论文解读《LightRAG: Simple and Fast Retrieval-Augmented Generation》

博客:https://learnopencv.com/lightrag

视频:https://www.youtube.com/watch?v=oageL-1I0GE

代码:https://github.com/HKUDS/LightRAG

论文:https://arxiv.org/abs/2410.05779

​ - 时间:2024.10

​ - 单位:University of Hong Kong、Beijing University of Posts and Telecommunications

RAG基本

好文:https://zhuanlan.zhihu.com/p/675509396

数据准备阶段:数据提取------>文本分块(Chunking)------>向量化(embedding)------>数据入库

应用阶段:用户提问------>数据检索(召回)------>重排/过滤------>注入Prompt------>LLM生成答案

RAG发展

  • 传统RAG:

    (1)痛点:

    • 它们无法捕捉分散在多个文档中的碎片信息之间的相互联系,这使得概括全面的见解变得具有挑战性。
    • 有限的上下文理解源于缺乏对检索到的块的整体概述。
    • 当数据量不断增长时,就会出现可扩展性效率低下的问题,从而导致检索质量不佳。
  • GraphRAG:

    好文:GraphRAG学习小结(2)_graphrag local-CSDN博客

    论文:https://arxiv.org/pdf/2404.16130

    • 时间:2024.4

    • 单位:Microsoft

    (1)创新:

    • 为了构建结构化索引知识图谱KG,GraphRAG 使用 LLM 从源文档中提取实体和关系。KG有什么好处呢?
      • 而传统 RAG 主要依赖于文本块的相似匹配,如果查询问题与存储文本不完全匹配或不在同一个文本块中,就可能无法返回最相关的结果,导致生成结果不够精准。GraphRAG可以有效地连接和聚合多个相关特性,即使它们在不同的节点中表达,也能通过图谱中的关系找到最佳检索结果。因此,它能更精确地满足复杂、多维的用户查询。

      • 传统RAG的一大痛点是它虽然能够较好的回答一些事实性问题,但是在面对一些统计性、总结性、概要性的QFS问题(Query-Focused Summarization)时却无能为力,或需要采用更复杂的技术手段来解决。GraphRAG的一个主要出发点也是能够更好的回答基于高层语义理解的总结性查询问题。

    (2)痛点:

    • GraphRAG 运行速度通常非常慢,因为它需要多个 LLM API 调用,可能会达到速率限制。

    • 它的成本极高。

    • 为了将新数据合并到现有的图形索引中,我们还需要为以前的数据重建整个 KG,这是一种低效的方法。

    • 没有对重复元素执行明确的重复数据删除步骤,这会导致图形索引嘈杂。

快速总结(vs GraphRAG)

(1)LightRAG 的知识图谱可以增量更新。为了将新数据合并到现有的图形索引中,GraphRAG 还需要为以前的数据重建整个 KG

(2)结合了 graph indexing 和 standard embedding 方法构建知识图谱。相对于 GraphRAG 以社区遍历的方法,LightRAG 专注于实体和关系的检索,进而减少检索开销

(3)Hybrid Query双层检索策略,结合了 local 和 global 方法检索

(4)推理速度更快。在检索阶段,LightRAG 需要少于 100 个token和一个 API 调用,而 GraphRAG 需要 社区数量 x 每个社区的平均令牌数量token

详细步骤

构建基于图的文本索引(知识图谱 KG)

(1)R(.),提取实体和关系: 正如 GraphRAG 中看到的那样,这里精心设计的提示词被发送到 LLM 以获取节点和边

python 复制代码
"""Example 1:

Entity_types: [person, technology, mission, organization, location]
Text:
while Alex clenched his jaw, the buzz of frustration dull against the backdrop of Taylor's authoritarian certainty. It was this competitive undercurrent that kept him alert, the sense that his and Jordan's shared commitment to discovery was an unspoken rebellion against Cruz's narrowing vision of control and order.

Then Taylor did something unexpected. They paused beside Jordan and, for a moment, observed the device with something akin to reverence. "If this tech can be understood..." Taylor said, their voice quieter, "It could change the game for us. For all of us."

The underlying dismissal earlier seemed to falter, replaced by a glimpse of reluctant respect for the gravity of what lay in their hands. Jordan looked up, and for a fleeting heartbeat, their eyes locked with Taylor's, a wordless clash of wills softening into an uneasy truce.

It was a small transformation, barely perceptible, but one that Alex noted with an inward nod. They had all been brought here by different paths
################
Output:
("entity"{tuple_delimiter}"Alex"{tuple_delimiter}"person"{tuple_delimiter}"Alex is a character who experiences frustration and is observant of the dynamics among other characters."){record_delimiter}
("entity"{tuple_delimiter}"Taylor"{tuple_delimiter}"person"{tuple_delimiter}"Taylor is portrayed with authoritarian certainty and shows a moment of reverence towards a device, indicating a change in perspective."){record_delimiter}
("entity"{tuple_delimiter}"Jordan"{tuple_delimiter}"person"{tuple_delimiter}"Jordan shares a commitment to discovery and has a significant interaction with Taylor regarding a device."){record_delimiter}
("entity"{tuple_delimiter}"Cruz"{tuple_delimiter}"person"{tuple_delimiter}"Cruz is associated with a vision of control and order, influencing the dynamics among other characters."){record_delimiter}
("entity"{tuple_delimiter}"The Device"{tuple_delimiter}"technology"{tuple_delimiter}"The Device is central to the story, with potential game-changing implications, and is revered by Taylor."){record_delimiter}
("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Taylor"{tuple_delimiter}"Alex is affected by Taylor's authoritarian certainty and observes changes in Taylor's attitude towards the device."{tuple_delimiter}"power dynamics, perspective shift"{tuple_delimiter}7){record_delimiter}
("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Jordan"{tuple_delimiter}"Alex and Jordan share a commitment to discovery, which contrasts with Cruz's vision."{tuple_delimiter}"shared goals, rebellion"{tuple_delimiter}6){record_delimiter}
("relationship"{tuple_delimiter}"Taylor"{tuple_delimiter}"Jordan"{tuple_delimiter}"Taylor and Jordan interact directly regarding the device, leading to a moment of mutual respect and an uneasy truce."{tuple_delimiter}"conflict resolution, mutual respect"{tuple_delimiter}8){record_delimiter}
("relationship"{tuple_delimiter}"Jordan"{tuple_delimiter}"Cruz"{tuple_delimiter}"Jordan's commitment to discovery is in rebellion against Cruz's vision of control and order."{tuple_delimiter}"ideological conflict, rebellion"{tuple_delimiter}5){record_delimiter}
("relationship"{tuple_delimiter}"Taylor"{tuple_delimiter}"The Device"{tuple_delimiter}"Taylor shows reverence towards the device, indicating its importance and potential impact."{tuple_delimiter}"reverence, technological significance"{tuple_delimiter}9){record_delimiter}
("content_keywords"{tuple_delimiter}"power dynamics, ideological conflict, discovery, rebellion"){completion_delimiter}"""

(2)**P(.),LLM 分析键值对生成:**键 (K) 是一个单词或短语,而值 (V) 是一段总结相关内容的段落。为每个实体及每条关系生成一个文本键值对(K, V),其中 K 是一个单词或短语,便于高效检索,V 是一个文本段落,用于文本片段的总结。

LightRAG combines graph indexing and standard embedding based approach. These KV data structures offer a more precise retrieval than less accurate standard embedding only RAG or inefficient chunk traversal techniques in GraphRAG.

python 复制代码
PROMPTS[
    "summarize_entity_descriptions"
] = """You are a helpful assistant responsible for generating a comprehensive summary of the data provided below.
Given one or two entities, and a list of descriptions, all related to the same entity or group of entities.
Please concatenate all of these into a single, comprehensive description. Make sure to include information collected from all the descriptions.
If the provided descriptions are contradictory, please resolve the contradictions and provide a single, coherent summary.
Make sure it is written in third person, and include the entity names so we the have full context.
Use {language} as output language.

#######
-Data-
Entities: {entity_name}
Description List: {description_list}
#######
Output:
"""

(3) **D(.),重复数据删除以优化图形操作:**通过去重函数识别并合并来自不同段落的相同实体和关系。

最后,我们得到了初始 KG 的最终优化版本:

双层检索机制

(1)对于给定的查询,LightRAG 的检索算法会提取局部查询关键词(low_level_keywords)和全局查询关键词(high_level_keywords):

(2)检索算法使用向量数据库来匹配局部查询关键词与候选实体,以及全局查询关键词与候选关系:

python 复制代码
async def _build_global_query_context(
    keywords,
    knowledge_graph_inst: BaseGraphStorage,
    entities_vdb: BaseVectorStorage,
    relationships_vdb: BaseVectorStorage,
    text_chunks_db: BaseKVStorage[TextChunkSchema],
    query_param: QueryParam,
):
    

# 用keywords匹配top_k个最相关的relationships

edge_datas = await relationships_vdb.query(keywords, top_k=query_param.top_k)

# 根据relationships找到entities

use_entities = await _find_most_related_entities_from_relationships(
edge_datas, query_param, knowledge_graph_inst
)

# 找到对应的文本块

use_text_units = await _find_related_text_unit_from_relationships(
edge_datas, query_param, text_chunks_db, knowledge_graph_inst
)

......
python 复制代码
from nano_vectordb import NanoVectorDB

......

self._client = NanoVectorDB(
     self.embedding_func.embedding_dim, storage_file=self._client_file_name
)

......

async def query(self, query: str, top_k=5):
    embedding = await self.embedding_func([query])
    embedding = embedding[0]
    results = self._client.query(
        query=embedding,
        top_k=top_k,
        better_than_threshold=self.cosine_better_than_threshold,
    )
    results = [
        {**dp, "id": dp["__id__"], "distance": dp["__metrics__"]} for dp in results
    ]
    return results

(3)增强高阶关联性:。LightRAG 进一步收集已检索到的实体或关系的局部子图,如实体或关系的一跳邻近节点,数学公式表示如下:

其中 Nv 和 Ne 分别表示检索到的节点 v 和边 e 的单跳相邻节点。

(4)检索到的信息分为三个部分:实体、关系和对应的文本块。然后将这个结构化数据输入到 LLM 中,使其能够为查询生成全面的答案:

评估指标

Win rates

跨越四个数据集和四个评估维度:

引入了一个综合评估框架,用于根据三个关键标准将两个答案与同一问题进行比较:Comprehensiveness、Diversity 和 Empowerment。其目的是通过为每个标准选择更好的答案来指导 LLM,然后进行整体评估。对于这三个标准中的每一个,LLM 必须确定哪些答案表现更好,并为其选择提供理由。最终,整体获胜者是根据所有三个维度的性能确定的,并附有证明决策的详细摘要:

Model cost

(1)索引阶段(Retrieval Phase):

Cmax 表示每个 API 调用允许的最大令牌数。在检索阶段,GraphRAG 生成 1,399 个社区,本实验积极使用 610 个 level-2 社区进行检索。因为 GraphRAG 的 map-reduce 机制,导致总内存消耗为 610,000 个 token(610 个社区 × 每个社区 1,000 个 token)。此外,GraphRAG 单独遍历每个社区的要求会导致数百个 API 调用,显著增加检索开销。相比之下,LightRAG 通过使用少于 100 个标记的关键字生成和检索来优化这个过程,只需要整个过程的单个 API 调用。

(2)更新 KG 阶段(Incremental Text Update):

Textract 表示实体和关系提取的令牌开销,Cextract 表示 extraction 所需的 API 调用数。在增量数据更新阶段,GraphRAG 必须消除其现有的社区结构以合并新的实体和关系,然后进行完全再生。GraphRAG 将需要大约 1399 个社区 × 2 级 level × 每个社区产生约 5,000 个 token 成本来重建原始和新的社区报告。相比之下,LightRAG 无缝地将新提取的实体和关系集成到现有图中,而无需完全重建。