🚀 LightRAG: Simple and Fast Retrieval-Augmented Generation

This repository hosts the code of LightRAG. The structure of this code is based on nano-graphrag.

🎉 News

• [2024.10.29]🎯🎯📢📢Multi-file types are now supported by textract.
• [2024.10.20]🎯🎯📢📢We’ve added a new feature to LightRAG: Graph Visualization.
• [2024.10.18]🎯🎯📢📢We’ve added a link to a LightRAG Introduction Video. Thanks to the author!
• [2024.10.17]🎯🎯📢📢We have created a Discord channel! Welcome to join for sharing and discussions! 🎉🎉
• [2024.10.16]🎯🎯📢📢LightRAG now supports Ollama models!
• [2024.10.15]🎯🎯📢📢LightRAG now supports Hugging Face models!

Algorithm Flowchart

Install

• Install from source (Recommend)

cd LightRAG
pip install -e .

• Install from PyPI

pip install lightrag-hku

Quick Start

• Video demo of running LightRAG locally.
• All the code can be found in the examples.
• Set OpenAI API key in environment if using OpenAI models: export OPENAI_API_KEY="sk-...".
• Download the demo text "A Christmas Carol by Charles Dickens":

curl https://raw.githubusercontent.com/gusye1234/nano-graphrag/main/tests/mock_data.txt > ./book.txt

Use the below Python snippet (in a script) to initialize LightRAG and perform queries:

import os
from lightrag import LightRAG, QueryParam
from lightrag.llm import gpt_4o_mini_complete, gpt_4o_complete

#########
# Uncomment the below two lines if running in a jupyter notebook to handle the async nature of rag.insert()
# import nest_asyncio
# nest_asyncio.apply()
#########

WORKING_DIR = "./dickens"


if not os.path.exists(WORKING_DIR):
    os.mkdir(WORKING_DIR)

rag = LightRAG(
    working_dir=WORKING_DIR,
    llm_model_func=gpt_4o_mini_complete  # Use gpt_4o_mini_complete LLM model
    # llm_model_func=gpt_4o_complete  # Optionally, use a stronger model
)

with open("./book.txt") as f:
    rag.insert(f.read())

# Perform naive search
print(rag.query("What are the top themes in this story?", param=QueryParam(mode="naive")))

# Perform local search
print(rag.query("What are the top themes in this story?", param=QueryParam(mode="local")))

# Perform global search
print(rag.query("What are the top themes in this story?", param=QueryParam(mode="global")))

# Perform hybrid search
print(rag.query("What are the top themes in this story?", param=QueryParam(mode="hybrid")))

Using Open AI-like APIs

• LightRAG also supports Open AI-like chat/embeddings APIs:

async def llm_model_func(
    prompt, system_prompt=None, history_messages=[], **kwargs
) -> str:
    return await openai_complete_if_cache(
        "solar-mini",
        prompt,
        system_prompt=system_prompt,
        history_messages=history_messages,
        api_key=os.getenv("UPSTAGE_API_KEY"),
        base_url="https://api.upstage.ai/v1/solar",
        **kwargs
    )

async def embedding_func(texts: list[str]) -> np.ndarray:
    return await openai_embedding(
        texts,
        model="solar-embedding-1-large-query",
        api_key=os.getenv("UPSTAGE_API_KEY"),
        base_url="https://api.upstage.ai/v1/solar"
    )

rag = LightRAG(
    working_dir=WORKING_DIR,
    llm_model_func=llm_model_func,
    embedding_func=EmbeddingFunc(
        embedding_dim=4096,
        max_token_size=8192,
        func=embedding_func
    )
)

Using Hugging Face Models

• If you want to use Hugging Face models, you only need to set LightRAG as follows:

from lightrag.llm import hf_model_complete, hf_embedding
from transformers import AutoModel, AutoTokenizer

# Initialize LightRAG with Hugging Face model
rag = LightRAG(
    working_dir=WORKING_DIR,
    llm_model_func=hf_model_complete,  # Use Hugging Face model for text generation
    llm_model_name='meta-llama/Llama-3.1-8B-Instruct',  # Model name from Hugging Face
    # Use Hugging Face embedding function
    embedding_func=EmbeddingFunc(
        embedding_dim=384,
        max_token_size=5000,
        func=lambda texts: hf_embedding(
            texts,
            tokenizer=AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2"),
            embed_model=AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
        )
    ),
)

Using Ollama Models

Overview

If you want to use Ollama models, you need to pull model you plan to use and embedding model, for example nomic-embed-text.

Then you only need to set LightRAG as follows:

from lightrag.llm import ollama_model_complete, ollama_embedding

# Initialize LightRAG with Ollama model
rag = LightRAG(
    working_dir=WORKING_DIR,
    llm_model_func=ollama_model_complete,  # Use Ollama model for text generation
    llm_model_name='your_model_name', # Your model name
    # Use Ollama embedding function
    embedding_func=EmbeddingFunc(
        embedding_dim=768,
        max_token_size=8192,
        func=lambda texts: ollama_embedding(
            texts,
            embed_model="nomic-embed-text"
        )
    ),
)

Increasing context size

In order for LightRAG to work context should be at least 32k tokens. By default Ollama models have context size of 8k. You can achieve this using one of two ways:

Increasing the num_ctx parameter in Modelfile.

1. Pull the model:

ollama pull qwen2

2. Display the model file:

ollama show --modelfile qwen2 > Modelfile

3. Edit the Modelfile by adding the following line:

PARAMETER num_ctx 32768

4. Create the modified model:

ollama create -f Modelfile qwen2m

Setup num_ctx via Ollama API.

Tiy can use llm_model_kwargs param to configure ollama:

rag = LightRAG(
    working_dir=WORKING_DIR,
    llm_model_func=ollama_model_complete,  # Use Ollama model for text generation
    llm_model_name='your_model_name', # Your model name
    llm_model_kwargs={"options": {"num_ctx": 32768}},
    # Use Ollama embedding function
    embedding_func=EmbeddingFunc(
        embedding_dim=768,
        max_token_size=8192,
        func=lambda texts: ollama_embedding(
            texts,
            embed_model="nomic-embed-text"
        )
    ),
)

Fully functional example

There fully functional example examples/lightrag_ollama_demo.py that utilizes gemma2:2b model, runs only 4 requests in parallel and set context size to 32k.

Low RAM GPUs

In order to run this experiment on low RAM GPU you should select small model and tune context window (increasing context increase memory consumption). For example, running this ollama example on repurposed mining GPU with 6Gb of RAM required to set context size to 26k while using gemma2:2b. It was able to find 197 entities and 19 relations on book.txt.

Query Param

class QueryParam:
    mode: Literal["local", "global", "hybrid", "naive"] = "global"
    only_need_context: bool = False
    response_type: str = "Multiple Paragraphs"
    # Number of top-k items to retrieve; corresponds to entities in "local" mode and relationships in "global" mode.
    top_k: int = 60
    # Number of tokens for the original chunks.
    max_token_for_text_unit: int = 4000
    # Number of tokens for the relationship descriptions
    max_token_for_global_context: int = 4000
    # Number of tokens for the entity descriptions
    max_token_for_local_context: int = 4000

Batch Insert

# Batch Insert: Insert multiple texts at once
rag.insert(["TEXT1", "TEXT2",...])

Incremental Insert

# Incremental Insert: Insert new documents into an existing LightRAG instance
rag = LightRAG(
     working_dir=WORKING_DIR,
     llm_model_func=llm_model_func,
     embedding_func=EmbeddingFunc(
          embedding_dim=embedding_dimension,
          max_token_size=8192,
          func=embedding_func,
     ),
)

with open("./newText.txt") as f:
    rag.insert(f.read())

Multi-file Type Support

The testract supports reading file types such as TXT, DOCX, PPTX, CSV, and PDF.

import textract

file_path = 'TEXT.pdf'
text_content = textract.process(file_path)

rag.insert(text_content.decode('utf-8'))

Graph Visualization

Graph visualization with html

• The following code can be found in examples/graph_visual_with_html.py

import networkx as nx
from pyvis.network import Network

# Load the GraphML file
G = nx.read_graphml('./dickens/graph_chunk_entity_relation.graphml')

# Create a Pyvis network
net = Network(notebook=True)

# Convert NetworkX graph to Pyvis network
net.from_nx(G)

# Save and display the network
net.show('knowledge_graph.html')

Graph visualization with Neo4j

• The following code can be found in examples/graph_visual_with_neo4j.py

import os
import json
from lightrag.utils import xml_to_json
from neo4j import GraphDatabase

# Constants
WORKING_DIR = "./dickens"
BATCH_SIZE_NODES = 500
BATCH_SIZE_EDGES = 100

# Neo4j connection credentials
NEO4J_URI = "bolt://localhost:7687"
NEO4J_USERNAME = "neo4j"
NEO4J_PASSWORD = "your_password"

def convert_xml_to_json(xml_path, output_path):
    """Converts XML file to JSON and saves the output."""
    if not os.path.exists(xml_path):
        print(f"Error: File not found - {xml_path}")
        return None

    json_data = xml_to_json(xml_path)
    if json_data:
        with open(output_path, 'w', encoding='utf-8') as f:
            json.dump(json_data, f, ensure_ascii=False, indent=2)
        print(f"JSON file created: {output_path}")
        return json_data
    else:
        print("Failed to create JSON data")
        return None

def process_in_batches(tx, query, data, batch_size):
    """Process data in batches and execute the given query."""
    for i in range(0, len(data), batch_size):
        batch = data[i:i + batch_size]
        tx.run(query, {"nodes": batch} if "nodes" in query else {"edges": batch})

def main():
    # Paths
    xml_file = os.path.join(WORKING_DIR, 'graph_chunk_entity_relation.graphml')
    json_file = os.path.join(WORKING_DIR, 'graph_data.json')

    # Convert XML to JSON
    json_data = convert_xml_to_json(xml_file, json_file)
    if json_data is None:
        return

    # Load nodes and edges
    nodes = json_data.get('nodes', [])
    edges = json_data.get('edges', [])

    # Neo4j queries
    create_nodes_query = """
    UNWIND $nodes AS node
    MERGE (e:Entity {id: node.id})
    SET e.entity_type = node.entity_type,
        e.description = node.description,
        e.source_id = node.source_id,
        e.displayName = node.id
    REMOVE e:Entity
    WITH e, node
    CALL apoc.create.addLabels(e, [node.entity_type]) YIELD node AS labeledNode
    RETURN count(*)
    """

    create_edges_query = """
    UNWIND $edges AS edge
    MATCH (source {id: edge.source})
    MATCH (target {id: edge.target})
    WITH source, target, edge,
         CASE
            WHEN edge.keywords CONTAINS 'lead' THEN 'lead'
            WHEN edge.keywords CONTAINS 'participate' THEN 'participate'
            WHEN edge.keywords CONTAINS 'uses' THEN 'uses'
            WHEN edge.keywords CONTAINS 'located' THEN 'located'
            WHEN edge.keywords CONTAINS 'occurs' THEN 'occurs'
           ELSE REPLACE(SPLIT(edge.keywords, ',')[0], '\"', '')
         END AS relType
    CALL apoc.create.relationship(source, relType, {
      weight: edge.weight,
      description: edge.description,
      keywords: edge.keywords,
      source_id: edge.source_id
    }, target) YIELD rel
    RETURN count(*)
    """

    set_displayname_and_labels_query = """
    MATCH (n)
    SET n.displayName = n.id
    WITH n
    CALL apoc.create.setLabels(n, [n.entity_type]) YIELD node
    RETURN count(*)
    """

    # Create a Neo4j driver
    driver = GraphDatabase.driver(NEO4J_URI, auth=(NEO4J_USERNAME, NEO4J_PASSWORD))

    try:
        # Execute queries in batches
        with driver.session() as session:
            # Insert nodes in batches
            session.execute_write(process_in_batches, create_nodes_query, nodes, BATCH_SIZE_NODES)

            # Insert edges in batches
            session.execute_write(process_in_batches, create_edges_query, edges, BATCH_SIZE_EDGES)

            # Set displayName and labels
            session.run(set_displayname_and_labels_query)

    except Exception as e:
        print(f"Error occurred: {e}")

    finally:
        driver.close()

if __name__ == "__main__":
    main()

API Server Implementation

LightRAG also provides a FastAPI-based server implementation for RESTful API access to RAG operations. This allows you to run LightRAG as a service and interact with it through HTTP requests.

Setting up the API Server

Click to expand setup instructions

1. First, ensure you have the required dependencies:

pip install fastapi uvicorn pydantic

2. Set up your environment variables:

export RAG_DIR="your_index_directory"  # Optional: Defaults to "index_default"

3. Run the API server:

python examples/lightrag_api_openai_compatible_demo.py

The server will start on http://0.0.0.0:8020.

API Endpoints

The API server provides the following endpoints:

1. Query Endpoint

Click to view Query endpoint details

• URL: /query
• Method: POST
• Body:

{
    "query": "Your question here",
    "mode": "hybrid"  // Can be "naive", "local", "global", or "hybrid"
}

• Example:

curl -X POST "http://127.0.0.1:8020/query" \
     -H "Content-Type: application/json" \
     -d '{"query": "What are the main themes?", "mode": "hybrid"}'

2. Insert Text Endpoint

Click to view Insert Text endpoint details

• URL: /insert
• Method: POST
• Body:

{
    "text": "Your text content here"
}

• Example:

curl -X POST "http://127.0.0.1:8020/insert" \
     -H "Content-Type: application/json" \
     -d '{"text": "Content to be inserted into RAG"}'

3. Insert File Endpoint

Click to view Insert File endpoint details

• URL: /insert_file
• Method: POST
• Body:

{
    "file_path": "path/to/your/file.txt"
}

• Example:

curl -X POST "http://127.0.0.1:8020/insert_file" \
     -H "Content-Type: application/json" \
     -d '{"file_path": "./book.txt"}'

4. Health Check Endpoint

Click to view Health Check endpoint details

• URL: /health
• Method: GET
• Example:

curl -X GET "http://127.0.0.1:8020/health"

Configuration

The API server can be configured using environment variables:

• RAG_DIR: Directory for storing the RAG index (default: "index_default")
• API keys and base URLs should be configured in the code for your specific LLM and embedding model providers

Error Handling

Click to view error handling details

The API includes comprehensive error handling:

• File not found errors (404)
• Processing errors (500)
• Supports multiple file encodings (UTF-8 and GBK)

Evaluation

Dataset

The dataset used in LightRAG can be downloaded from TommyChien/UltraDomain.

Generate Query

LightRAG uses the following prompt to generate high-level queries, with the corresponding code in example/generate_query.py.

Prompt

Given the following description of a dataset:

{description}

Please identify 5 potential users who would engage with this dataset. For each user, list 5 tasks they would perform with this dataset. Then, for each (user, task) combination, generate 5 questions that require a high-level understanding of the entire dataset.

Output the results in the following structure:
- User 1: [user description]
    - Task 1: [task description]
        - Question 1:
        - Question 2:
        - Question 3:
        - Question 4:
        - Question 5:
    - Task 2: [task description]
        ...
    - Task 5: [task description]
- User 2: [user description]
    ...
- User 5: [user description]
    ...

Batch Eval

To evaluate the performance of two RAG systems on high-level queries, LightRAG uses the following prompt, with the specific code available in example/batch_eval.py.

Prompt

---Role---
You are an expert tasked with evaluating two answers to the same question based on three criteria: **Comprehensiveness**, **Diversity**, and **Empowerment**.
---Goal---
You will evaluate two answers to the same question based on three criteria: **Comprehensiveness**, **Diversity**, and **Empowerment**.

- **Comprehensiveness**: How much detail does the answer provide to cover all aspects and details of the question?
- **Diversity**: How varied and rich is the answer in providing different perspectives and insights on the question?
- **Empowerment**: How well does the answer help the reader understand and make informed judgments about the topic?

For each criterion, choose the better answer (either Answer 1 or Answer 2) and explain why. Then, select an overall winner based on these three categories.

Here is the question:
{query}

Here are the two answers:

**Answer 1:**
{answer1}

**Answer 2:**
{answer2}

Evaluate both answers using the three criteria listed above and provide detailed explanations for each criterion.

Output your evaluation in the following JSON format:

{{
    "Comprehensiveness": {{
        "Winner": "[Answer 1 or Answer 2]",
        "Explanation": "[Provide explanation here]"
    }},
    "Empowerment": {{
        "Winner": "[Answer 1 or Answer 2]",
        "Explanation": "[Provide explanation here]"
    }},
    "Overall Winner": {{
        "Winner": "[Answer 1 or Answer 2]",
        "Explanation": "[Summarize why this answer is the overall winner based on the three criteria]"
    }}
}}

Overall Performance Table

	Agriculture		CS		Legal		Mix
	NaiveRAG	LightRAG	NaiveRAG	LightRAG	NaiveRAG	LightRAG	NaiveRAG	LightRAG
Comprehensiveness	32.69%	67.31%	35.44%	64.56%	19.05%	80.95%	36.36%	63.64%
Diversity	24.09%	75.91%	35.24%	64.76%	10.98%	89.02%	30.76%	69.24%
Empowerment	31.35%	68.65%	35.48%	64.52%	17.59%	82.41%	40.95%	59.05%
Overall	33.30%	66.70%	34.76%	65.24%	17.46%	82.54%	37.59%	62.40%
	RQ-RAG	LightRAG	RQ-RAG	LightRAG	RQ-RAG	LightRAG	RQ-RAG	LightRAG
Comprehensiveness	32.05%	67.95%	39.30%	60.70%	18.57%	81.43%	38.89%	61.11%
Diversity	29.44%	70.56%	38.71%	61.29%	15.14%	84.86%	28.50%	71.50%
Empowerment	32.51%	67.49%	37.52%	62.48%	17.80%	82.20%	43.96%	56.04%
Overall	33.29%	66.71%	39.03%	60.97%	17.80%	82.20%	39.61%	60.39%
	HyDE	LightRAG	HyDE	LightRAG	HyDE	LightRAG	HyDE	LightRAG
Comprehensiveness	24.39%	75.61%	36.49%	63.51%	27.68%	72.32%	42.17%	57.83%
Diversity	24.96%	75.34%	37.41%	62.59%	18.79%	81.21%	30.88%	69.12%
Empowerment	24.89%	75.11%	34.99%	65.01%	26.99%	73.01%	45.61%	54.39%
Overall	23.17%	76.83%	35.67%	64.33%	27.68%	72.32%	42.72%	57.28%
	GraphRAG	LightRAG	GraphRAG	LightRAG	GraphRAG	LightRAG	GraphRAG	LightRAG
Comprehensiveness	45.56%	54.44%	45.98%	54.02%	47.13%	52.87%	51.86%	48.14%
Diversity	19.65%	80.35%	39.64%	60.36%	25.55%	74.45%	35.87%	64.13%
Empowerment	36.69%	63.31%	45.09%	54.91%	42.81%	57.19%	52.94%	47.06%
Overall	43.62%	56.38%	45.98%	54.02%	45.70%	54.30%	51.86%	48.14%

Reproduce

All the code can be found in the ./reproduce directory.

Step-0 Extract Unique Contexts

First, we need to extract unique contexts in the datasets.

Code

def extract_unique_contexts(input_directory, output_directory):

    os.makedirs(output_directory, exist_ok=True)

    jsonl_files = glob.glob(os.path.join(input_directory, '*.jsonl'))
    print(f"Found {len(jsonl_files)} JSONL files.")

    for file_path in jsonl_files:
        filename = os.path.basename(file_path)
        name, ext = os.path.splitext(filename)
        output_filename = f"{name}_unique_contexts.json"
        output_path = os.path.join(output_directory, output_filename)

        unique_contexts_dict = {}

        print(f"Processing file: {filename}")

        try:
            with open(file_path, 'r', encoding='utf-8') as infile:
                for line_number, line in enumerate(infile, start=1):
                    line = line.strip()
                    if not line:
                        continue
                    try:
                        json_obj = json.loads(line)
                        context = json_obj.get('context')
                        if context and context not in unique_contexts_dict:
                            unique_contexts_dict[context] = None
                    except json.JSONDecodeError as e:
                        print(f"JSON decoding error in file {filename} at line {line_number}: {e}")
        except FileNotFoundError:
            print(f"File not found: {filename}")
            continue
        except Exception as e:
            print(f"An error occurred while processing file {filename}: {e}")
            continue

        unique_contexts_list = list(unique_contexts_dict.keys())
        print(f"There are {len(unique_contexts_list)} unique `context` entries in the file {filename}.")

        try:
            with open(output_path, 'w', encoding='utf-8') as outfile:
                json.dump(unique_contexts_list, outfile, ensure_ascii=False, indent=4)
            print(f"Unique `context` entries have been saved to: {output_filename}")
        except Exception as e:
            print(f"An error occurred while saving to the file {output_filename}: {e}")

    print("All files have been processed.")

Step-1 Insert Contexts

For the extracted contexts, we insert them into the LightRAG system.

Code

def insert_text(rag, file_path):
    with open(file_path, mode='r') as f:
        unique_contexts = json.load(f)

    retries = 0
    max_retries = 3
    while retries < max_retries:
        try:
            rag.insert(unique_contexts)
            break
        except Exception as e:
            retries += 1
            print(f"Insertion failed, retrying ({retries}/{max_retries}), error: {e}")
            time.sleep(10)
    if retries == max_retries:
        print("Insertion failed after exceeding the maximum number of retries")

Step-2 Generate Queries

We extract tokens from the first and the second half of each context in the dataset, then combine them as dataset descriptions to generate queries.

Code

tokenizer = GPT2Tokenizer.from_pretrained('gpt2')

def get_summary(context, tot_tokens=2000):
    tokens = tokenizer.tokenize(context)
    half_tokens = tot_tokens // 2

    start_tokens = tokens[1000:1000 + half_tokens]
    end_tokens = tokens[-(1000 + half_tokens):1000]

    summary_tokens = start_tokens + end_tokens
    summary = tokenizer.convert_tokens_to_string(summary_tokens)

    return summary

Step-3 Query

For the queries generated in Step-2, we will extract them and query LightRAG.

Code

def extract_queries(file_path):
    with open(file_path, 'r') as f:
        data = f.read()

    data = data.replace('**', '')

    queries = re.findall(r'- Question \d+: (.+)', data)

    return queries

Code Structure

.
├── examples
│   ├── batch_eval.py
│   ├── generate_query.py
│   ├── graph_visual_with_html.py
│   ├── graph_visual_with_neo4j.py
│   ├── lightrag_api_openai_compatible_demo.py
│   ├── lightrag_azure_openai_demo.py
│   ├── lightrag_bedrock_demo.py
│   ├── lightrag_hf_demo.py
│   ├── lightrag_lmdeploy_demo.py
│   ├── lightrag_ollama_demo.py
│   ├── lightrag_openai_compatible_demo.py
│   ├── lightrag_openai_demo.py
│   ├── lightrag_siliconcloud_demo.py
│   └── vram_management_demo.py
├── lightrag
│   ├── __init__.py
│   ├── base.py
│   ├── lightrag.py
│   ├── llm.py
│   ├── operate.py
│   ├── prompt.py
│   ├── storage.py
│   └── utils.py
├── reproduce
│   ├── Step_0.py
│   ├── Step_1_openai_compatible.py
│   ├── Step_1.py
│   ├── Step_2.py
│   ├── Step_3_openai_compatible.py
│   └── Step_3.py
├── .gitignore
├── .pre-commit-config.yaml
├── LICENSE
├── README.md
├── requirements.txt
└── setup.py

Star History

Citation

@article{guo2024lightrag,
title={LightRAG: Simple and Fast Retrieval-Augmented Generation},
author={Zirui Guo and Lianghao Xia and Yanhua Yu and Tu Ao and Chao Huang},
year={2024},
eprint={2410.05779},
archivePrefix={arXiv},
primaryClass={cs.IR}
}