LightRAG出自2024年10月的論文《LIGHTRAG: SIMPLE AND FASTRETRIEVAL-AUGMENTED GENERATION》(github)，也是使用圖結(jié)構(gòu)來索引和搜索相關(guān)文本。

LightRAG作者認(rèn)為已有的RAG系統(tǒng)有如下兩個(gè)限制，導(dǎo)致難以回答類似"How does the rise of electric vehicles influence urban air quality and public transportation infrastructure?"的問題。

• 僅使用扁平數(shù)據(jù)表征(flat data representation)，這限制了它們基于實(shí)體之間的復(fù)雜關(guān)系理解和檢索信息的能力。
• 這些系統(tǒng)通常缺乏維持各種實(shí)體及其相互關(guān)系一致性所需的上下文意識(shí)，導(dǎo)致生成的回復(fù)可能無法完全滿足用戶查詢的需求。

為了解決上述問題，LightRAG的解決辦法是在現(xiàn)有RAG系統(tǒng)中引入圖結(jié)構(gòu)，其流程示意如論文圖1。

索引

先來看一下LightRAG的數(shù)據(jù)索引步驟：

1. 將一個(gè)文檔分塊(chunk)，并將chunk存入到一個(gè)KV存儲(chǔ)庫中（key是由前綴和chunk文本hash之后的id組成，value是chunk的文本和token長(zhǎng)度等內(nèi)容）。
2. 用LLM對(duì)每一個(gè)文本chunk提取實(shí)體和關(guān)系。(所有prompt在github的prompt.py文件)。實(shí)體由實(shí)體名稱唯一標(biāo)識(shí)(屬性有實(shí)體名稱、實(shí)體類型、實(shí)體描述、源chunk_id），關(guān)系由會(huì)由排序后的(首實(shí)體名稱、尾實(shí)體名稱)元祖來唯一標(biāo)識(shí)，除了描述外，還包括多個(gè)關(guān)鍵詞(屬性有頭實(shí)體名稱、尾實(shí)體名稱、關(guān)鍵詞、描述、權(quán)重、源chunk_id)。

## 第一次對(duì)chunk提取實(shí)體和關(guān)系的prompt
PROMPTS["entity_extraction"] = """-Goal-
Given a text document that is potentially relevant to this activity and a list of entity types, identify all entities of those types from the text and all relationships among the identified entities.-Steps-
1. Identify all entities. For each identified entity, extract the following information:
- entity_name: Name of the entity, use same language as input text. If English, capitalized the name.
- entity_type: One of the following types: [{entity_types}]
- entity_description: Comprehensive description of the entity's attributes and activities
Format each entity as ("entity"{tuple_delimiter}<entity_name>{tuple_delimiter}<entity_type>{tuple_delimiter}<entity_description>2. From the entities identified in step 1, identify all pairs of (source_entity, target_entity) that are *clearly related* to each other.
For each pair of related entities, extract the following information:
- source_entity: name of the source entity, as identified in step 1
- target_entity: name of the target entity, as identified in step 1
- relationship_description: explanation as to why you think the source entity and the target entity are related to each other
- relationship_strength: a numeric score indicating strength of the relationship between the source entity and target entity
- relationship_keywords: one or more high-level key words that summarize the overarching nature of the relationship, focusing on concepts or themes rather than specific details
Format each relationship as ("relationship"{tuple_delimiter}<source_entity>{tuple_delimiter}<target_entity>{tuple_delimiter}<relationship_description>{tuple_delimiter}<relationship_keywords>{tuple_delimiter}<relationship_strength>)3. Identify high-level key words that summarize the main concepts, themes, or topics of the entire text. These should capture the overarching ideas present in the document.
Format the content-level key words as ("content_keywords"{tuple_delimiter}<high_level_keywords>)4. Return output in English as a single list of all the entities and relationships identified in steps 1 and 2. Use **{record_delimiter}** as the list delimiter.5. When finished, output {completion_delimiter}######################
-Examples-
######################
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}
#############################
Example 2:Entity_types: [person, technology, mission, organization, location]
Text:
They were no longer mere operatives; they had become guardians of a threshold, keepers of a message from a realm beyond stars and stripes. This elevation in their mission could not be shackled by regulations and established protocols—it demanded a new perspective, a new resolve.Tension threaded through the dialogue of beeps and static as communications with Washington buzzed in the background. The team stood, a portentous air enveloping them. It was clear that the decisions they made in the ensuing hours could redefine humanity's place in the cosmos or condemn them to ignorance and potential peril.Their connection to the stars solidified, the group moved to address the crystallizing warning, shifting from passive recipients to active participants. Mercer's latter instincts gained precedence— the team's mandate had evolved, no longer solely to observe and report but to interact and prepare. A metamorphosis had begun, and Operation: Dulce hummed with the newfound frequency of their daring, a tone set not by the earthly
#############
Output:
("entity"{tuple_delimiter}"Washington"{tuple_delimiter}"location"{tuple_delimiter}"Washington is a location where communications are being received, indicating its importance in the decision-making process."){record_delimiter}
("entity"{tuple_delimiter}"Operation: Dulce"{tuple_delimiter}"mission"{tuple_delimiter}"Operation: Dulce is described as a mission that has evolved to interact and prepare, indicating a significant shift in objectives and activities."){record_delimiter}
("entity"{tuple_delimiter}"The team"{tuple_delimiter}"organization"{tuple_delimiter}"The team is portrayed as a group of individuals who have transitioned from passive observers to active participants in a mission, showing a dynamic change in their role."){record_delimiter}
("relationship"{tuple_delimiter}"The team"{tuple_delimiter}"Washington"{tuple_delimiter}"The team receives communications from Washington, which influences their decision-making process."{tuple_delimiter}"decision-making, external influence"{tuple_delimiter}7){record_delimiter}
("relationship"{tuple_delimiter}"The team"{tuple_delimiter}"Operation: Dulce"{tuple_delimiter}"The team is directly involved in Operation: Dulce, executing its evolved objectives and activities."{tuple_delimiter}"mission evolution, active participation"{tuple_delimiter}9){completion_delimiter}
("content_keywords"{tuple_delimiter}"mission evolution, decision-making, active participation, cosmic significance"){completion_delimiter}
#############################
Example 3:Entity_types: [person, role, technology, organization, event, location, concept]
Text:
their voice slicing through the buzz of activity. "Control may be an illusion when facing an intelligence that literally writes its own rules," they stated stoically, casting a watchful eye over the flurry of data."It's like it's learning to communicate," offered Sam Rivera from a nearby interface, their youthful energy boding a mix of awe and anxiety. "This gives talking to strangers' a whole new meaning."Alex surveyed his team—each face a study in concentration, determination, and not a small measure of trepidation. "This might well be our first contact," he acknowledged, "And we need to be ready for whatever answers back."Together, they stood on the edge of the unknown, forging humanity's response to a message from the heavens. The ensuing silence was palpable—a collective introspection about their role in this grand cosmic play, one that could rewrite human history.The encrypted dialogue continued to unfold, its intricate patterns showing an almost uncanny anticipation
#############
Output:
("entity"{tuple_delimiter}"Sam Rivera"{tuple_delimiter}"person"{tuple_delimiter}"Sam Rivera is a member of a team working on communicating with an unknown intelligence, showing a mix of awe and anxiety."){record_delimiter}
("entity"{tuple_delimiter}"Alex"{tuple_delimiter}"person"{tuple_delimiter}"Alex is the leader of a team attempting first contact with an unknown intelligence, acknowledging the significance of their task."){record_delimiter}
("entity"{tuple_delimiter}"Control"{tuple_delimiter}"concept"{tuple_delimiter}"Control refers to the ability to manage or govern, which is challenged by an intelligence that writes its own rules."){record_delimiter}
("entity"{tuple_delimiter}"Intelligence"{tuple_delimiter}"concept"{tuple_delimiter}"Intelligence here refers to an unknown entity capable of writing its own rules and learning to communicate."){record_delimiter}
("entity"{tuple_delimiter}"First Contact"{tuple_delimiter}"event"{tuple_delimiter}"First Contact is the potential initial communication between humanity and an unknown intelligence."){record_delimiter}
("entity"{tuple_delimiter}"Humanity's Response"{tuple_delimiter}"event"{tuple_delimiter}"Humanity's Response is the collective action taken by Alex's team in response to a message from an unknown intelligence."){record_delimiter}
("relationship"{tuple_delimiter}"Sam Rivera"{tuple_delimiter}"Intelligence"{tuple_delimiter}"Sam Rivera is directly involved in the process of learning to communicate with the unknown intelligence."{tuple_delimiter}"communication, learning process"{tuple_delimiter}9){record_delimiter}
("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"First Contact"{tuple_delimiter}"Alex leads the team that might be making the First Contact with the unknown intelligence."{tuple_delimiter}"leadership, exploration"{tuple_delimiter}10){record_delimiter}
("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Humanity's Response"{tuple_delimiter}"Alex and his team are the key figures in Humanity's Response to the unknown intelligence."{tuple_delimiter}"collective action, cosmic significance"{tuple_delimiter}8){record_delimiter}
("relationship"{tuple_delimiter}"Control"{tuple_delimiter}"Intelligence"{tuple_delimiter}"The concept of Control is challenged by the Intelligence that writes its own rules."{tuple_delimiter}"power dynamics, autonomy"{tuple_delimiter}7){record_delimiter}
("content_keywords"{tuple_delimiter}"first contact, control, communication, cosmic significance"){completion_delimiter}
#############################
-Real Data-
######################
Entity_types: {entity_types}
Text: {input_text}
######################
Output:
"""### 防止第一次提取的實(shí)體不全，讓LLM繼續(xù)提取的prompt
PROMPTS["entiti_continue_extraction"
] = """MANY entities were missed in the last extraction.  Add them below using the same format:
"""

3. 將提取的多個(gè)相同名稱的實(shí)體或關(guān)系用LLM來總結(jié)其描述（只有當(dāng)描述長(zhǎng)度超過配置長(zhǎng)度時(shí)，才會(huì)觸發(fā)LLM總結(jié)）

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

4. 將實(shí)體和關(guān)系的數(shù)據(jù)存儲(chǔ)為圖，同時(shí)將實(shí)體和關(guān)系的數(shù)據(jù)分別存入獨(dú)立的向量數(shù)據(jù)庫（實(shí)體的名稱作為meta_fields，實(shí)體的名稱和描述拼接成字符串作為向量編碼文本；關(guān)系的頭尾實(shí)體名稱作為meta_fileds，關(guān)系的關(guān)鍵詞、頭尾實(shí)體名稱、描述拼接成字符串作為向量編碼文本）。

5. 增量更新機(jī)制：每次有新數(shù)據(jù)進(jìn)來時(shí)，會(huì)檢查數(shù)據(jù)在已有的KV存儲(chǔ)或者圖存儲(chǔ)中是否已存在，再進(jìn)行更新合并或新增。

查詢

LightRAG將用戶查詢分為兩大類：

• 具體查詢(Specific Queries)：涉及到圖中具體實(shí)體的問題，比如”Who wrote ‘Pride and Prejudice’?“。
• 抽象查詢(Abstract Queries)：更概念化，涉及到不直接與具體實(shí)體關(guān)聯(lián)的更大的話題、主旨等，比如"How does artificial intelligence influence modern education?"。

所以為了能夠處理不同類型的用戶查詢，LightRAG也有兩種檢索策略：Low-Level Retrieval(Local)和High-Level Retrieval(Global)，用戶選擇其中一種或者同時(shí)使用這兩種檢索策略。下面介紹兩種檢索策略的詳細(xì)步驟。

Low-Level Retrieval:

1. 先讓LLM對(duì)給定的query q，提取local關(guān)鍵字（提取global關(guān)鍵詞的prompt是同一個(gè)）。

PROMPTS["keywords_extraction"] = """---Role---You are a helpful assistant tasked with identifying both high-level and low-level keywords in the user's query.---Goal---Given the query, list both high-level and low-level keywords. High-level keywords focus on overarching concepts or themes, while low-level keywords focus on specific entities, details, or concrete terms.---Instructions---- Output the keywords in JSON format.
- The JSON should have two keys:- "high_level_keywords" for overarching concepts or themes.- "low_level_keywords" for specific entities or details.######################
-Examples-
######################
Example 1:Query: "How does international trade influence global economic stability?"
################
Output:
{{"high_level_keywords": ["International trade", "Global economic stability", "Economic impact"],"low_level_keywords": ["Trade agreements", "Tariffs", "Currency exchange", "Imports", "Exports"]
}}
#############################
Example 2:Query: "What are the environmental consequences of deforestation on biodiversity?"
################
Output:
{{"high_level_keywords": ["Environmental consequences", "Deforestation", "Biodiversity loss"],"low_level_keywords": ["Species extinction", "Habitat destruction", "Carbon emissions", "Rainforest", "Ecosystem"]
}}
#############################
Example 3:Query: "What is the role of education in reducing poverty?"
################
Output:
{{"high_level_keywords": ["Education", "Poverty reduction", "Socioeconomic development"],"low_level_keywords": ["School access", "Literacy rates", "Job training", "Income inequality"]
}}
#############################
-Real Data-
######################
Query: {query}
######################
Output:"""

2. 如果沒有關(guān)鍵詞則返回?zé)o法回答問題，否則將LLM提取的關(guān)鍵字用逗號(hào)拼接起來作為新的query。
3. 用上一步生成的新query來檢索存儲(chǔ)實(shí)體的向量數(shù)據(jù)庫，召回top-k個(gè)實(shí)體。
4. 根據(jù)top-k實(shí)體獲取上下文，上下文分為三部分，用帶head的csv格式表示： 1. 實(shí)體在圖譜中存儲(chǔ)的詳情(名稱、類型描述、度)； 2. 實(shí)體相關(guān)的源chunk文本，這些chunk數(shù)據(jù)根據(jù)實(shí)體的top-k位置正序以及與實(shí)體的一度鄰居實(shí)體共享這個(gè)chunk的次數(shù)倒序排序(all_text_units = sorted(all_text_units, key=lambda x: (x["order"], -x["relation_counts"])))； 3. 實(shí)體相關(guān)的邊詳情（頭實(shí)體名稱、尾實(shí)體名稱、關(guān)鍵詞、描述、權(quán)重、度)，這些邊根據(jù)度和權(quán)重倒序排序。
5. 如果前一步?jīng)]有上下文則返回?zé)o法回答問題，否則讓LLM根據(jù)上下文來回答用戶的問題。

PROMPTS["rag_response"] = """---Role---You are a helpful assistant responding to questions about data in the tables provided.---Goal---Generate a response of the target length and format that responds to the user's question, summarizing all information in the input data tables appropriate for the response length and format, and incorporating any relevant general knowledge.
If you don't know the answer, just say so. Do not make anything up.
Do not include information where the supporting evidence for it is not provided.---Target response length and format---{response_type}---Data tables---{context_data}Add sections and commentary to the response as appropriate for the length and format. Style the response in markdown.
"""

High-Level Retrieval:

1. 讓LLM對(duì)給定的query q，提取global關(guān)鍵字（與前面提取local關(guān)鍵詞的prompt是同一個(gè)）。
2. 如果沒有關(guān)鍵詞則返回?zé)o法回答問題，否則將LLM提取的關(guān)鍵字用逗號(hào)拼接起來作為新的query。
3. 用上一步生成的新query來檢索存儲(chǔ)關(guān)系的向量數(shù)據(jù)庫，召回top-k個(gè)關(guān)系。
4. 根據(jù)top-k關(guān)系獲取上下文，上下文分為三部分，用帶head的csv格式表示： 1. 關(guān)系對(duì)應(yīng)的首尾實(shí)體在圖譜中存儲(chǔ)的詳情(名稱、類型描述、度)； 2. 關(guān)系相關(guān)的源chunk文本，這些chunk數(shù)據(jù)根據(jù)關(guān)系的top-k位置正序； 3. 關(guān)系詳情（頭實(shí)體名稱、尾實(shí)體名稱、關(guān)鍵詞、描述、權(quán)重、度)，這些邊根據(jù)度和權(quán)重倒序排序（感覺這里排序直接就按照向量召回的top-k排序就可以了，不需要與low-level rerieval保持一致）。
5. 如果前一步?jīng)]有上下文則返回?zé)o法回答問題，否則讓LLM根據(jù)上下文來回答用戶的問題（prompt與low-level rerieval一致）。

總結(jié)

LightRAG像是GraphRAG的簡(jiǎn)化版。它利用圖結(jié)構(gòu)來存儲(chǔ)和檢索文本數(shù)據(jù)，對(duì)于實(shí)體和關(guān)系不僅使用圖存儲(chǔ)，同時(shí)使用向量存儲(chǔ)。檢索時(shí)先用向量進(jìn)行實(shí)體或關(guān)系召回，再借助圖結(jié)構(gòu)找到對(duì)應(yīng)的chunk文本。