【大模子开发指南】llamaindex配置deepseek、jina embedding及chromadb实现 ...

说一些坑，原来之前准备用milvus，但是发现win搞不了（docker都配好了）。然后转头搞chromadb。这里面尚有就是embedding一般都是本地摆设，但我电脑是cpu的没法玩，我就选了jina的embedding性能较优（也可以换glm的embedding但是要改代码）。最后问题出在deepseek与llamaindex的适配，因为接纳openai的接口，这里面改了openai库的源码然后对llamaindex加了配置项才完全跑通。国内小伙伴如果使用我这套方案可以抄，给我点个赞谢谢。
  主要环境：

1. os：win11
2. python3.10
3. llamaindex  0.11.20
4. chromadb   0.5.15
5. 这个文件是官方例子，自己弄个也成

复制代码

源码如下：

1. # %%
2. from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
3. from llama_index.vector_stores.chroma import ChromaVectorStore
4. from llama_index.core import StorageContext
5. from llama_index.embeddings.huggingface import HuggingFaceEmbedding
6. from IPython.display import Markdown, display
8. from llama_index.llms.openai import OpenAI
9. import chromadb
11. # %%
13. import openai
14. openai.api_key = "sk"
16. openai.api_base = "https://api.deepseek.com/v1"
17. llm = OpenAI(model='deepseek-chat',api_key=openai.api_key, base_url=openai.base_url)
20. from llama_index.core import Settings
23. # llm = OpenAI(api_key=openai.api_key, base_url=openai.base_url)
24. Settings.llm = OpenAI(model="deepseek-chat",api_key=openai.api_key, base_url=openai.base_url)
25. # %%
26. import os
28. jinaai_api_key = "jina"
29. os.environ["JINAAI_API_KEY"] = jinaai_api_key
31. from llama_index.embeddings.jinaai import JinaEmbedding
33. text_embed_model = JinaEmbedding(
34.     api_key=jinaai_api_key,
35.     model="jina-embeddings-v3",
36.     # choose `retrieval.passage` to get passage embeddings
37.     task="retrieval.passage",
38. )
40. # %%
41. # create client and a new collection
42. chroma_client = chromadb.EphemeralClient()
43. chroma_collection = chroma_client.create_collection("quickstart")
45. # %%
48. # define embedding function
49. embed_model = text_embed_model
51. # load documents
52. documents = SimpleDirectoryReader("./data/paul_graham/").load_data()
54. # save to disk
56. db = chromadb.PersistentClient(path="./chroma_db")
57. chroma_collection = db.get_or_create_collection("quickstart")
58. vector_store = ChromaVectorStore(chroma_collection=chroma_collection)
59. storage_context = StorageContext.from_defaults(vector_store=vector_store)
61. index = VectorStoreIndex.from_documents(
62.     documents, storage_context=storage_context, embed_model=embed_model
63. )
65. # load from disk
66. db2 = chromadb.PersistentClient(path="./chroma_db")
67. chroma_collection = db2.get_or_create_collection("quickstart")
68. vector_store = ChromaVectorStore(chroma_collection=chroma_collection)
69. index = VectorStoreIndex.from_vector_store(
70.     vector_store,
71.     embed_model=embed_model,
72. )
74. # Query Data from the persisted index
75. query_engine = index.as_query_engine()
76. response = query_engine.query("What did the author do growing up?")
77. print('response:',response)

复制代码

1.llamaindex如何配置deepseek

找到llama_index下面的openai的utils配置里，参加"deepseek-chat":128000,
路径C:\Users\USER.conda\envs\workspace\lib\site-packages\llama_index\llms\openai\utils.py

1. from llama_index.llms.openai import OpenAI
3. llm = OpenAI(model="deepseek-chat", base_url="https://api.deepseek.com/v1", api_key="sk-")
5. response = llm.complete("见到你很高兴")
6. print(str(response))

复制代码

2.llama使用jina

1. # Initilise with your api key
2. import os
4. jinaai_api_key = "jina_"
5. os.environ["JINAAI_API_KEY"] = jinaai_api_key
7. from llama_index.embeddings.jinaai import JinaEmbedding
9. text_embed_model = JinaEmbedding(
10.     api_key=jinaai_api_key,
11.     model="jina-embeddings-v3",
12.     # choose `retrieval.passage` to get passage embeddings
13.     task="retrieval.passage",
14. )
16. embeddings = text_embed_model.get_text_embedding("This is the text to embed")
17. print("Text dim:", len(embeddings))
18. print("Text embed:", embeddings[:5])
20. query_embed_model = JinaEmbedding(
21.     api_key=jinaai_api_key,
22.     model="jina-embeddings-v3",
23.     # choose `retrieval.query` to get query embeddings, or choose your desired task type
24.     task="retrieval.query",
25.     # `dimensions` allows users to control the embedding dimension with minimal performance loss. by default it is 1024.
26.     # A number between 256 and 1024 is recommended.
27.     dimensions=512,
28. )
30. embeddings = query_embed_model.get_query_embedding(
31.     "This is the query to embed"
32. )
33. print("Query dim:", len(embeddings))
34. print("Query embed:", embeddings[:5])

复制代码

3.llamaindex 使用chromadb

1. # %%
2. from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
3. from llama_index.vector_stores.chroma import ChromaVectorStore
4. from llama_index.core import StorageContext
5. from llama_index.embeddings.huggingface import HuggingFaceEmbedding
6. from IPython.display import Markdown, display
8. from llama_index.llms.openai import OpenAI
9. import chromadb
11. # %%
13. import openai
14. openai.api_key = "sk-"
16. openai.api_base = "https://api.deepseek.com/v1"
19. from llama_index.core import Settings
22. # llm = OpenAI(api_key=openai.api_key, base_url=openai.base_url)
23. Settings.llm = OpenAI(model="deepseek-chat",api_key=openai.api_key, base_url=openai.base_url)
26. # %%
27. import os
29. jinaai_api_key = "jina_"
30. os.environ["JINAAI_API_KEY"] = jinaai_api_key
32. from llama_index.embeddings.jinaai import JinaEmbedding
34. text_embed_model = JinaEmbedding(
35.     api_key=jinaai_api_key,
36.     model="jina-embeddings-v3",
37.     # choose `retrieval.passage` to get passage embeddings
38.     task="retrieval.passage",
39. )
41. # %%
42. # create client and a new collection
43. chroma_client = chromadb.EphemeralClient()
44. chroma_collection = chroma_client.create_collection("quickstart")
46. # %%
49. # define embedding function
50. embed_model = text_embed_model
52. # load documents
53. documents = SimpleDirectoryReader("./data/paul_graham/").load_data()
55. # %%
56. # set up ChromaVectorStore and load in data
57. vector_store = ChromaVectorStore(chroma_collection=chroma_collection)
59. # %%
61. storage_context = StorageContext.from_defaults(vector_store=vector_store)
63. # %%
64. index = VectorStoreIndex.from_documents(
65.     documents, storage_context=storage_context, embed_model=embed_model
66. )
69. # Settings.llm = llm
71. # Query Data
72. query_engine = index.as_query_engine()
73. response = query_engine.query("What did the author do growing up?")
74. print('response:',response)

复制代码

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。