目录

GraphRAG

GraphRAG 项目是一个数据管道和转换套件，旨在利用大型语言模型（LLMs）的力量从非结构化文本中提取有意义的结构化数据。

若要了解更多关于 GraphRAG 以及它如何用于增强您的大型语言模型（LLMs）对您的私有数据进行推理的能力，请访问 Microsoft Research Blog Post。

Get Started

构建虚拟环境

cd /Users/junjian/GitHub/microsoft/graphrag

python -m venv env
source env/bin/activate

安装 GraphRAG

pip install graphrag

准备数据

mkdir -p ./ragtest/input
curl https://www.gutenberg.org/cache/epub/24022/pg24022.txt > ./ragtest/input/book.txt

运行索引器

设置您的工作区变量

初始化您的工作区，运行 graphrag.index --init 命令

python -m graphrag.index --init --root ./ragtest

这将在 ./ragtest 目录中创建两个文件：.env 和 settings.yaml 。

  • .env 包含运行 GraphRAG pipeline 所需的环境变量。如果您检查该文件，您将看到定义了一个单一的环境变量，GRAPHRAG_API_KEY=<API_KEY> 。这是 OpenAI API 或 Azure OpenAI 端点的 API 密钥。您可以用您自己的 API 密钥替换它。
  • settings.yaml 包含 pipeline 的设置。您可以修改此文件来更改 pipeline 的设置。

修改配置 ragtest/settings.yaml

encoding_model: cl100k_base
skip_workflows: []
llm:
  api_key: ${GRAPHRAG_API_KEY}
  model: mistral
  api_base: http://127.0.0.1:11434/v1

embeddings:
  llm:
    api_key: ${GRAPHRAG_API_KEY}
    model: nomic-embed-text
    api_base: http://127.0.0.1:11434/v1
推理服务配置
推理服务 OpenAI API (api_base)
Ollama http://127.0.0.1:11434/v1
FastChat http://127.0.0.1:8000/v1
XInference http://127.0.0.1:9997/v1
模型配置
模型 (model) 类型 语言 下载
mistral:v0.2 llm 英文 ollama pull mistral:v0.2
aya:35b llm 多语言 ollama pull aya:35b
phi3:medium llm 英文 ollama pull phi3:medium
yi:9b llm 中英文 ollama pull yi:9b
codestral:22b llm 代码 ollama pull codestral:22b
nomic-embed-text embeddings 英文 ollama pull nomic-embed-text
mxbai-embed-large embeddings 英文 ollama pull mxbai-embed-large
bge-base-zh-v1.5 embeddings 中英文 ollama pull quentinz/bge-base-zh-v1.5
dmeta-embedding-zh embeddings 中英文 ollama pull herald/dmeta-embedding-zh

使用了 Ollama 的很多模型进行了测试，最终得到了上面的结果。

下面的模型在使用过程中出现了问题：

  • mistral:v0.3 最初发布的可以，最新的版本有问题。
  • llama3:8b
  • llama3.1:8b
  • llama3.1:70b
  • aya:8b
  • gemma2:9b
  • gemma2:27b
  • glm4:9b
  • internlm2:7b
  • yi:6b
  • yi:34b
  • qwen2:0.5b
  • qwen2:1.5b
  • qwen2:7b
  • qwen2:72b
  • deepseek-v2:16b
  • deepseek-coder-v2:16b
  • codeqwen:7b
  • codegeex4:9b
错误处理

❌ embeddings 使用 Ollama 容易出现错误 ZeroDivisionError，如下：

Error embedding chunk {'OpenAIEmbedding': "Error code: 400 - {'error': {'message': 'invalid input type', 'type': 'api_error', 'param': None, 'code': None}}"}
Traceback (most recent call last):
  File "/opt/miniconda/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/miniconda/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/__main__.py", line 76, in <module>
    run_local_search(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/cli.py", line 154, in run_local_search
    result = search_engine.search(query=query)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/structured_search/local_search/search.py", line 118, in search
    context_text, context_records = self.context_builder.build_context(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/structured_search/local_search/mixed_context.py", line 139, in build_context
    selected_entities = map_query_to_entities(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/context_builder/entity_extraction.py", line 55, in map_query_to_entities
    search_results = text_embedding_vectorstore.similarity_search_by_text(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/vector_stores/lancedb.py", line 118, in similarity_search_by_text
    query_embedding = text_embedder(text)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/context_builder/entity_extraction.py", line 57, in <lambda>
    text_embedder=lambda t: text_embedder.embed(t),
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/query/llm/oai/embedding.py", line 96, in embed
    chunk_embeddings = np.average(chunk_embeddings, axis=0, weights=chunk_lens)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/numpy/lib/function_base.py", line 550, in average
    raise ZeroDivisionError(
ZeroDivisionError: Weights sum to zero, can't be normalized

可以把 embeddings 的 api_base 换成 XInference的

❌ 使用 Qwen2:7b 模型在 create_base_entity_graph 阶段出现错误，如下：

datashaper.workflow.workflow ERROR Error executing verb "cluster_graph" in create_base_entity_graph: Columns must be same length as key
Traceback (most recent call last):
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/datashaper/workflow/workflow.py", line 410, in _execute_verb
    result = node.verb.func(**verb_args)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/cluster_graph.py", line 102, in cluster_graph
    output_df[[level_to, to]] = pd.DataFrame(
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/frame.py", line 4299, in __setitem__
    self._setitem_array(key, value)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/frame.py", line 4341, in _setitem_array
    check_key_length(self.columns, key, value)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/indexers/utils.py", line 391, in check_key_length
    raise ValueError("Columns must be same length as key")
ValueError: Columns must be same length as key

换上面的模型

❌ 使用 Llama3.1:8b 模型在 create_base_entity_graph 阶段出现错误，如下：

datashaper.workflow.workflow ERROR Error executing verb "cluster_graph" in create_base_entity_graph: EmptyNetworkError
Traceback (most recent call last):
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/datashaper/workflow/workflow.py", line 410, in _execute_verb
    result = node.verb.func(**verb_args)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/cluster_graph.py", line 61, in cluster_graph
    results = output_df[column].apply(lambda graph: run_layout(strategy, graph))
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/series.py", line 4924, in apply
    ).apply()
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/apply.py", line 1427, in apply
    return self.apply_standard()
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/apply.py", line 1507, in apply_standard
    mapped = obj._map_values(
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/base.py", line 921, in _map_values
    return algorithms.map_array(arr, mapper, na_action=na_action, convert=convert)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/pandas/core/algorithms.py", line 1743, in map_array
    return lib.map_infer(values, mapper, convert=convert)
  File "lib.pyx", line 2972, in pandas._libs.lib.map_infer
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/cluster_graph.py", line 61, in <lambda>
    results = output_df[column].apply(lambda graph: run_layout(strategy, graph))
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/cluster_graph.py", line 167, in run_layout
    clusters = run_leiden(graph, strategy)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/strategies/leiden.py", line 26, in run
    node_id_to_community_map = _compute_leiden_communities(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/verbs/graph/clustering/strategies/leiden.py", line 61, in _compute_leiden_communities
    community_mapping = hierarchical_leiden(
  File "<@beartype(graspologic.partition.leiden.hierarchical_leiden) at 0x319eddd80>", line 304, in hierarchical_leiden
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/graspologic/partition/leiden.py", line 588, in hierarchical_leiden
    hierarchical_clusters_native = gn.hierarchical_leiden(
leiden.EmptyNetworkError: EmptyNetworkError

换上面的模型

❌ 使用 Llama3.1:70b 模型在 create_final_community_reports 阶段出现错误，如下：

graphrag.index.graph.extractors.community_reports.community_reports_extractor ERROR error generating community report
Traceback (most recent call last):
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/index/graph/extractors/community_reports/community_reports_extractor.py", line 58, in __call__
    await self._llm(
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/openai/json_parsing_llm.py", line 34, in __call__
    result = await self._delegate(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/openai/openai_token_replacing_llm.py", line 37, in __call__
    return await self._delegate(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/openai/openai_history_tracking_llm.py", line 33, in __call__
    output = await self._delegate(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/caching_llm.py", line 104, in __call__
    result = await self._delegate(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/rate_limiting_llm.py", line 177, in __call__
    result, start = await execute_with_retry()
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/rate_limiting_llm.py", line 159, in execute_with_retry
    async for attempt in retryer:
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/tenacity/asyncio/__init__.py", line 166, in __anext__
    do = await self.iter(retry_state=self._retry_state)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/tenacity/asyncio/__init__.py", line 153, in iter
    result = await action(retry_state)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/tenacity/_utils.py", line 99, in inner
    return call(*args, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/env/lib/python3.10/site-packages/tenacity/__init__.py", line 398, in <lambda>
    self._add_action_func(lambda rs: rs.outcome.result())
  File "/opt/miniconda/lib/python3.10/concurrent/futures/_base.py", line 451, in result
    return self.__get_result()
  File "/opt/miniconda/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
    raise self._exception
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/rate_limiting_llm.py", line 165, in execute_with_retry
    return await do_attempt(), start
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/rate_limiting_llm.py", line 147, in do_attempt
    return await self._delegate(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/base/base_llm.py", line 48, in __call__
    return await self._invoke_json(input, **kwargs)
  File "/Users/junjian/GitHub/microsoft/graphrag/graphrag/llm/openai/openai_chat_llm.py", line 90, in _invoke_json
    raise RuntimeError(FAILED_TO_CREATE_JSON_ERROR)
RuntimeError: Failed to generate valid JSON output

换上面的模型

运行索引 pipeline

python -m graphrag.index --root ./ragtest

⠦ GraphRAG Indexer 
├── Loading Input (text) - 1 files loaded (0 filtered) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00 0:00:00
├── create_base_text_units
├── create_base_extracted_entities
├── create_summarized_entities
├── create_base_entity_graph
├── create_final_entities
├── create_final_nodes
├── create_final_communities
├── join_text_units_to_entity_ids
├── create_final_relationships
├── join_text_units_to_relationship_ids
├── create_final_community_reports
├── create_final_text_units
├── create_base_documents
└── create_final_documents
🚀 All workflows completed successfully.

此过程运行将需要一些时间。这取决于您的输入数据的大小、您正在使用的模型以及所使用的文本块大小（这些可以在您的 .env 文件中进行配置）。一旦 pipeline 完成，您应该会看到一个名为 ./ragtest/output/<timestamp>/artifacts 的新文件夹，其中包含一系列的 parquet 文件。

运行查询引擎

使用 全局(global)搜索 提出高层次问题

What are the top themes in this story?（这个故事的主题是什么？）

python -m graphrag.query \
    --root ./ragtest \
    --method global \
    "What are the top themes in this story?"

INFO: Reading settings from ragtest/settings.yaml
creating llm client with {'api_key': 'REDACTED,len=9', 'type': "openai_chat", 'model': 'mistral:v0.3', 'max_tokens': 4000, 'temperature': 0.0, 'top_p': 1.0, 'n': 1, 'request_timeout': 180.0, 'api_base': 'http://127.0.0.1:11434/v1', 'api_version': None, 'organization': None, 'proxy': None, 'cognitive_services_endpoint': None, 'deployment_name': None, 'model_supports_json': True, 'tokens_per_minute': 0, 'requests_per_minute': 0, 'max_retries': 10, 'max_retry_wait': 10.0, 'sleep_on_rate_limit_recommendation': True, 'concurrent_requests': 25}

SUCCESS: Global Search Response:  The story primarily revolves around the themes of redemption and transformation, as depicted in Charles Dickens' classic tale 'A Christmas Carol'. This narrative centers around Ebenezer Scrooge, a man who undergoes a profound change after being visited by three ghosts on Christmas Eve [Data: Reports 1].

In addition to the story, the Verdant Oasis Plaza community benefits from a digital library that offers electronic resources such as books and texts for free use under the Project Gutenberg License [Data: Reports 2].

Furthermore, the Unity March taking place in Verdant Oasis Plaza has garnered media attention from Tribune Spotlight, suggesting it may have a significant impact on the community dynamics [Data: Reports 3]. This event could potentially shape the social landscape of the area.

使用 本地(local)搜索 针对特定角色提出更具体问题

Who is Scrooge, and what are his main relationships?（斯克鲁奇是谁，他的主要关系是什么？）

python -m graphrag.query \
    --root ./ragtest \
    --method local \
    "Who is Scrooge, and what are his main relationships?"

INFO: Reading settings from ragtest/settings.yaml
creating llm client with {'api_key': 'REDACTED,len=9', 'type': "openai_chat", 'model': 'mistral:v0.3', 'max_tokens': 4000, 'temperature': 0.0, 'top_p': 1.0, 'n': 1, 'request_timeout': 180.0, 'api_base': 'http://127.0.0.1:11434/v1', 'api_version': None, 'organization': None, 'proxy': None, 'cognitive_services_endpoint': None, 'deployment_name': None, 'model_supports_json': True, 'tokens_per_minute': 0, 'requests_per_minute': 0, 'max_retries': 10, 'max_retry_wait': 10.0, 'sleep_on_rate_limit_recommendation': True, 'concurrent_requests': 25}
creating embedding llm client with {'api_key': 'REDACTED,len=9', 'type': "openai_embedding", 'model': 'bge-base-zh-v1.5', 'max_tokens': 4000, 'temperature': 0, 'top_p': 1, 'n': 1, 'request_timeout': 180.0, 'api_base': 'http://127.0.0.1:9997/v1', 'api_version': None, 'organization': None, 'proxy': None, 'cognitive_services_endpoint': None, 'deployment_name': None, 'model_supports_json': None, 'tokens_per_minute': 0, 'requests_per_minute': 0, 'max_retries': 10, 'max_retry_wait': 10.0, 'sleep_on_rate_limit_recommendation': True, 'concurrent_requests': 25}

SUCCESS: Local Search Response:  Ebenezer Scrooge is a fictional character created by Charles Dickens, who first appeared in the novel "A Christmas Carol," published in 1843. Scrooge is a miserly old man who values money above all else and is known for being cold-hearted and uncaring towards others.

Throughout the story, we learn about Scrooge's main relationships:

1. Jacob Marley: Scrooge's former business partner who died seven years prior to the events of the novel. Marley appears to Scrooge as a ghostly figure and warns him about his fate if he does not change his ways.

2. Bob Cratchit: Scrooge employs Bob Cratchit as his clerk, and they have a master-servant relationship. Scrooge is harsh towards Cratchit, but after his transformation, he becomes kinder and more generous to him.

3. Tiny Tim: Bob Cratchit's youngest son, who is sickly and disabled. Scrooge initially shows no concern for Tiny Tim's well-being but later expresses regret for not having been more compassionate towards him.

4. Fred: Scrooge's nephew, who invites him to Christmas dinner every year but is always rejected. Despite this, Fred remains friendly and hopeful that his uncle will change. After Scrooge's transformation, he welcomes Fred warmly into his life.

5. Belle: Scrooge was once engaged to Belle, but their relationship ended due to Scrooge's obsession with money. She represents the love and warmth that Scrooge has lost over the years.

6. The Ghosts of Christmas Past, Present, and Yet to Come: These supernatural beings visit Scrooge on Christmas Eve and show him visions of his past, present, and future, respectively. They play a crucial role in helping Scrooge understand the error of his ways and inspiring him to change.

OpenAI API

Ollama

  • completions 
    curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
      "model": "mistral:v0.3",
      "messages": [
          {
              "role": "user",
              "content": "Hello!"
          }
      ]
  }'|jq

    {
"id": "chatcmpl-14",
"object": "chat.completion",
"created": 1722220031,
"model": "mistral:v0.3",
"system_fingerprint": "fp_ollama",
"choices": [
  {
    "index": 0,
    "message": {
      "role": "assistant",
      "content": " Hello there! How can I assist you today? Feel free to ask me anything related to programming, data analysis, or machine learning. I'm here to help!\n\nIf you have any general questions or just need someone to bounce ideas off of, I'll do my best to help out or guide you in the right direction. Let's get started! 😊"
    },
    "finish_reason": "stop"
  }
],
"usage": {
  "prompt_tokens": 7,
  "completion_tokens": 79,
  "total_tokens": 86
}
}
  • embeddings 
    curl http://localhost:11434/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
      "input": "Hello!",
      "model": "nomic-embed-text"
  }'|jq

    {
"object": "list",
"data": [
  {
    "object": "embedding",
    "embedding": [
      0.020163044,
      -0.0036894183,
      //......
      -0.04950454,
      -0.010347797
    ],
    "index": 0
  }
],
"model": "nomic-embed-text"
}

XIinference

  • completions 
    curl http://localhost:9997/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
      "model": "qwen-chat",
      "messages": [
          {
              "role": "user",
              "content": "Hello!"
          }
      ]
  }'|jq

    {
"id": "chat27fca670-4d54-11ef-bbf6-8e835058c3ee",
"object": "chat.completion",
"created": 1722220931,
"model": "qwen-chat",
"choices": [
  {
    "index": 0,
    "message": {
      "role": "assistant",
      "content": "Hello! How can I help you today?"
    },
    "finish_reason": "stop"
  }
],
"usage": {
  "prompt_tokens": 21,
  "completion_tokens": 9,
  "total_tokens": 30
}
}
  • embeddings 
    curl http://127.0.0.1:9997/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
      "input": "Hello!",
      "model": "bge-base-zh"
  }'|jq

    {
"object": "list",
"model": "bge-base-zh-1-0",
"data": [
  {
    "index": 0,
    "object": "embedding",
    "embedding": [
      0.007897183299064636,
      -0.012519387528300285,
      //......
      -0.01635596714913845,
      -0.08653949201107025
    ]
  }
],
"usage": {
  "prompt_tokens": 37,
  "total_tokens": 37
}
}

参考资料