1
支持的输入格式:
2
├── PDF → Markdown
3
├── DOCX → Markdown
4
├── PPTX → Markdown
5
├── XLSX → Markdown（表格转 Markdown 表格）
6
├── HTML → Markdown
7
├── 图片 (OCR) → Markdown（含文字提取）
8
├── 音频 → Markdown（语音转文字）
9
└── Zip → Markdown（解压后逐文件处理）

1
# 使用 MarkItDown 处理文档（Python API）
2
from markitdown import MarkItDown
3

4
md_converter = MarkItDown()
5

6
# 转换 PDF
7
result = md_converter.convert("document.pdf")
8
print(result.text_content)  # 标准化的 Markdown 输出
9

10
# 批量处理目录
11
import pathlib
12

13
docs_dir = pathlib.Path("./documents/")
14
for ext in ["*.pdf", "*.docx", "*.pptx"]:
15
    for file_path in docs_dir.glob(ext):
16
        result = md_converter.convert(str(file_path))
17
        output_path = docs_dir / f"{file_path.stem}.md"
18
        output_path.write_text(result.text_content)
19
        print(f"✅ {file_path.name} → {output_path.name}")

1
# GitHub Actions 工作流：自动转换文档并提交
2
name: Document Conversion Pipeline
3

4
on:
5
  push:
6
    paths:
7
      - "docs/source/**"
8

9
jobs:
10
  convert:
11
    runs-on: ubuntu-latest
12
    steps:
13
      - uses: actions/checkout@v4
14

15
      - uses: actions/setup-python@v5
16
        with:
17
          python-version: "3.12"
18

19
      - name: Install MarkItDown
20
        run: pip install markitdown
21

22
      - name: Convert all documents
23
        run: |
24
          mkdir -p docs/markdown
25
          markitdown batch ./docs/source --out-dir ./docs/markdown
26

27
      - name: Check diff
28
        run: git diff --stat
29

30
      - name: Commit and push
31
        run: |
32
          git config user.name "Doc Bot"
33
          git config user.email "docbot@contoso.com"
34
          git add docs/markdown/
35
          git commit -m "docs: auto-convert source docs to Markdown [skip ci]"
36
          git push

1
# 完整的 RAG Pipeline：MarkItDown → 分块 → 向量化 → 搜索
2
from markitdown import MarkItDown
3
from azure.search.documents import SearchClient
4
from azure.search.documents.indexes import SearchIndexClient
5
from azure.core.credentials import AzureKeyCredential
6
from langchain.text_splitter import MarkdownHeaderTextSplitter
7
import os
8

9
# Step 1: 用 MarkItDown 批量转换文档
10
md_converter = MarkItDown()
11

12
def process_document(file_path: str) -> list[dict]:
13
    """处理单个文档，返回分块后的内容"""
14
    result = md_converter.convert(file_path)
15
    markdown_content = result.text_content
16

17
    # 按 Markdown 标题层级分块（保留结构信息）
18
    headers_to_split_on = [
19
        ("#", "H1"),
20
        ("##", "H2"),
21
        ("###", "H3"),
22
    ]
23
    splitter = MarkdownHeaderTextSplitter(
24
        headers_to_split_on=headers_to_split_on,
25
        strip_headers=False
26
    )
27

28
    chunks = splitter.split_text(markdown_content)
29

30
    documents = []
31
    for i, chunk in enumerate(chunks):
32
        documents.append({
33
            "id": f"{os.path.basename(file_path)}-{i}",
34
            "title": chunk.metadata.get("H1", "") or chunk.metadata.get("H2", ""),
35
            "content": chunk.page_content,
36
            "section": " > ".join(chunk.metadata.values()),
37
            "source": file_path,
38
            # 这里会让 embedding 模型生成向量
39
            "content_vector": None,  # 稍后由 AI Search 自动填充
40
        })
41

42
    return documents
43

44
# Step 2: 处理整个文档目录
45
all_docs = []
46
docs_dir = "./product-docs/"
47
for ext in ["*.pdf", "*.docx", "*.md"]:
48
    for file_path in sorted(glob.glob(os.path.join(docs_dir, ext))):
49
        docs = process_document(file_path)
50
        all_docs.extend(docs)
51
        print(f"✅ {file_path}: {len(docs)} chunks")
52

53
# Step 3: 上传到 Azure AI Search（自动向量化）
54
search_client = SearchClient(
55
    endpoint=os.environ["AZURE_SEARCH_ENDPOINT"],
56
    index_name="product-knowledge-base",
57
    credential=AzureKeyCredential(os.environ["AZURE_SEARCH_API_KEY"]),
58
)
59

60
# 批量上传（自动触发内置的向量化 pipeline）
61
result = search_client.upload_documents(documents=all_docs)
62
print(f"上传完成: {len(result)} 个文档")
63

64
# Step 4: 搜索查询
65
def search_docs(query: str, top: int = 5):
66
    results = search_client.search(
67
        search_text=query,
68
        query_type="semantic",
69
        semantic_configuration_name="default",
70
        top=top,
71
        select=["title", "content", "section", "source"],
72
        highlight_fields="content",
73
    )
74

75
    for r in results:
76
        print(f"\n📄 {r['title']}")
77
        print(f"   来源: {r['section']}")
78
        print(f"   匹配片段: {r['content'][:200]}...")
79
        print(f"   相关度: {r['@search.score']:.3f}")

1
# .gitlab-ci.yml — 2026 年的 Agentic CI/CD
2
stages:
3
  - agentic-plan
4
  - build
5
  - agentic-review
6
  - deploy
7

8
# Agent 自动规划任务
9
agentic-plan:
10
  stage: agentic-plan
11
  script:
12
    - gitlab-agent plan --from-diff --output plan.md
13
    - cat plan.md | llm "审查这个部署计划，给出风险分析"
14
  artifacts:
15
    paths: [plan.md]
16

17
build:
18
  stage: build
19
  script:
20
    - npm ci
21
    - npm run build
22
    - npm run test:coverage
23

24
# AI Agent 自动代码审查
25
agentic-review:
26
  stage: agentic-review
27
  script:
28
    - gitlab-agent review --diff origin/main...HEAD \
29
        --focus security,performance \
30
        --auto-approve-minor
31
  only:
32
    - merge_requests
33

34
deploy:
35
  stage: deploy
36
  script:
37
    - gitlab-agent deploy --environment production \
38
        --rollback-on-failure \
39
        --canary-percent 10
40
  environment:
41
    name: production
42
  needs: [build, agentic-review]