入到LLM上下文窗口直到达到token上限。

优先级定义为将社区中的边按照首尾节点的度之和来降序排序将首节点、尾节点、相关covariate、边的描述加入LLM上下文。

对更高级别的社区(Higher-level

communities):如果所有元素信息都可以放入LLM的上下文窗口则按叶子级别一样的处理逻辑。

否则将子社区按照元素摘要token数目降序排序并迭代地用更短的子社区摘要来替换更长的元素摘要直到长度满足LLM上下文要求。

(代码在graphrag/index/graph/extractors/community_reports/prep_community_report_context.py)

GraphRAG的查询有Local和Global两种模式Local适用于回答关于某个实体相关的问题Global模式适合回答关于整个数据集相关的问题。

将query在存储实体信息的向量库中检索出相关实体。

将第一步实体相关的chunk信息、社区摘要、实体详情、实体关系、实体Covarites按一定的格式组织作为上下文。

如果有历史聊天记录想历史聊天记录也作为上下文的一部分。

让LLM根据上下文生成回答prompt路径为graphrag/query/structured_search/local_search/system_prompt.py)。

Global查询的步骤如下如下图所示prompt在graphrag/query/structured_search/global_search/map_system_prompt.py

graphrag/query/structured_search/global_search/reduce_system_prompt.py

将所有社区摘要shuffle并分块作为上下文另将历史对话构成的上下文与这些社区摘要块拼接在一起作为上下文。

用map机制将前一步的多个上下文让LLM评估它们对于回答用户问题是否有帮助并进行0-100的打分。

过滤掉分数为0的上下文。

将前一步得到的结果合并且按照分数大小进行降序排序并将这些信息加入到LLM上下文窗口让LLM生成最终的回答。

GraphRAG在代码实现上有workflow概念如果需要修改运行流程只需要修改配置就可。

构建索引时的流程定义在graphrag/index/create_pipeline_config.py中(pipeline基于微软开源的另一个包DataShaper实现的)实体抽取、社区发现等操作被定义为verb代码在graphrag/index/verbs目录下。

但也因为它的workflow概念整个项目的代码可读性并不好。

参考资料