pdf-inspector 完全指南：Firecrawl 开源的 Rust 高性能 PDF 解析引擎

[dependencies]
pdf-inspector = "0.2"

cargo add pdf-inspector

pip install maturin
git clone https://github.com/firecrawl/pdf-inspector.git
cd pdf-inspector
maturin develop --release

npm install @firecrawl/pdf-inspector

npm install @firecrawl/pdf-inspector-wasm

import pdf_inspector

# 处理单个 PDF 文件
result = pdf_inspector.process_pdf("document.pdf")

# 查看 PDF 类型分类
print(f"PDF 类型: {result.pdf_type}")
# 输出: "text_based", "scanned", "image_based", 或 "mixed"

# 获取 Markdown 输出
if result.markdown:
    print(result.markdown)

const { processPdf } = require('@firecrawl/pdf-inspector');
const fs = require('fs');

async function analyzePdf(filePath) {
  const pdfBuffer = fs.readFileSync(filePath);
  const result = await processPdf(pdfBuffer);

  console.log('PDF 类型:', result.pdfType);
  console.log('置信度:', result.confidence);
  console.log('--- Markdown 输出 ---');
  console.log(result.markdown);
}

analyzePdf('report.pdf');

use pdf_inspector::{PdfInspector, ProcessPdfResult};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let inspector = PdfInspector::new();
    let result: ProcessPdfResult = inspector.process_pdf_path("document.pdf")?;

    println!("PDF 类型: {:?}", result.pdf_type);
    println!("置信度: {:.2}", result.confidence);

    if let Some(markdown) = result.markdown {
        println!("Markdown:\n{}", markdown);
    }

    Ok(())
}

import pdf_inspector
# 假设你有自己的 OCR 函数
def smart_pdf_pipeline(pdf_path, ocr_function):
    result = pdf_inspector.process_pdf(pdf_path)

    if result.pdf_type == "text_based":
        # 原生文本，直接提取——快速、免费、准确
        print("✅ 原生文本 PDF，直接提取")
        return result.markdown

    elif result.pdf_type == "scanned":
        # 扫描件，调用 OCR
        print("🔍 扫描件 PDF，调用 OCR")
        return ocr_function(pdf_path)

    elif result.pdf_type == "mixed":
        # 混合型——逐页处理
        print("📄 混合型 PDF，逐页路由")
        pages = []
        for page_info in result.page_results:
            if page_info.needs_ocr:
                pages.append(ocr_function_for_page(pdf_path, page_info.page_number))
            else:
                pages.append(page_info.markdown)
        return "\n\n".join(pages)

    else:
        # 纯图片型，无法提取文本
        print("❌ 图片型 PDF，无法提取文本")
        return None

# Deep Learning Approaches for Natural Language Processing

## 1. Introduction

Natural language processing (NLP) has seen remarkable progress in recent years...

### 1.1 Background

The transformer architecture, introduced by Vaswani et al., has become...

## 2. Methodology

| Model | BLEU Score | Training Time |
|-------|-----------|---------------|
| Transformer | 38.2 | 12 hours |
| BERT | 41.0 | 24 hours |
| GPT-4 | 45.7 | 72 hours |

## 3. Results

The experimental results demonstrate...

> **Note:** All experiments were conducted on...

result = pdf_inspector.process_pdf("newspaper.pdf")

# pdf-inspector 自动检测双栏/三栏布局
# 并按从上到下、从左到右的正确顺序输出
print(result.markdown)

# pdf_service.py
import os
import json
from datetime import datetime
import pdf_inspector

class PdfProcessingService:
    def __init__(self, output_dir="./processed"):
        self.output_dir = output_dir
        os.makedirs(output_dir, exist_ok=True)

    def process_batch(self, pdf_files):
        """批量处理 PDF 文件"""
        results = []
        for pdf_path in pdf_files:
            result = self.process_single(pdf_path)
            results.append(result)
        return results

    def process_single(self, pdf_path):
        """处理单个 PDF 文件"""
        filename = os.path.basename(pdf_path)
        pdf_result = pdf_inspector.process_pdf(pdf_path)

        # 提取元信息
        metadata = {
            "filename": filename,
            "pdf_type": pdf_result.pdf_type,
            "confidence": pdf_result.confidence,
            "processed_at": datetime.now().isoformat(),
            "pages_count": len(pdf_result.page_results) if hasattr(pdf_result, 'page_results') else 0,
            "needs_ocr": pdf_result.pdf_type in ("scanned", "image_based"),
        }

        # 保存 Markdown 输出
        if pdf_result.markdown:
            md_filename = filename.replace(".pdf", ".md")
            md_path = os.path.join(self.output_dir, md_filename)
            with open(md_path, "w", encoding="utf-8") as f:
                f.write(pdf_result.markdown)
            metadata["markdown_path"] = md_path

        # 保存元信息
        meta_path = os.path.join(
            self.output_dir,
            filename.replace(".pdf", ".meta.json")
        )
        with open(meta_path, "w", encoding="utf-8") as f:
            json.dump(metadata, f, ensure_ascii=False, indent=2)

        return metadata

# 使用示例
service = PdfProcessingService(output_dir="./output")

# 批量处理
pdf_files = [
    "contracts/contract_2026_001.pdf",
    "invoices/invoice_aug_2026.pdf",
    "reports/q2_financial_report.pdf",
]

results = service.process_batch(pdf_files)
for r in results:
    status = "需要 OCR" if r["needs_ocr"] else "已直接提取"
    print(f"{r['filename']}: {r['pdf_type']} ({status})")

import pdf_inspector
# 假设你使用 OpenAI 兼容 API
from openai import OpenAI

def analyze_pdf_with_llm(pdf_path, prompt="总结这份文档的主要内容"):
    """用 LLM 分析 PDF 内容"""
    # 1. 用 pdf-inspector 提取 Markdown
    result = pdf_inspector.process_pdf(pdf_path)

    if not result.markdown:
        return "无法提取文本内容，可能需要 OCR"

    # 2. 如果文档太长，截取前 8000 字符
    markdown_content = result.markdown[:8000]

    # 3. 调用 LLM 分析
    client = OpenAI(api_key="YOUR_API_KEY")
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": "你是一个文档分析助手。"},
            {"role": "user", "content": f"请基于以下文档内容{prompt}：\n\n{markdown_content}"}
        ]
    )

    return response.choices[0].message.content

# 使用
summary = analyze_pdf_with_llm("annual_report.pdf", "提取关键财务指标")
print(summary)