1
输入: 124,832 tokens (一个完整的 CI 构建日志)
2
输出: 18,724 tokens (经过 headroom 处理后)
3
节省: ≈85%
4
回答质量: 人眼评估无差异

1
// headroom 的基础用法（Library 模式）
2
import { Headroom } from "@headroom/compressor"
3

4
const headroom = new Headroom({
5
  // 压缩级别：conservative | balanced | aggressive
6
  mode: "balanced",
7
  // 是否保留错误信息（推荐保留）
8
  preserveErrors: true,
9
  // 最大输出比例（相对于输入）
10
  maxOutputRatio: 0.3,  // 最多保留 30%
11
})
12

13
// 压缩一段构建日志
14
const buildLog = `
15
[2026-06-03T10:23:01.234Z] Starting build pipeline #43821
16
[2026-06-03T10:23:01.456Z] Checking out branch: feature/vector-search-optimization
17
[2026-06-03T10:23:02.001Z] Running: git checkout feature/vector-search-optimization
18
[2026-06-03T10:23:02.890Z] SUCCESS: Checkout completed in 0.889s
19
[2026-06-03T10:23:03.100Z] Running: npm ci
20
... (省略 15,000 行)
21
[2026-06-03T10:28:01.200Z] SUCCESS: Build completed in 4m 59s
22
`
23

24
const compressed = await headroom.compress(buildLog)
25
console.log(`压缩前: ${buildLog.length} chars → 压缩后: ${compressed.length} chars`)
26
// 输出: 压缩前: 682341 chars → 压缩后: 102354 chars

1
import os
2
from headroom_compressor import Headroom, CompressionMode
3
from langchain.text_splitter import RecursiveCharacterTextSplitter
4

5
headroom = Headroom(mode=CompressionMode.BALANCED)
6

7
def prepare_documents_for_rag(file_paths: list[str]) -> list[str]:
8
    chunks = []
9

10
    for path in file_paths:
11
        with open(path, "r", encoding="utf-8") as f:
12
            content = f.read()
13

14
        # 1. 先用 headroom 压缩
15
        compressed = headroom.compress(content)
16

17
        # 2. 按段落分块
18
        splitter = RecursiveCharacterTextSplitter(
19
            chunk_size=1000,
20
            chunk_overlap=200,
21
            separators=["\n\n", "\n", "。", ".", " ", ""]
22
        )
23
        doc_chunks = splitter.split_text(compressed)
24

25
        original_tokens = len(content) // 4  # 粗略估计
26
        compressed_tokens = len(compressed) // 4
27
        savings = (1 - compressed_tokens / original_tokens) * 100
28

29
        print(f"  {os.path.basename(path)}: {original_tokens}→{compressed_tokens} tokens (节省 {savings:.1f}%)")
30
        chunks.extend(doc_chunks)
31

32
    return chunks
33

34
# 处理一批文档
35
docs = prepare_documents_for_rag([
36
    "./docs/api-reference-v3.md",      # 1.2M
37
    "./docs/deployment-guide.md",      # 3.5M
38
    "./docs/troubleshooting.md",       # 2.1M
39
])

1
# headroom 的高级分析模式
2
from headroom_compressor import HeadroomAnalyzer
3

4
analyzer = HeadroomAnalyzer()
5

6
# 分析 CI 日志
7
log_data = open("ci-build.log").read()
8

9
report = analyzer.analyze(log_data, mode="ci_pipeline")
10

11
print("=== 构建摘要 ===")
12
print(f"构建状态: {'✅ 成功' if report.success else '❌ 失败'}")
13
print(f"总耗时: {report.duration}")
14
print(f"模块数量: {len(report.modules)}")
15
print(f"失败模块: {[m.name for m in report.failed_modules]}")
16
print(f"测试通过率: {report.test_pass_rate:.1f}%")
17
print(f"压缩比: {report.compression_ratio:.1f}x")
18

19
# 输出压缩后的精简版本（适合直接喂给 LLM）
20
compressed = report.to_llm_prompt(max_tokens=4000)
21
print(f"压缩后 LLM 输入 ({len(compressed)} tokens):")
22
print(compressed)