在我之前的文章 "Elasticsearch：与多个 PDF 聊天 | LangChain Python 应用教程（免费 LLMs 和嵌入）" 里，我详述如何使用 Streamlit，Langchain, Elasticsearch 及 OpenAI 来针对 PDF 进行聊天。在今天的文章中，我将使用 Chainlit 来展示如使用 Langchain 及 Elasticsearch 针对 PDF 文件进行查询。
为方便大家学习，我的代码在地址 GitHub - liu-xiao-guo/langchain-openai-chainlit: Chat with your documents (pdf, csv, text) using Openai model, LangChain and Chainlit 进行下载。
安装
安装 Elasticsearch 及 Kibana
如果你还没有安装好自己的 Elasticsearch 及 Kibana，那么请参考一下的文章来进行安装：
在安装的时候，请选择 Elastic Stack 8.x 进行安装。在安装的时候，我们可以看到如下的安装信息：
拷贝 Elasticsearch 证书
我们把 Elasticsearch 的证书拷贝到当前的目录下：
bash
1. $ pwd
2. /Users/liuxg/python/elser
3. $ cp ~/elastic/elasticsearch-8.12.0/config/certs/http_ca.crt .
4. $ ls http_ca.crt
5. http_ca.crt安装 Python 依赖包
我们在当前的目录下打入如下的命令：
bash
1. python3 -m venv .venv
2. source .venv/bin/activate然后，我们再打入如下的命令：
bash
1. $ pwd
2. /Users/liuxg/python/langchain-openai-chainlit
3. $ source .venv/bin/activate
4. (.venv) $ pip3 install -r requirements.txt运行应用
有关 Chainlit 的更多知识请参考 Overview - Chainlit。这里就不再赘述。有关 pdf_qa.py 的代码如下：
pdf_qa.py
ini
1. # Import necessary modules and define env variables
3. # from langchain.embeddings.openai import OpenAIEmbeddings
4. from langchain_openai import OpenAIEmbeddings
5. from langchain.text_splitter import RecursiveCharacterTextSplitter
6. from langchain.chains import RetrievalQAWithSourcesChain
7. from langchain_openai import ChatOpenAI
8. from langchain.prompts.chat import (
9. ChatPromptTemplate,
10. SystemMessagePromptTemplate,
11. HumanMessagePromptTemplate,
12. )
13. import os
14. import io
15. import chainlit as cl
16. import PyPDF2
17. from io import BytesIO
19. from pprint import pprint
20. import inspect
21. # from langchain.vectorstores import ElasticsearchStore
22. from langchain_community.vectorstores import ElasticsearchStore
23. from elasticsearch import Elasticsearch
25. from dotenv import load_dotenv
27. # Load environment variables from .env file
28. load_dotenv()
30. OPENAI_API_KEY= os.getenv("OPENAI_API_KEY")
31. ES_USER = os.getenv("ES_USER")
32. ES_PASSWORD = os.getenv("ES_PASSWORD")
33. elastic_index_name='pdf_docs'
35. # text_splitter and system template
37. text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
39. system_template = """Use the following pieces of context to answer the users question.
40. If you don't know the answer, just say that you don't know, don't try to make up an answer.
41. ALWAYS return a "SOURCES" part in your answer.
42. The "SOURCES" part should be a reference to the source of the document from which you got your answer.
44. Example of your response should be:
46. ```
47. The answer is foo
48. SOURCES: xyz
49. ```
51. Begin!
52. ----------------
53. {summaries}"""
56. messages = [
57. SystemMessagePromptTemplate.from_template(system_template),
58. HumanMessagePromptTemplate.from_template("{question}"),
59. ]
60. prompt = ChatPromptTemplate.from_messages(messages)
61. chain_type_kwargs = {"prompt": prompt}
64. @cl.on_chat_start
65. async def on_chat_start():
67. # Sending an image with the local file path
68. elements = [
69. cl.Image()
70. ]
71. await cl.Message(content="Hello there, Welcome to AskAnyQuery related to Data!", elements=elements).send()
72. files = None
74. # Wait for the user to upload a PDF file
75. while files is None:
76. files = await cl.AskFileMessage(
77. content="Please upload a PDF file to begin!",
78. accept=["application/pdf"],
79. max_size_mb=20,
80. timeout=180,
81. ).send()
83. file = files[0]
85. # print("type: ", type(file))
86. # print("file: ", file)
87. # pprint(vars(file))
88. # print(file.content)
90. msg = cl.Message(content=f"Processing `{file.name}`...")
91. await msg.send()
93. # Read the PDF file
94. # pdf_stream = BytesIO(file.content)
95. with open(file.path, 'rb') as f:
96. pdf_content = f.read()
97. pdf_stream = BytesIO(pdf_content)
98. pdf = PyPDF2.PdfReader(pdf_stream)
99. pdf_text = ""
100. for page in pdf.pages:
101. pdf_text += page.extract_text()
103. # Split the text into chunks
104. texts = text_splitter.split_text(pdf_text)
106. # Create metadata for each chunk
107. metadatas = [{"source": f"{i}-pl"} for i in range(len(texts))]
109. # Create a Chroma vector store
110. embeddings = OpenAIEmbeddings()
112. url = f"https://{ES_USER}:{ES_PASSWORD}@localhost:9200"
114. connection = Elasticsearch(
115. hosts=[url],
116. ca_certs = "./http_ca.crt",
117. verify_certs = True
118. )
120. docsearch = None
122. if not connection.indices.exists(index=elastic_index_name):
123. print("The index does not exist, going to generate embeddings")
124. docsearch = await cl.make_async(ElasticsearchStore.from_texts)(
125. texts,
126. embedding = embeddings,
127. es_url = url,
128. es_connection = connection,
129. index_name = elastic_index_name,
130. es_user = ES_USER,
131. es_password = ES_PASSWORD,
132. metadatas=metadatas
133. )
134. else:
135. print("The index already existed")
137. docsearch = ElasticsearchStore(
138. es_connection=connection,
139. embedding=embeddings,
140. es_url = url,
141. index_name = elastic_index_name,
142. es_user = ES_USER,
143. es_password = ES_PASSWORD
144. )
146. # Create a chain that uses the Chroma vector store
147. chain = RetrievalQAWithSourcesChain.from_chain_type(
148. ChatOpenAI(temperature=0),
149. chain_type="stuff",
150. retriever=docsearch.as_retriever(search_kwargs={"k": 4}),
151. )
153. # Save the metadata and texts in the user session
154. cl.user_session.set("metadatas", metadatas)
155. cl.user_session.set("texts", texts)
157. # Let the user know that the system is ready
158. msg.content = f"Processing `{file.name}` done. You can now ask questions!"
159. await msg.update()
161. cl.user_session.set("chain", chain)
164. @cl.on_message
165. async def main(message:str):
167. chain = cl.user_session.get("chain") # type: RetrievalQAWithSourcesChain
168. print("chain type: ", type(chain))
169. cb = cl.AsyncLangchainCallbackHandler(
170. stream_final_answer=True, answer_prefix_tokens=["FINAL", "ANSWER"]
171. )
172. cb.answer_reached = True
174. print("message: ", message)
175. pprint(vars(message))
176. print(message.content)
177. res = await chain.acall(message.content, callbacks=[cb])
179. answer = res["answer"]
180. sources = res["sources"].strip()
181. source_elements = []
183. # Get the metadata and texts from the user session
184. metadatas = cl.user_session.get("metadatas")
185. all_sources = [m["source"] for m in metadatas]
186. texts = cl.user_session.get("texts")
188. print("texts: ", texts)
190. if sources:
191. found_sources = []
193. # Add the sources to the message
194. for source in sources.split(","):
195. source_name = source.strip().replace(".", "")
196. # Get the index of the source
197. try:
198. index = all_sources.index(source_name)
199. except ValueError:
200. continue
201. text = texts[index]
202. found_sources.append(source_name)
203. # Create the text element referenced in the message
204. source_elements.append(cl.Text(content=text, name=source_name))
206. if found_sources:
207. answer += f"\nSources: {', '.join(found_sources)}"
208. else:
209. answer += "\nNo sources found"
211. if cb.has_streamed_final_answer:
212. cb.final_stream.elements = source_elements
213. await cb.final_stream.update()
214. else:
215. await cl.Message(content=answer, elements=source_elements).send()我们可以使用如下的命令来运行：
ini
1. export ES_USER="elastic"
2. export ES_PASSWORD="xnLj56lTrH98Lf_6n76y"
3. export OPENAI_API_KEY="YourOpenAiKey"
5. chainlit run pdf_qa.py -w
arduino
1. (.venv) $ chainlit run pdf_qa.py -w
2. 2024-02-14 10:58:30 - Loaded .env file
3. 2024-02-14 10:58:33 - Your app is available at http://localhost:8000
4. 2024-02-14 10:58:34 - Translation file for en not found. Using default translation en-US.
5. 2024-02-14 10:58:35 - 2 changes detected我们先选择项目自带的 pdf 文件：
vbnet
Is sample PDF download critical to an organization?
Does comprehensive PDF testing have various advantages?