【AIGC学习笔记】——GraphRAG当地摆设

【GraphRAG+Ollama当地摆设】奇怪滚热辣的小白利用

---

  

---

情况准备+源码下载

1.利用系统：Ubuntu20.04
2.VSCode编译器
3.Python 3.11
4.Ollama安装
5.GraphRAG代码下载：https://github.com/microsoft/graphrag.git
一、Anaconda虚拟情况创建及配置

1.conda create --name graphR python=3.11
2.pip install graphrag==0.3.6（新版本的graphrag轻易报错：No module named graphrag.index.main）
3.pip install ollama
（接下来利用期间若出现未提及包的未安装提示，就按照提示安装即可）
二、Ollama模型下载

1.ollama serve （启动ollama）
2.ollama pull mistral：v0.2
3.ollama pull nomic-embed-text:latest
三、创建数据目次

在graphrag代码文件夹中创建一个ragtest文件夹，并在ragtest文件夹中创建一个input文件夹，把txt数据放在input文件夹中
四、项目初始化

1.python -m graphrag.index --init --root ./ragtest
（在graphrag代码文件夹中打开终端，并激活graphR虚拟情况，运行上述代码，随后在ragtest文件夹中会出现setting.yaml，prompts，.env等文件，正常情况下是有图中6个东西的，但是有时候一开始只有几个，背面创建索引的时候其余的也会主动天生）

四、修改配置文件settings.yaml

按照下面的代码修改即可

1. encoding_model: cl100k_base
2. skip_workflows: []
3. llm:
4.   api_key: ollama
5.   type: openai_chat # or azure_openai_chat
6.   model: mistral:v0.2
7.   model_supports_json: true # recommended if this is available for your model.
8.   max_tokens: 1024
9.   # request_timeout: 180.0
10.   api_base: http://localhost:11434/v1
11.   # api_version: 2024-02-15-preview
12.   # organization: <organization_id>
13.   # deployment_name: <azure_model_deployment_name>
14.   # tokens_per_minute: 150_000 # set a leaky bucket throttle
15.   # requests_per_minute: 10_000 # set a leaky bucket throttle
16.   # max_retries: 10
17.   # max_retry_wait: 10.0
18.   # sleep_on_rate_limit_recommendation: true # whether to sleep when azure suggests wait-times
19.   # concurrent_requests: 25 # the number of parallel inflight requests that may be made
20.   # temperature: 0 # temperature for sampling
21.   # top_p: 1 # top-p sampling
22.   # n: 1 # Number of completions to generate
24. parallelization:
25.   stagger: 0.3
26.   # num_threads: 50 # the number of threads to use for parallel processing
28. async_mode: threaded # or asyncio
30. embeddings:
31.   ## parallelization: override the global parallelization settings for embeddings
32.   async_mode: threaded # or asyncio
33.   # target: required # or all
34.   # batch_size: 16 # the number of documents to send in a single request
35.   # batch_max_tokens: 8191 # the maximum number of tokens to send in a single request
36.   llm:
37.     api_key: ollama
38.     type: openai_embedding # or azure_openai_embedding
39.     model: nomic-embed-text:latest
40.     api_base: http://localhost:11434/api
41.     # api_version: 2024-02-15-preview
42.     # organization: <organization_id>
43.     # deployment_name: <azure_model_deployment_name>
44.     # tokens_per_minute: 150_000 # set a leaky bucket throttle
45.     # requests_per_minute: 10_000 # set a leaky bucket throttle
46.     # max_retries: 10
47.     # max_retry_wait: 10.0
48.     # sleep_on_rate_limit_recommendation: true # whether to sleep when azure suggests wait-times
49.     # concurrent_requests: 25 # the number of parallel inflight requests that may be made
51. chunks:
52.   size: 200
53.   overlap: 100
54.   group_by_columns: [id] # by default, we don't allow chunks to cross documents
56. input:
57.   type: file # or blob
58.   file_type: text # or csv
59.   base_dir: "input"
60.   file_encoding: utf-8
61.   file_pattern: ".*\\.txt$"
63. cache:
64.   type: file # or blob
65.   base_dir: "cache"
66.   # connection_string: <azure_blob_storage_connection_string>
67.   # container_name: <azure_blob_storage_container_name>
69. storage:
70.   type: file # or blob
71.   base_dir: "output/${timestamp}/artifacts"
72.   # connection_string: <azure_blob_storage_connection_string>
73.   # container_name: <azure_blob_storage_container_name>
75. reporting:
76.   type: file # or console, blob
77.   base_dir: "output/${timestamp}/reports"
78.   # connection_string: <azure_blob_storage_connection_string>
79.   # container_name: <azure_blob_storage_container_name>
81. entity_extraction:
82.   ## strategy: fully override the entity extraction strategy.
83.   ##   type: one of graph_intelligence, graph_intelligence_json and nltk
84.   ## llm: override the global llm settings for this task
85.   ## parallelization: override the global parallelization settings for this task
86.   ## async_mode: override the global async_mode settings for this task
87.   prompt: "prompts/entity_extraction.txt"
88.   entity_types: [organization,person,geo,event]
89.   max_gleanings: 0
91. summarize_descriptions:
92.   ## llm: override the global llm settings for this task
93.   ## parallelization: override the global parallelization settings for this task
94.   ## async_mode: override the global async_mode settings for this task
95.   prompt: "prompts/summarize_descriptions.txt"
96.   max_length: 500
98. claim_extraction:
99.   ## llm: override the global llm settings for this task
100.   ## parallelization: override the global parallelization settings for this task
101.   ## async_mode: override the global async_mode settings for this task
102.   # enabled: true
103.   prompt: "prompts/claim_extraction.txt"
104.   description: "Any claims or facts that could be relevant to information discovery."
105.   max_gleanings: 0
107. community_reports:
108.   ## llm: override the global llm settings for this task
109.   ## parallelization: override the global parallelization settings for this task
110.   ## async_mode: override the global async_mode settings for this task
111.   prompt: "prompts/community_report.txt"
112.   max_length: 2000
113.   max_input_length: 8000
115. cluster_graph:
116.   max_cluster_size: 10
118. embed_graph:
119.   enabled: false # if true, will generate node2vec embeddings for nodes
120.   # num_walks: 10
121.   # walk_length: 40
122.   # window_size: 2
123.   # iterations: 3
124.   # random_seed: 597832
126. umap:
127.   enabled: false # if true, will generate UMAP embeddings for nodes
129. snapshots:
130.   graphml: false
131.   raw_entities: false
132.   top_level_nodes: false
134. local_search:
135.   # text_unit_prop: 0.5
136.   # community_prop: 0.1
137.   # conversation_history_max_turns: 5
138.   # top_k_mapped_entities: 10
139.   # top_k_relationships: 10
140.   # llm_temperature: 0 # temperature for sampling
141.   # llm_top_p: 1 # top-p sampling
142.   # llm_n: 1 # Number of completions to generate
143.   # max_tokens: 12000
145. global_search:
146.   # llm_temperature: 0 # temperature for sampling
147.   # llm_top_p: 1 # top-p sampling
148.   # llm_n: 1 # Number of completions to generate
149.   # max_tokens: 12000
150.   # data_max_tokens: 12000
151.   # map_max_tokens: 1000
152.   # reduce_max_tokens: 2000
153.   # concurrency: 32

复制代码

五、修改.env文件

按照下面的代码修改即可

1. GRAPHRAG_API_KEY=ollama
2. GRAPHRAG_CLAIM_EXTRACTION_ENABLED=True

复制代码

六、修改虚拟情况graphR里graphrag包的源码

1.找到graphrag包位置：/home//.conda/envs/graphR/lib/python3.11/site-packages
2.修改第一个文件：/home//.conda/envs/graphR/lib/python3.11/site-packages/graphrag/llm/openai/openai_embeddings_llm.py

1. """The EmbeddingsLLM class."""
3. from typing_extensions import Unpack
5. from graphrag.llm.base import BaseLLM
6. from graphrag.llm.types import (
7.     EmbeddingInput,
8.     EmbeddingOutput,
9.     LLMInput,
10. )
12. from .openai_configuration import OpenAIConfiguration
13. from .types import OpenAIClientTypes
14. import ollama # 增加依赖
17. class OpenAIEmbeddingsLLM(BaseLLM[EmbeddingInput, EmbeddingOutput]):
18.     """A text-embedding generator LLM."""
20.     _client: OpenAIClientTypes
21.     _configuration: OpenAIConfiguration
23.     def __init__(self, client: OpenAIClientTypes, configuration: OpenAIConfiguration):
24.         self.client = client
25.         self.configuration = configuration
27.     async def _execute_llm(
28.         self, input: EmbeddingInput, **kwargs: Unpack[LLMInput]
29.     ) -> EmbeddingOutput | None:
30.         args = {
31.             "model": self.configuration.model,
32.             **(kwargs.get("model_parameters") or {}),
33.         }
34.         # 修改此处
35.         #embedding = await self.client.embeddings.create(
36.         #    input=input,
37.         #    **args,
38.         #)
39.         #return [d.embedding for d in embedding.data]
40.         
41.         embedding_list = []
42.         for inp in input:
43.             embedding = ollama.embeddings(model="nomic-embed-text:latest", prompt=inp)
44.             embedding_list.append(embedding["embedding"])
45.         return embedding_list

复制代码

3.修改第二个文件：/home/***/.conda/envs/graphR/lib/python3.11/site-packages/graphrag/query/llm/oai/embedding.py

1. """OpenAI Embedding model implementation."""
3. import asyncio
4. from collections.abc import Callable
5. from typing import Any
7. import numpy as np
8. import tiktoken
9. from tenacity import (
10.     AsyncRetrying,
11.     RetryError,
12.     Retrying,
13.     retry_if_exception_type,
14.     stop_after_attempt,
15.     wait_exponential_jitter,
16. )
18. from graphrag.logging import StatusLogger
19. from graphrag.query.llm.base import BaseTextEmbedding
20. from graphrag.query.llm.oai.base import OpenAILLMImpl
21. from graphrag.query.llm.oai.typing import (
22.     OPENAI_RETRY_ERROR_TYPES,
23.     OpenaiApiType,
24. )
25. from graphrag.query.llm.text_utils import chunk_text
26. # 增加依赖
27. import ollama
30. class OpenAIEmbedding(BaseTextEmbedding, OpenAILLMImpl):
31.     """Wrapper for OpenAI Embedding models."""
33.     def __init__(
34.         self,
35.         api_key: str | None = None,
36.         azure_ad_token_provider: Callable | None = None,
37.         model: str = "text-embedding-3-small",
38.         deployment_name: str | None = None,
39.         api_base: str | None = None,
40.         api_version: str | None = None,
41.         api_type: OpenaiApiType = OpenaiApiType.OpenAI,
42.         organization: str | None = None,
43.         encoding_name: str = "cl100k_base",
44.         max_tokens: int = 8191,
45.         max_retries: int = 10,
46.         request_timeout: float = 180.0,
47.         retry_error_types: tuple[type[BaseException]] = OPENAI_RETRY_ERROR_TYPES,  # type: ignore
48.         reporter: StatusLogger | None = None,
49.     ):
50.         OpenAILLMImpl.__init__(
51.             self=self,
52.             api_key=api_key,
53.             azure_ad_token_provider=azure_ad_token_provider,
54.             deployment_name=deployment_name,
55.             api_base=api_base,
56.             api_version=api_version,
57.             api_type=api_type,  # type: ignore
58.             organization=organization,
59.             max_retries=max_retries,
60.             request_timeout=request_timeout,
61.             reporter=reporter,
62.         )
64.         self.model = model
65.         self.encoding_name = encoding_name
66.         self.max_tokens = max_tokens
67.         self.token_encoder = tiktoken.get_encoding(self.encoding_name)
68.         self.retry_error_types = retry_error_types
70.     def embed(self, text: str, **kwargs: Any) -> list[float]:
71.         """
72.         Embed text using OpenAI Embedding's sync function.
74.         For text longer than max_tokens, chunk texts into max_tokens, embed each chunk, then combine using weighted average.
75.         Please refer to: https://github.com/openai/openai-cookbook/blob/main/examples/Embedding_long_inputs.ipynb
76.         """
77.         token_chunks = chunk_text(
78.             text=text, token_encoder=self.token_encoder, max_tokens=self.max_tokens
79.         )
80.         chunk_embeddings = []
81.         chunk_lens = []
82.         for chunk in token_chunks:
83.             try:
84.                 #embedding, chunk_len = self._embed_with_retry(chunk, **kwargs)
85.                 #修改embedding、chunk_len
86.                 embedding = ollama.embeddings(model='nomic-embed-text:latest', prompt=chunk)['embedding']
87.                 chunk_len = len(chunk)
88.                 chunk_embeddings.append(embedding)
89.                 chunk_lens.append(chunk_len)
90.             # TODO: catch a more specific exception
91.             except Exception as e:  # noqa BLE001
92.                 self._reporter.error(
93.                     message="Error embedding chunk",
94.                     details={self.__class__.__name__: str(e)},
95.                 )
97.                 continue
98.         #chunk_embeddings = np.average(chunk_embeddings, axis=0, weights=chunk_lens)
99.         #chunk_embeddings = chunk_embeddings / np.linalg.norm(chunk_embeddings)
100.         #return chunk_embeddings.tolist()
101.         return chunk_embeddings
103.    
105.     async def aembed(self, text: str, **kwargs: Any) -> list[float]:
106.         """
107.         Embed text using OpenAI Embedding's async function.
109.         For text longer than max_tokens, chunk texts into max_tokens, embed each chunk, then combine using weighted average.
110.         """
111.         token_chunks = chunk_text(
112.             text=text, token_encoder=self.token_encoder, max_tokens=self.max_tokens
113.         )
114.         chunk_embeddings = []
115.         chunk_lens = []
116.         embedding_results = await asyncio.gather(*[
117.             self._aembed_with_retry(chunk, **kwargs) for chunk in token_chunks
118.         ])
119.         embedding_results = [result for result in embedding_results if result[0]]
120.         chunk_embeddings = [result[0] for result in embedding_results]
121.         chunk_lens = [result[1] for result in embedding_results]
122.         chunk_embeddings = np.average(chunk_embeddings, axis=0, weights=chunk_lens)  # type: ignore
123.         chunk_embeddings = chunk_embeddings / np.linalg.norm(chunk_embeddings)
124.         return chunk_embeddings.tolist()
126.     def _embed_with_retry(
127.         self, text: str | tuple, **kwargs: Any
128.     ) -> tuple[list[float], int]:
129.         try:
130.             retryer = Retrying(
131.                 stop=stop_after_attempt(self.max_retries),
132.                 wait=wait_exponential_jitter(max=10),
133.                 reraise=True,
134.                 retry=retry_if_exception_type(self.retry_error_types),
135.             )
136.             for attempt in retryer:
137.                 with attempt:
138.                     embedding = (
139.                         self.sync_client.embeddings.create(  # type: ignore
140.                             input=text,
141.                             model=self.model,
142.                             **kwargs,  # type: ignore
143.                         )
144.                         .data[0]
145.                         .embedding
146.                         or []
147.                     )
148.                     return (embedding, len(text))
149.         except RetryError as e:
150.             self._reporter.error(
151.                 message="Error at embed_with_retry()",
152.                 details={self.__class__.__name__: str(e)},
153.             )
154.             return ([], 0)
155.         else:
156.             # TODO: why not just throw in this case?
157.             return ([], 0)
159.     async def _aembed_with_retry(
160.         self, text: str | tuple, **kwargs: Any
161.     ) -> tuple[list[float], int]:
162.         try:
163.             retryer = AsyncRetrying(
164.                 stop=stop_after_attempt(self.max_retries),
165.                 wait=wait_exponential_jitter(max=10),
166.                 reraise=True,
167.                 retry=retry_if_exception_type(self.retry_error_types),
168.             )
169.             async for attempt in retryer:
170.                 with attempt:
171.                     embedding = (
172.                         await self.async_client.embeddings.create(  # type: ignore
173.                             input=text,
174.                             model=self.model,
175.                             **kwargs,  # type: ignore
176.                         )
177.                     ).data[0].embedding or []
178.                     return (embedding, len(text))
179.         except RetryError as e:
180.             self._reporter.error(
181.                 message="Error at embed_with_retry()",
182.                 details={self.__class__.__name__: str(e)},
183.             )
184.             return ([], 0)
185.         else:
186.             # TODO: why not just throw in this case?
187.             return ([], 0)

复制代码

4.修改第三个文件：/home/***/.conda/envs/graphR/lib/python3.11/site-packages/graphrag/query/llm/text_utils.py

1. """Text Utilities for LLM."""
3. from collections.abc import Iterator
4. from itertools import islice
6. import tiktoken
9. def num_tokens(text: str, token_encoder: tiktoken.Encoding | None = None) -> int:
10.     """Return the number of tokens in the given text."""
11.     if token_encoder is None:
12.         token_encoder = tiktoken.get_encoding("cl100k_base")
13.     return len(token_encoder.encode(text))  # type: ignore
16. def batched(iterable: Iterator, n: int):
17.     """
18.     Batch data into tuples of length n. The last batch may be shorter.
20.     Taken from Python's cookbook: https://docs.python.org/3/library/itertools.html#itertools.batched
21.     """
22.     # batched('ABCDEFG', 3) --> ABC DEF G
23.     if n < 1:
24.         value_error = "n must be at least one"
25.         raise ValueError(value_error)
26.     it = iter(iterable)
27.     while batch := tuple(islice(it, n)):
28.         yield batch
31. def chunk_text(
32.     text: str, max_tokens: int, token_encoder: tiktoken.Encoding | None = None
33. ):
34.     """Chunk text by token length."""
35.     if token_encoder is None:
36.         token_encoder = tiktoken.get_encoding("cl100k_base")
37.     tokens = token_encoder.encode(text)  # type: ignore
38.     # 增加下行代码，将tokens解码成字符串
39.     tokens = token_encoder.decode(tokens)
40.     chunk_iterator = batched(iter(tokens), max_tokens)
41.     #yield from (token_encoder.decode(list(chunk)) for chunk in chunk_iterator)
42.     yield from chunk_iterator

复制代码

七、创建索引

1.打开ollama，可以看到模型运行状态
2.python -m graphrag.index --root ./ragtest
（回到第四步打开的终端页面，运行这行代码，并出现图中效果即创建完成）

八、开始查询

1.局部查询：python -m graphrag.query --root ./ragtest --method local "who is Marley?"

2.全局查询：python -m graphrag.query --root ./ragtest --method global "who is Marley?"

（留意：我在查询的时候会出现没有graphrag.logging的错误提示，直接把源码中graphrag文件夹里面的logging文件夹复制到虚拟情况中graphrag包里相应位置就解决了）
参考博文：
1.https://blog.csdn.net/weixin_42107217/article/details/141649920
2.https://blog.csdn.net/gaotianhao123/article/details/140640415
3.https://blog.csdn.net/m0_56378800/article/details/140319467

免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。