本文介绍了利用 Python、Ollama、ChromaDB 和 Streamlit 构建本地检索增强生成（RAG）系统的完整流程。通过 OCR 处理 PDF 文档，结合多模态数据处理技术，在无需 GPU 的笔记本电脑上实现私有知识库问答。内容涵盖环境搭建、数据预处理、向量数据库存储、后端逻辑及前端交互界面开发，强调数据隐私安全与本地化部署优势。
本文详细介绍如何利用 Python 和大型语言模型（LLM）构建一个检索增强生成（Retrieval-Augmented Generation, RAG）系统。该系统包括完整的数据处理、向量数据库、前端以及后端结构，能够从个人文档中学习知识，并在无需 GPU 的情况下于笔记本电脑上运行。
自然语言处理（NLP）是人工智能领域研究机器与人类语言交互的学科。NLP 的巅峰目前当属大型语言模型（LLM）的出现。LLM 在海量文本数据上进行训练，能够学习语言模式和变体。'语言模型'一词随着深度学习和神经网络的兴起而普及。特别是 2018 年，谷歌引入了基于 Transformer 架构的模型，极大地提升了 NLP 模型的性能（例如谷歌的 BERT 和 OpenAI 的 GPT）。
如今，LLM 通常用于文本到文本任务或'文本生成'（如翻译、摘要、聊天机器人和虚拟助手，甚至撰写整本书）。这催生了生成式人工智能（GenAI）的兴起，GenAI 专注于生成新内容（如文本、图像、音频、视频）。
目前最先进的 LLM 包括：
ChatGPT 是最常用的 LLM，但企业面临一个问题：无法将敏感数据上传到 OpenAI（主要出于隐私和安全考虑）。因此，企业正在创建内部 AI 服务，以便在其私有知识库上利用 LLM 的强大功能。这种任务被称为检索增强生成（RAG）。RAG 是一种结合检索和生成模型的技术，通过从外部来源获取知识事实来增强 LLM。
在组织中，知识库通常由包含多模态内容（如文本、图像、电子表格）的文档组成。因此，最大的挑战是如何以机器能够理解的方式处理所有这些内容。简而言之，首先将每个文档转换为嵌入向量，然后将用户查询转换为相同的向量空间，从而实现余弦相似度搜索。
本文将介绍如何使用 LLM 和多模态数据构建一个 RAG 应用程序，该应用程序可以在普通笔记本电脑上运行，无需 GPU。文中将提供一些有用的 Python 代码，这些代码可以轻松应用于其他类似案例，并对每一行代码进行注释，以便读者可以复现该示例。
在企业环境中，PDF 是使用最广泛的文档格式，因为大多数文档在共享前都会转换为 PDF 格式。此外，PDF 也是一个很好的案例，因为它们包含图像、表格和文本。因此，本文将以一份 PDF 格式的上市公司财务报表作为数据集。
处理 PDF 文档有两种方法：将其作为文本读取或将其解析为图像。这两种方法都不完美，具体取决于用例，但 OCR（光学字符识别）往往效果更好，因此本文采用这种方法。
首先，需要将文档转换为图像：
import pdf2image # 1.17.0
doc_img = pdf2image.convert_from_path("data/doc_nvidia.pdf", dpi=300)
然后，使用 Tesseract 识别图像中的文本。Tesseract 是 HP 在 1985 年开发的主要 OCR 系统，目前由谷歌维护。
import pytesseract # 0.3.10
doc_txt = []
for page in doc_img:
text = pytesseract.image_to_string(page)
doc_txt.append(text)
现在，我们有了原始文本，在进行任何处理之前，必须对其进行预处理。我们将使用语言模型。目前，有几种库可以利用 LLM 的强大功能：Ollama、HuggingFace、DSpy、LangChain。我喜欢使用 Ollama，因为它功能强大且用户友好。
首先，需要从网站下载 Ollama 程序。然后，在提示符下运行以下命令以在笔记本电脑上运行 LLM（本文使用 Phi3 和 LLaVa）：
最后，安装 Python 库（pip install ollama）即可。建议创建一个虚拟环境以管理依赖：
python -m venv rag_env
source rag_env/bin/activate # Windows: rag_env\Scripts\activate
pip install -r requirements.txt
在任何数据科学项目中，最困难的部分始终是数据集。无论投入多少机器学习，如果数据不好，AI 就无法正常工作。
对于 LLM，由于涉及嵌入，向量空间中的所有内容都应转换为文本。我们将展示如何处理最常见的多模态内容（文本、表格、图像、图表）。
文本 - 由于文档可能具有不同的结构和布局，因此应将数据集拆分为段落。除了原始文本外，还可以使用额外的信息（如标题、标签、链接）来丰富数据。这被称为元数据增强。
# 使用目录为段落添加标签
title_map = {
"4-12" : "Business",
"13-33" : "Risk Factors",
"34-44" : "Financials",
"45-46" : "Directors",
"47-83" : "Data"
}
lst_docs, lst_ids, lst_metadata = [], [], []
for n, page in enumerate(doc_txt):
try:
## 获取标题
title = [v for k,v in title_map.items() if n in range(int(k.split("-")[0]), int(k.split("-")[1])+1)][0]
## 清理页面
page = page.replace("Table of Contents","")
## 获取段落
for i,p in enumerate(page.split('\n\n')):
if len(p.strip())>5:
lst_docs.append(p.strip())
lst_ids.append(str(n)+"_"+str(i))
lst_metadata.append({"title":title})
except:
continue
元数据增强可以显著提高文档检索效率。例如，可以使用 Phi3 将每个段落总结为几个关键词。
import ollama # 0.2.1
def keyword_generator(p, top=3):
prompt = "summarize the following paragraph in 3 keywords separated by , : "
res = ollama.generate(model="phi3", prompt=prompt)["response"]
return res.replace("\n", " ").strip()
对所有段落应用该函数：
from tqdm.notebook import tqdm
for i, doc in tqdm(enumerate(lst_docs)):
lst_metadata[i]["keywords"] = keyword_generator(doc)
表格 - LLM 可以处理顺序文本，但也可以理解结构化数据，因此可以将表格作为普通文本传递。
table = lst_docs[376]
print("Table: \n", table)
prompt = f"Summarize the following table: {table}"
res = ollama.generate(model="phi3", prompt=prompt)["response"]
print("\nSummary : \n", res)
图像 - 图像嵌入比较棘手（涉及将 3D 数组的维度降低到向量）。因此，一种常见的做法是使用 AI 生成描述图像的标题，并将其作为文本处理。
from matplotlib import image, pyplot
image_file = "data/image.jpeg"
pyplot.imshow(image.imread(image_file))
pyplot.show()
为了使 LLM 能够处理图像，必须将其转换为字符串：
import base64
def encode_image(path):
with open(path, "rb") as file:
return base64.b64encode(file.read()).decode('utf-8')
img = encode_image(image_file)
LLM 无法直接处理图像，因此需要使用视觉模型来描述图像。微软的 LLaVa 是一个不错的选择，因为它也可以在无 GPU 的情况下运行。
prompt = "describe the image"
res = ollama.generate(model="llava", prompt=prompt, images=[img])["response"]
print(res)
图表 - 同样，视觉模型可以读取和总结图表，但结果不如图像准确。
image_file = "data/plot.png"
pyplot.imshow(image.imread(image_file))
pyplot.show()
img = encode_image(image_file)
prompt = "Describe the image in detail. Be specific about graphs, such as bar plots, line graphs, etc."
res = ollama.generate(model="llava", prompt=prompt, images=[img])["response"]
print(res)
向量数据库是存储、索引和搜索非结构化数据的最佳解决方案。当前，最常用的向量数据库是微软的 AISearch，而最好的开源工具是 ChromaDB，它功能强大、易于使用，并且是免费的。
安装 ChromaDB 后，可以使用 Python 以三种不同的方式与数据库交互：
import chromadb # 0.5.0
# 创建一个持久化的数据库实例
db = chromadb.PersistentClient()
# 获取或创建一个名为 "nvidia" 的集合
collection_name = "nvidia"
collection = db.get_or_create_collection(
name=collection_name,
embedding_function=chromadb.utils.embedding_functions.DefaultEmbed
)
# 将文档、ID 和元数据添加到集合中
collection.add(
documents=lst_docs,
ids=lst_ids,
metadatas=lst_metadata,
images=None,
embeddings=None
)
在 ChromaDB 中存储文档时，数据以向量形式保存，因此可以使用查询向量检索最匹配的记录。需要注意的是，除非另有说明，否则默认的嵌入函数是一个句子转换模型（Sentence Transformer model），例如 all-MiniLM-L6-v2。
# 查看集合中的一个样本
collection.peek(1)
输出示例：
{
'embeddings': [
[-0.06092095375061035, -0.01741098240017891, 0.0484163761138916, ...]
],
'metadatas': [
{
'keywords': 'Renewable Energy Adoption, Supplier Engagement, Emission Reduction Goals',
'title': 'Business'
}
],
'documents': [
'We aim to generate enough renewable energy to match 100% of our global electricity usage for our offices and data centers. In fiscal year 2023, we increased the percentage of our total electricity use matched by renewable energy purchases to 44%. By fiscal year 2026, we aim to engage manufacturing suppliers comprising at least 67% of NVIDIA's scope 3 category 1 GHG emissions with the goal of effecting supplier adoption of science-based targets.'
],
'uris': None,
'data': None
}
文档作为包含向量和元数据的字典保存。
接下来，尝试查询一些信息：
query = "how much is the revenue?"
res_db = collection.query(query_texts=[query])["documents"][0][0:10]
context = ' '.join(res_db).replace("\n", " ")
print(context)
输出示例：
Total revenue for fiscal year 2024 was $60.9 billion, up 126% from a year ago. Data Center revenue for fiscal year 2024 was up 217%. Strong demand was driven by enterprise software and consumer internet applications, and multiple industry verticals including automotive, financial services, and healthcare. Gaming revenue for fiscal year 2024 was up 15%. Professional Visualization revenue for fiscal year 2024 was up 1%. Automotive revenue for the fiscal year 2024 was up 21%. The increase primarily reflected growth in self-driving platforms. Gross margin increased in fiscal year 2024, primarily driven by Data Center revenue growth and lower net inventory provisions as a percentage of revenue. Operating expenses increased for fiscal year 2024, driven by growth in employees and compensation increases.
数据库的响应可以作为上下文，即 LLM 可以利用其记忆模块存储和访问的一小部分信息。使用提示，我们可以指定模型是仅使用上下文中的信息，还是也使用其在训练过程中学习到的知识。
仅使用上下文进行聊天：
res = ollama.chat(
model="phi3",
messages=[
{"role": "system", "content": "Give the most accurate answer using only the following information:"},
{"role": "user", "content": query}
],
stream=True
)
print(res["message"]["content"])
输出示例：
The total recognized revenue for fiscal year 2024 was $60.9 billion, which represents an increase of 126% from the previous year. The breakdown by category in millions of dollars is as follows:
- Data Center: $47,525 million
- Gaming: $10,447 million
- Professional Visualization: $1,553 million
- Automotive: $1,091 million
- OEM and Other: $306 million
结合知识和上下文进行聊天：
res = ollama.chat(
model="phi3",
messages=[
{"role": "system", "content": "Give the most accurate answer using your knowledge and the following information:"},
{"role": "user", "content": query}
],
stream=True
)
print(res["message"]["content"])
输出示例：
The total recognized revenue for fiscal year 2024 was $60,922 million (or $60.9 billion when expressed in billions). This amount represents an increase of 126% from the previous year's revenue. Additionally, there is a breakdown by product categories as follows:
- Data Center: $47,525 million
- Gaming: $10,447 million
- Professional Visualization: $1,553 million
- Automotive: $1,091 million
- OEM and Other: $306 million
It's also important to note that there are deferred revenue amounts of $233 million in fiscal 2024 and $35 million in fiscal 2023 related to customer advances, which will be recognized as revenue over future periods. The remaining performance obligations account for approximately $1.1 billion, with an expectation that about 40% of this amount will be recognized within the next twelve months.
数据库设置完成并运行良好。接下来，我们可以构建聊天机器人应用程序以执行 RAG，或者更简单地说，回答有关我们文档的问题。
如今，Streamlit 是构建快速 Web 应用程序最常用的 Python 库，因为它通过其流式功能简化了 NLP 应用程序的开发。
首先，定义布局：我的屏幕应有一个侧边栏，用户可以在其中查看聊天历史记录。
import streamlit as st # 1.35.0
## 布局
st.title("Write your questions")
st.sidebar.title("Chat History")
app = st.session_state
if 'messages' not in app:
app['messages'] = [{"role": "assistant", "content": "I'm ready to retrieve information"}]
if 'history' not in app:
app['history'] = []
if 'full_response' not in app:
app['full_response'] = ''
应用程序对象（或会话状态）是一个具有以下结构的字典：
{
'history': [
': how much is the revenue?',
': The total revenue reported in the given information is 60 million'
],
'messages': [
{'role': 'assistant', 'content': "I'm ready to retrieve information"},
{'role': 'user', 'content': 'how much is the revenue?'},
{'role': 'assistant', 'content': 'The total revenue reported in the given information is 60 million'}
],
'full_response': 'The total revenue reported in the given information is 60 million'
}
如果不想聊天中的消息消失，请添加以下代码：
## 保持消息在聊天中
for msg in app["messages"]:
if msg["role"] == "user":
st.chat_message(msg["role"], avatar="🧑").write(msg["content"])
elif msg["role"] == "assistant":
st.chat_message(msg["role"], avatar="🤖").write(msg["content"])
接下来，必须编写聊天核心：用户的问题将被添加到消息列表中，传递给 AI，应用程序将流式传输响应。
## 聊天
if txt := st.chat_input():
### 用户写入
app["messages"].append({"role": "user", "content": txt})
st.chat_message("user", avatar="🧑").write(txt)
### AI 使用聊天流式响应
app["full_response"] = ""
with st.chat_message("assistant", avatar="🤖"):
for chunk in ai.respond(app["messages"], use_knowledge=True):
app["full_response"] += chunk
st.write(chunk)
### 显示历史记录
app['history'].append(": " + txt)
app['history'].append(": " + app["full_response"])
st.sidebar.markdown("<br />".join(app['history']) + "<br /><br />", unsafe_allow_html=True)
在用户输入消息后，AI 将调用 respond 函数，该函数将查询数据库并使用 LLM 生成响应。
为了增强用户体验，建议在 Streamlit 中添加错误处理机制，例如当数据库连接失败时显示友好的提示信息，或者限制用户输入长度以防止 Token 溢出。
后端部分我们已经熟悉，只需将其放入一个类中即可。以下是需要强调的几个要点：
stream 参数应设置为 True，以返回文本块而不是完整响应。import chromadb # 0.5.0
import ollama # 0.5.0
class AI:
def __init__(self):
db = chromadb.PersistentClient()
self.collection = db.get_or_create_collection("nvidia")
def query(self, q, top=10):
res_db = self.collection.query(query_texts=[q])["documents"][0][0:top]
context = ' '.join(res_db).replace("\n", " ")
return context
def respond(self, lst_messages, use_knowledge=False):
q = lst_messages[-1]["content"]
context = self.query(q)
if use_knowledge:
prompt = "Give the most accurate answer using your knowledge and the following information:"
else:
prompt = "Give the most accurate answer using only the following information:"
res_ai = ollama.chat(
model="phi3",
messages=[
{"role": "system", "content": prompt},
*lst_messages
],
stream=True
)
for res in res_ai:
chunk = res["message"]["content"]
app["full_response"] += chunk
yield chunk
ai = AI()
要运行 Streamlit 应用程序，必须使用以下命令：
streamlit run rag_app.py
在生产环境中，建议使用 Docker 容器化部署，以确保环境一致性，并通过反向代理（如 Nginx）处理流量。
以下是该应用程序的运行示例：
用户输入: 'how much is the revenue?'
AI 响应:
The total reported revenue for fiscal year 2024 was 60,922 million. This represents an increase of 126% from a year ago. Additionally, there is a breakdown by product categories as follows:
- Data Center: 47,525 million
- Gaming: 10,447 million
- Professional Visualization: 1,553 million
- Automotive: 1,091 million
- OEM and Other: 306 million
用户输入: 'how did they make it?'
AI 响应:
NVIDIA achieved the reported revenue through a combination of factors, including:
- **Data Center Revenue Growth**: The Data Center segment saw a significant increase of 217%, driven by strong demand from enterprise software and consumer internet applications, as well as various industry verticals such as automotive, financial services, and healthcare. Customers access NVIDIA AI infrastructure both through the cloud and on-premises, with Data Center compute revenue growing by 244% and Networking revenue by 133%.
- **Gaming Revenue Increase**: Gaming revenue rose by 15%, reflecting higher sales to partners following the normalization of channel inventory levels and growing demand.
- **Professional Visualization and Automotive Revenue**: These segments also saw growth, with Professional Visualization revenue increasing by 1% and Automotive revenue by 21%, primarily due to growth in self-driving platforms.
- **Gross Margin Improvement**: Gross margin increased due to the growth in Data Center revenue and lower net inventory provisions as a percentage of revenue.
- **Operating Expenses**: Operating expenses increased due to the growth in employees and compensation increases.
用户输入: 'What is the significance of their Data Center revenue growth?'
AI 响应:
The significant increase in Data Center revenue, up 217% from the previous year, highlights NVIDIA's growing dominance in the AI and data center markets. This growth is driven by several factors:
- **Enterprise Demand**: Strong demand from enterprise customers for AI infrastructure to support applications such as machine learning, data analytics, and cloud computing.
- **Diverse Industry Verticals**: The adoption of NVIDIA's AI solutions across various industries, including automotive, financial services, and healthcare, indicates a broad-based demand for AI capabilities.
- **Cloud and On-Premises**: The ability to serve customers through both cloud-based and on-premises solutions provides flexibility and meets different customer needs.
- **Compute and Networking**: The substantial growth in both Data Center compute revenue (up 244%) and Networking revenue (up 133%) underscores the importance of NVIDIA's comprehensive AI platform, which includes both hardware and software components.
用户输入: 'What challenges did they face in the Automotive segment?'
AI 响应:
While the Automotive revenue increased by 21%, the document does not explicitly list specific challenges faced in this segment. However, potential challenges in the automotive industry for NVIDIA could include:
- **Market Competition**: Intense competition from other technology companies and traditional automotive suppliers in the autonomous driving and AI solutions space.
- **Regulatory Hurdles**: Navigating the complex regulatory landscape for autonomous vehicles and AI technologies.
- **Technological Development**: The need to continuously innovate and improve AI algorithms and hardware to meet the demanding requirements of the automotive industry.
- **Supply Chain Issues**: Potential disruptions in the supply chain for critical components needed for NVIDIA's automotive solutions.
用户输入: 'What is NVIDIA's strategy for future growth?'
AI 响应:
NVIDIA's strategy for future growth is likely to focus on several key areas:
- **Continued Expansion in Data Center**: Leveraging the strong growth in the Data Center segment by expanding its AI infrastructure offerings and targeting new industries and applications.
- **Gaming Market**: Maintaining and growing its presence in the gaming market by continuing to innovate and meet the demands of gamers and developers.
- **Professional Visualization**: Further developing its professional visualization solutions to support industries such as architecture, engineering, and design.
- **Automotive Innovation**: Investing in research and development to advance its autonomous driving and AI technologies for the automotive industry.
- **Sustainability**: Focusing on sustainability initiatives, such as increasing the use of renewable energy and setting science-based targets for emissions reductions.
- **Strategic Partnerships**: Forming strategic partnerships with other companies to expand its reach and capabilities in various markets.
本文展示了如何使用 Python 和大型语言模型构建一个检索增强生成（RAG）系统。通过以下步骤，我们成功构建了一个能够从多模态文档中学习知识并回答问题的 AI 应用程序：
该系统展示了如何利用 LLM 和多模态数据构建一个强大的 AI 应用程序，即使在普通笔记本电脑上也能高效运行。该方法具有以下优势：
在实际应用中，建议定期更新向量数据库以反映最新文档，并监控 LLM 的输出质量以防幻觉。通过合理配置硬件资源，此类本地 RAG 系统可广泛应用于企业内部知识库、个人笔记管理及智能客服等场景。
微信公众号「极客日志」，在微信中扫描左侧二维码关注。展示文案：极客日志 zeeklog
使用加密算法（如AES、TripleDES、Rabbit或RC4）加密和解密文本明文。 在线工具，加密/解密文本在线工具，online
生成新的随机RSA私钥和公钥pem证书。 在线工具，RSA密钥对生成器在线工具，online
基于 Mermaid.js 实时预览流程图、时序图等图表，支持源码编辑与即时渲染。 在线工具，Mermaid 预览与可视化编辑在线工具，online
解析常见 curl 参数并生成 fetch、axios、PHP curl 或 Python requests 示例代码。 在线工具，curl 转代码在线工具，online
将字符串编码和解码为其 Base64 格式表示形式即可。 在线工具，Base64 字符串编码/解码在线工具，online
将字符串、文件或图像转换为其 Base64 表示形式。 在线工具，Base64 文件转换器在线工具，online