Try   HackMD

AI / ML領域相關學習筆記入口頁面

Deeplearning.ai GenAI/LLM系列課程筆記

GenAI
RAG

Preprocessing Unstructured Data for LLM Applications
大型語言模型(LLM)應用的非結構化資料前處理

Introduction

課程重點

  • 如何為 LLM 應用開發預處理數據、處理不同的文件類型
  • 如何將各種文件提取和規範化為通用的 JSON 格式，並用metadata豐富其內容，以改善搜索結果
  • 文件圖像分析技術，包括排版(layout)檢測和視覺轉換器(vision transformers)，以提取和理解 PDF、圖片和表格
  • 如何構建一個能夠讀取 PDF、PowerPoint 和 Markdown 文件等不同文件的RAG bot

Overview of LLM Data Preprocessing

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

2023.08。unstructured.io/。How to Build an End-to-End RAG Pipeline with Unstructured’s AP

Data Preprocessing and LLMs

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

  • Retrieval Augmented Generation (RAG): A technique for grounding LLM responses on validated external information.

  • Contextual Integration

    • RAG apps load context into a database, then retrieve content to insert into a prompt.

Preprocessing Outputs

  • Document Content
    Text content from the documents
    Used for keyword or similarity search in RAG apps

    • Document Elements
      The basic building blocks of a document.
      Useful for various RAG tasks, such as filtering and chunking.

      • Title
      • Narrative Text
      • List Item
      • Table
      • Image
    • Element Metadata
      Additional information about an element.
      Image Not Showing Possible Reasons
      • The image file may be corrupted
      • The server hosting the image is unavailable
      • The image path is incorrect
      • The image format is not supported
      Learn More →
      Useful for filtering in hybrid search and for identifying the source of a response.
      • Filename
      • Filetype
      • Page Number
      • Section
      • keywords
      • summary

資料前處裡的難點

  • 內容提示(Content Cues)

    • 在處理不同類型的文件時，需要根據文件類型（如圖片或Markdown文檔）來識別不同的資訊或結構標記

  • 標準化需求(Standardization Need)

    • 不同類型的文件（如Word文檔、PDF等）在處理前需轉化為一種通用格式，以便統一處理和分析(標準化)

  • 提取變異性(Extraction Variability)

    • 根據文件的格式和結構，數據提取的方法會有所不同，如從表格中提取數據與從學術文章中提取內容可能採用不同技術

  • 後設資料洞察(Metadata Insight)

    • 提取後設資料（如作者、發布日期等資訊）常需分析文件的特定結構，以確定相應的後設資料位於哪裡

標準化文件內容。Normalizing the Content

Normalizing Diverse Documents

  • 格式多樣性(Format Diversity)

    • 文件可能以多種不同的電子格式存在，每種格式都有其特定的文件結構和呈現方式（PDF、Word、EPUB、Markdown 等）

  • 通用格式(Common Format)

    • 在處理不同格式的文件前，為了便於統一處理和分析，需要將它們轉換成一個標準化的格式
    • 這種通用格式能夠識別並提取文件中的共通元素，例如標題和正文，這樣不同來源的文件在接下來的處理階段可以被一視同仁

  • 標準化好處(Normalization Benefit)
    標準化格式讓任何文件都能以相同的方式處理，無論其來源格式如何

    • 過濾不需要的元素，如頁眉(headers)和頁腳
    • 將文件元素切塊到不同的章節

  • 降低處理成本(Reduced Processing Cost)
    文件重新處理的初步步驟是整個過程中最昂貴的部分

    • 例如切塊這樣的下游任務，在標準化的輸出上是成本較低的操作
    • 使實驗不同的切塊技術成為可能，無需重新處理文件

資料序列化Data Serialization

資料序列化 (Data Serialization)

資料序列化是指在電腦科學中將資料結構或物件轉換為位元組序列的過程，以便可以在檔案、記憶體或網絡中存儲或傳輸。序列化後的資料可以在需要時進行反序列化，即將位元組序列還原回原始的資料結構或物件。

主要功能和好處：

  1. 持久性(Persistence): 序列化允許數據持久保存在硬碟或其他存儲媒介上，即使程式結束後仍能再次讀取數據。
  2. 通信(Communication): 在不同的計算機系統或網絡中，序列化允許數據在節點間進行交換和傳輸，支持分散式計算應用。
  3. 資料交換(Data Exchange): 序列化使得不同應用程式或不同平台間可以交換數據，即使這些程式是用不同的程式語言編寫的。

實現方式：

  • 二進制格式(Binary Format): 將資料轉換為二進制代碼，這種格式通常更緊湊、效率更高，但可讀性差。
  • 文字格式(Textual Format): 如XML、JSON等，這些格式可讀性好，便於人類理解和編輯，但通常比二進制格式佔用更多空間。

應用實例：

  • Web API通信：網絡應用程式常用JSON或XML格式序列化數據，以便在網際網路上交換資料。
  • 應用程式設定：許多應用程式將配置選項序列化到配置文件（如.ini或.cfg文件），以便在應用啟動時載入。

  • 序列化好處(Serialization Benefits):

    • 允許文件前處理的結果能夠在未來被再次使用

  • JSON的優勢

    • 結構普遍且易於理解，是標準的HTTP回應格式
    • 能夠在多種程式語言中使用
      • JSON作為一種輕量級的數據交換格式，被廣泛支持於各種程式語言，如Python、Java、JavaScript等
    • 可以轉換為JSONL格式，用於串流的應用場景
      • JSONL（JSON Line）格式是一種將多個JSON對象分行存儲的方式，適合於串流數據處理和實時數據載入

    Image Not Showing Possible Reasons
    • The image was uploaded to a note which you don't have access to
    • The note which the image was originally uploaded to has been deleted
    Learn More →

HTML 頁面提取

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

  • LLM相關性(LLM Relevance)
    為了使大型語言模型能夠反映當前的語言使用和知識，重要的是定期更新其訓練數據。包括最新的網絡內容可以幫助模型更好地理解當前流行的話題和語言變化

  • HTML理解(HTML Understanding)
    HTML 是網頁的基礎，其結構化元素幫助定義了網頁上資訊的呈現方式，使用元素標籤如 <h1> 作為標題和 <p> 作為文本。

  • 數據提取和分類(Data Extraction and Categorization)
    通過分析HTML元素，可以從網頁中提取有用的資訊並將其組織成結構化的數據。這包括識別標題、段落、列表等，並將這些內容分類以便於後續的數據分析或應用
    這種結構化的提取是資訊檢索、內容管理和數據分析等領域的關鍵技術

    Image Not Showing Possible Reasons
    • The image file may be corrupted
    • The server hosting the image is unavailable
    • The image path is incorrect
    • The image format is not supported
    Learn More →
    這不就是網頁爬蟲在做的事? 可以用LLM進行網頁解析?

code - lesson/3/normalizing-the-content

  • 環境設定






















from IPython.display import JSON

import json

from unstructured_client import UnstructuredClient
from unstructured_client.models import shared
from unstructured_client.models.errors import SDKError

from unstructured.partition.html import partition_html
from unstructured.partition.pptx import partition_pptx
from unstructured.staging.base import dict_to_elements, elements_to_json

from Utils import Utils
utils = Utils()

DLAI_API_KEY = utils.get_dlai_api_key()
DLAI_API_URL = utils.get_dlai_url()

s = UnstructuredClient(
    api_key_auth=DLAI_API_KEY,
    server_url=DLAI_API_URL,
)



filename = "example_files/medium_blog.html"
elements = partition_html(filename=filename)

image









from unstructured.partition.html import partition_html
filename = "example_files/medium_blog.html"
elements = partition_html(filename=filename)

element_dict = [el.to_dict() for el in elements]
example_output = json.dumps(element_dict[11:15], indent=2)
print(example_output)


​  {
​   "type": "Title",
​   "element_id": "29887a5ff9846ccc23327565a07e17fa",
​   "text": "Share",
​   "metadata": {
​     "category_depth": 0,
​     "last_modified": "2024-03-30T04:25:39",
​     "page_number": 1,
​     "languages": [
​       "eng"
​     ],
​     "file_directory": "example_files",
​     "filename": "medium_blog.html",
​     "filetype": "text/html"
​   }
​   ...

JSON(example_output)

image

Powerpoint擷取

image

image







from unstructured.partition.pptx import partition_pptx
filename = "example_files/msft_openai.pptx"
elements = partition_pptx(filename=filename)

element_dict = [el.to_dict() for el in elements]
JSON(json.dumps(element_dict[:], indent=2))

image

PDF

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →
model-based。需要api key

image

Image(filename="images/cot_paper.png", height=600, width=600)

image




















filename = "example_files/CoT.pdf"
with open(filename, "rb") as f:
    files=shared.Files(
        content=f.read(), 
        file_name=filename,
    )

req = shared.PartitionParameters(
    files=files,
    strategy='hi_res',
    pdf_infer_table_structure=True,
    languages=["eng"],
)
try:
    resp = s.general.partition(req)
    print(json.dumps(resp.elements[:3], indent=2))
except SDKError as e:
    print(e)





























 {
    "type": "Title",
    "element_id": "bff1fd0ec25e78f1224ad7309a1e79c4",
    "text": "B All Experimental Results",
    "metadata": {
      "filetype": "application/pdf",
      "languages": [
        "eng"
      ],
      "page_number": 1,
      "filename": "CoT.pdf"
    }
  },
  {
    "type": "NarrativeText",
    "element_id": "ebf8dfb149bcbbd8c4b7f9a7046900a9",
    "text": "This section contains tables for experimental results for varying models and model sizes, on all benchmarks, for standard prompting vs. chain-of-thought prompting.",
    "metadata": {
      "filetype": "application/pdf",
      "languages": [
        "eng"
      ],
      "page_number": 1,
      "parent_id": "bff1fd0ec25e78f1224ad7309a1e79c4",
      "filename": "CoT.pdf"
    }
  },

image

Resources

https://unstructured.io/api-key-free

Get started in minutes with our free API hosted by Unstructured. Usage is capped at 1,000 pages per month.

Last changed by 
 
YH Hsu
www.linkedin.com/in/yung-hui-hsu-226b50a6
0
1
1670

Read more

[GenAI][AI Agents] Long-Term Agentic Memory With LangGraph - Introduction to Agent Memory

Long-Term Agentic Memory With LangGraph

Mar 19, 2025
[GenAI][AI Agents] Long-Term Agentic Memory With LangGraph - Baseline Email Assistant

Large Language Models with Semantic Search。大型語言模型與語義搜索

Mar 19, 2025
[AI Agents in LangGraph](https://learn.deeplearning.ai/courses/ai-agents-in-langgraph/lesson/1/introduction)

LanaGraph Components

Mar 19, 2025
AI / ML領域相關學習筆記入口頁面

Mar 19, 2025
Read more from YH Hsu
Published on HackMD