nano-GraphRAG

A simple, easy-to-hack GraphRAG implementation

😭 GraphRAG is good and powerful, but the official implementation is difficult/painful to read or hack.

😊 This project provides a smaller, faster, cleaner GraphRAG, while remaining the core functionality(see benchmark and issues ).

🎁 Excluding tests and prompts, nano-graphrag is about 800 lines of code.

👌 Small yet portable, asynchronous and fully typed.

Install

Install from PyPi

pip install nano-graphrag

Install from source

# clone this repo first
cd nano-graphrag
pip install -e .

Quick Start

Please set OpenAI API key in environment: export OPENAI_API_KEY="sk-...".

Tip

If you like to use another LLM: LLM component.

If you like to use another Embedding Model: Embedding.

download a copy of 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:

from nano_graphrag import GraphRAG, QueryParam

graph_func = GraphRAG(working_dir="./dickens")

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

# Perform global graphrag search
print(graph_func.query("What are the top themes in this story?"))

# Perform local graphrag search (I think is better and more scalable one)
print(graph_func.query("What are the top themes in this story?", param=QueryParam(mode="local")))

Next time you initialize a GraphRAG from the same working_dir, it will reload all the contexts automatically.

Incremental Insert

nano-graphrag supports incremental insert, no duplicated computation or data will be added:

with open("./book.txt") as f:
    book = f.read()
    half_len = len(book) // 2
    graph_func.insert(book[:half_len])
    graph_func.insert(book[half_len:])

nano-graphrag use md5-hash of the content as the key, so there is no duplicated chunk.

However, each time you insert, the communities of graph will be re-computed and the community reports will be re-generated

Async

For each method NAME(...) , there is a corresponding async method aNAME(...)

await graph_func.ainsert(...)
await graph_func.aquery(...)
...

Available Parameters

GraphRAG and QueryParam are dataclass in Python. Use help(GraphRAG) and help(QueryParam) to see all available parameters!

Advances

Prompt

nano-graphrag use prompts from nano_graphrag.prompt.PROMPTS dict object. You can play with it and replace any prompt inside.

Some important prompts:

• PROMPTS["entity_extraction"] is used to extract the entities and relations from a text chunk.
• PROMPTS["community_report"] is used to organize and summary the graph cluster's description.
• PROMPTS["local_rag_response"] is the system prompt template of the local search generation.
• PROMPTS["global_reduce_rag_response"] is the system prompt template of the global search generation.

Storage

You can replace all storage-related components to your own implementation, nano-graphrag mainly uses three kinds of storage:

• base.BaseKVStorage for storing key-json pairs of data.
  • By default we use disk file storage as the backend.
  • GraphRAG(.., key_string_value_json_storage_cls=YOURS,...)
• base.BaseVectorStorage for indexing embeddings.
  • By default we use milvus-lite as the backend.
  • GraphRAG(.., vector_db_storage_cls=YOURS,...)
• base.BaseGraphStorage for storing knowledge graph.
  • By default we use networkx as the backend.
  • GraphRAG(.., graph_storage_cls=YOURS,...)

You can refer to nano_graphrag.base to see detailed interfaces for each components.

LLM

In nano-graphrag, we requires two types of LLM, a great one and a cheap one. The former is used to plan and respond, the latter is used to summary. By default, the great one is gpt-4o and the cheap one is gpt-4o-mini

You can implement your own LLM function (refer to _llm.gpt_4o_complete):

async def my_llm_complete(
    prompt, system_prompt=None, history_messages=[], **kwargs
) -> str:
  # pop cache KV database if any
  hashing_kv: BaseKVStorage = kwargs.pop("hashing_kv", None)
  # the rest kwargs are for calling LLM, for example, `max_tokens=xxx`
	...
  # YOUR LLM calling
  response = await call_your_LLM(messages, **kwargs)
  return response

Replace the default one with:

# Adjust the max token size or the max async requests if needed
GraphRAG(best_model_func=my_llm_complete, best_model_max_token_size=..., best_model_max_async=...)
GraphRAG(cheap_model_func=my_llm_complete, cheap_model_max_token_size=..., cheap_model_max_async=...)

You can refer to an example that use deepseek-chat as the LLM model in nano-graphrag

Embedding

You can replace the default embedding functions with any _utils.EmbedddingFunc instance.

For example, the default one is using OpenAI embedding API:

@wrap_embedding_func_with_attrs(embedding_dim=1536, max_token_size=8192)
async def openai_embedding(texts: list[str]) -> np.ndarray:
    openai_async_client = AsyncOpenAI()
    response = await openai_async_client.embeddings.create(
        model="text-embedding-3-small", input=texts, encoding_format="float"
    )
    return np.array([dp.embedding for dp in response.data])

Replace default embedding function with:

GraphRAG(embedding_func=your_embed_func, embedding_batch_num=..., embedding_func_max_async=...)

You can refer to an example that use sentence-transformer to locally compute embeddings.

Benchmark

• benchmark for English
• benchmark for Chinese

Issues

• nano-graphrag didn't implement the covariates feature of GraphRAG
• nano-graphrag implements the global search different from the original. The original use a map-reduce-like style to fill all the communities into context, while nano-graphrag only use the top-K important and central communites (use QueryParam.global_max_conside_community to control, default to 512 communities).

TODO in Next Version

If a checkbox is filled meaning someone is on it

• nano-graphrag's Data Source Id is local, meaning it always starts at 0 at any response and you have to remap it into the current session. So it's kinda useless right now.
• nano-graphrag truncates the community's raw description if it exceed the maximun context size when generating community report, while GraphRAG uses a sub-communities iterative summary to include all.
• Add real benchmark with GraphRAG
• Add new components, see issue