feat: Release v0.10 - Modular Architecture & External Config

- Refactor codebase into src/ (preprocessing, translation, assembly)
- Add pipeline/ scripts for individual stages
- Externalize configuration to config/config.yaml
- Fix Cover Image preservation
- Update documentation and manuals
This commit is contained in:
谭凯
2026-01-31 22:49:44 +08:00
parent 9ef82393be
commit 7a93c52b42
306 changed files with 30313 additions and 1071 deletions
+90
View File
@@ -0,0 +1,90 @@
# Operation Manual & Change Log
## Core Principles
1. **Modularity**: The system is divided into three distinct phases (Preprocessing, Translation, Assembly) with clear boundaries.
2. **Immutability**: `book_structure.json` is generated once during preprocessing and should not be modified by subsequent steps.
3. **Source of Truth**: `manifest.json` is the single source of truth for translations.
4. **Idempotency**: Translation steps can be retried without side effects (existing translations are preserved).
## Directory Structure
* `pipeline/`: Executable scripts for each stage.
* `01_preprocess.py`: Clean EPUB, generate structure, extract text.
* `02_translate.py`: Translate text in manifest.
* `03_assemble.py`: Apply translations and build final EPUB.
* `src/`: Core logic modules.
* `preprocessing/`: Cleaning, extraction, profiling.
* `translation/`: LLM integration, manifest management.
* `assembly/`: Backfilling, EPUB building.
* `common/`: Shared data models and utils.
* `work/`: Working directory for intermediate files (ignored by git).
## Pipeline Usage
### Step 1: Preprocessing
```bash
python pipeline/01_preprocess.py inputs/my_book.epub
```
Generates `work/my_book/book_structure.json` and `manifest.json`.
### Step 2: Translation
```bash
python pipeline/02_translate.py --input-epub inputs/my_book.epub
```
Translates entries in `manifest.json`. ensuring `.env` has `OPENAI_API_KEY`.
### Step 3: Assembly
```bash
python pipeline/03_assemble.py inputs/my_book.epub --mode bilingual
```
Generates `output/my_book_bilingual.epub`.
## Configuration
System settings are managed via `config/config.yaml` and environment variables.
### `config/config.yaml`
Control LLM parameters and translation behavior:
```yaml
llm:
model: "gpt-3.5-turbo" # LLM Model Name
base_url: "https://api.openai.com/v1"
timeout: 60
requests_per_minute: 60 # Rate limiting
concurrent_requests: 5 # Parallel chunks
translation:
chunk_size: 4000 # Characters per chunk
```
### Environment Variables (`.env`)
Security-sensitive credentials must be set here:
```bash
OPENAI_API_KEY=sk-... # Required
OPENAI_BASE_URL=... # Optional override for config
```
## Known Issues & Troubleshooting
### Missing Placeholders Warning
During assembly, you may see logs like:
`WARNING - Restoration warning: missing placeholders {'1'}`
This indicates that the LLM translation missed a placeholder tag (e.g. `φ1φ`). The system attempts to recover, but this warning is logged for review. These are usually minor and do not prevent EPUB generation.
## Change Log
### [2026-01-28] Bug Fixes
* **Fix Cover Image**: Resolved issue where book cover execution was missing in the final EPUB. Added `cover_image_id` tracking in `BookStructure` and restored proper OPF metadata in `BilingualBuilder`.
### [2026-01-27] Externalized Configuration
* **Config**: Added `config/config.yaml` for tuning parameters (LLM model, RPM, Chunk Size).
* **Logic**: `pipeline/02_translate.py` now loads settings from `config.yaml`.
* **Dependency**: Added `PyYAML` to `requirements.txt`.
### [2026-01-27] Architecture Refactoring
* **Restructured**: Moved source files into `src/preprocessing`, `src/translation`, `src/assembly`, `src/common`.
* **Pipeline**: Created individual pipeline scripts in `pipeline/`.
* **Refactor**: Renamed `fine_grained_extractor` to `text_extractor`, `translator` to `translator_engine`, etc.
* **Logic**: Enforced 100% text coverage check in `format_extractor.py` (removed 95% threshold).
* **Docs**: Created this Operation Manual.
+100
View File
@@ -0,0 +1,100 @@
# ePub Bilingual Translator - Architecture & Data Flow (Revised)
本文档详细描述了程序处理一个 EPUB 文件的完整生命周期。整个流程旨在实现**结构稳定性**(不丢段落)与**内容精细度**(不丢格式)的最佳平衡。
## High Level Data Flow
```mermaid
graph TD
Input[Input EPUB] --> Cleaner[EpubCleaner]
Cleaner --> CleanedEPUB[1. Cleaned EPUB Temp]
CleanedEPUB --> Profiler[Book Profiler]
Profiler -->|Identify| Profile[Book Profile / Style]
CleanedEPUB --> Extractor[FineGrainedExtractor]
subgraph Extraction
Extractor -->|Step 1: P/H Tags| P[Source Paragraph]
P -->|Step 2: FormatExtract| PhText[Analyzed Text]
PhText -->|Register| Manifest[Manifest DB]
end
Manifest -->|Batch| LLM[LLM Translation]
Profile -->|Prompt Context| LLM
LLM -->|Translation| Manifest
Manifest --> Restorer[FormatRestorer]
Restorer -->|Reconstruct HTML| TargetHtml[Target HTML]
CleanedEPUB --> Backfiller[Backfill Engine]
TargetHtml --> Backfiller
Backfiller -->|Bilingual/Chinese Mode| DOM[Final DOM]
DOM --> Builder[BilingualBuilder]
Builder --> Output[Output EPUB]
```
## Detailed Workflow
### 1. Preprocessing (预处理)
**模块**: `src/epub_cleaner.py`
* **Flatten Structure**: 消除嵌套 `div`,统一转为 `<p>`,消除结构性漏译风险。
* **Auto-Fix**: 修复 TOC 死链、缺失 UID、由于 `ebooklib` bug 导致的样式丢失。
* **Result**: 产生一个标准的临时文件,后续所有操作基于此文件,不再受原始糟糕格式影响。
### 2. Intelligent Extraction (智能提取)
**模块**: `src/fine_grained_extractor.py` + `src/format_extractor.py`
#### A. 结构层 (Macro)
使用 `FineGrainedExtractor` 锁定所有正文元素 (`p`, `h1`-`h6`)。
* **Filter**: 排除页码、页眉脚。
* **Optimization**: 针对目录章节,识别 **罗马数字 (I, II)**、**单独数字 (1, 2)**、**修饰符 (***)**,这些内容**不送翻译**,直接在回填时保留原文,以维持原书排版美感。
#### B. 内容层 (Micro)
对每个提取的段落调用 `FormatExtractor`
* **Inline Style**: 将 `<b>`, `<i>` 转为配对占位符 `φ1φ...φ/1φ`
* **Formula Protection**: 识别 $E=mc^2$ 等数学公式,保护为不可变占位符。
* **Drop Cap Handling**:
- 原始: `<span class="dropcap">T</span>he`
- 提取给 LLM: "The" (完整单词,无格式干扰)
- 记录: Prefix 包含 Drop Cap 样式。
### 3. Manifest Management (清单管理)
**模块**: `src/manifest_manager.py`
Manifest 是系统的**核心状态中心 (Source of Truth)**。
* **作用**: 解耦提取和翻译。提取器只管往 Manifest 填数据,翻译器只管从 Manifest 取数据。
* **Persistence**: 支持中断续传,翻译进度实时保存。
### 4. Translation with Profiling (翻译)
**模块**: `src/book_profiler.py` & `src/translator.py`
* **Profiling**: 在翻译前,抽取部分文本分析书籍的类型(技术、小说、诗歌)、核心术语和语言风格,生成 `System Prompt`
* **Translation**: 这是纯文本层面的转换,LLM 处理的是带有 `φ` 占位符的文本。
### 5. Robust Restoration (健壮还原)
**模块**: `src/format_restorer.py`
负责将 LLM 返回的文本还原为 HTML。
* **Drop Cap Logic**:
- **原文回填**: 需要 Prefix `<span class="dropcap">T</span>`
- **译文回填**: **丢弃** Drop Cap Prefix。中文不需要首字母下沉,否则会出现 "T这本书..." 的怪诞结果。
* **Error Handling**:
- **Missing Placeholders**: 如果 LLM 丢了 `φ1φ`,自动在末尾补全或报错重试。
- **Hallucinated Placeholders**: 移除 LLM 臆造的不存在 ID。
### 6. Backfill Strategy (回填策略)
**模块**: `src/fine_grained_extractor.py` (backfill method)
支持多种模式,且**严格遵循一对一 (One-to-One) 映射**,绝不依赖顺序,而是依赖元素的内存引用或唯一 ID。
* **Mode A: Bilingual (双语)**
- 保留原文 DOM。
- 在原文后 `append` 一个新元素 `<p class="translation">译文</p>`
- 样式继承:译文元素复制原文的 `margin`, `text-align` 等关键样式。
* **Mode B: Chinese Only (仅中文)**
- **Replace**: 直接用译文元素替换原文元素。
- **Drop Cap 适配**: 此时译文通常为普通段落,不再保留首字母下沉样式,以符合中文排版习惯。
### 7. Assembly (组装)
**模块**: `src/bilingual_builder.py`
将内存中的 DOM 序列化,重新打包资源,生成最终 EPUB。
+180
View File
@@ -0,0 +1,180 @@
# 预处理与回填层技术文档
本文档详细描述 EPUB 双语翻译器的预处理(Preprocessing)和回填(Backfill)层的技术方案、数据结构、问题解决方案,用于指导代码开发、调试和维护。
---
## 1. 架构概览
```
┌─────────────┐ ┌────────────────┐ ┌───────────────────┐
│ EPUB 文件 │ ──► │ EpubCleaner │ ──► │ book_structure.json│
└─────────────┘ └────────────────┘ └───────────────────┘
┌────────────────┐ ┌───────────────────┐
│FineGrainedExt. │ ──► │ manifest.json │
└────────────────┘ └───────────────────┘
┌────────────────┐ │
│ BackfillEngine │ ◄───────────┘
└────────────────┘
┌────────────────┐ ┌───────────────────┐
│BilingualBuilder│ ──► │ 输出 EPUB │
└────────────────┘ └───────────────────┘
```
---
## 2. 目录结构
```
work/
├── {book_name}/ # 每本书独立的工作目录
│ ├── book_structure.json # 书籍结构(会被复用)
│ ├── manifest.json # 翻译清单(持久化,Source of Truth
│ └── assets/ # 解压出的 EPUB 资源文件
│ └── OEBPS/images/... # 保留原始路径结构
└── translations/ # (可选) 翻译记忆或其他中间文件
```
**关键设计**`book_structure.json` **默认复用**。为避免 UUID 不一致导致回填失败,除非显式指定强制清理,否则程序优先读取现有的结构文件。
---
## 3. 数据结构定义
### 3.1 BookStructure (book_structure.json)
记录书籍的“骨架”,确保翻译后的章节能按正确顺序和层级重组。
```json
{
"metadata": {
"title": "书名",
"author": "作者",
"language": "en",
"identifier": "ISBN或UUID"
},
"spine": ["item_id_1", "item_id_2", ...], // 阅读顺序
"resources": {
"item_id_1": {
"href": "OEBPS/chapter1.xhtml",
"media_type": "application/xhtml+xml",
"content": "<html>...</html>", // 清理并注入ID后的HTML内容
"properties": "nav"
},
...
}
}
```
### 3.2 ManifestEntry (manifest.json)
翻译清单是全生命周期的核心,记录了所有待翻译段落的状态。
```json
[
{
"entry_id": "OEBPS/c3Z.xhtml#uuid-603ff3fb",
"file_path": "OEBPS/c3Z.xhtml",
"element_id": "uuid-603ff3fb",
"original_text": "φ1φHelloφ/1φ world.",
"placeholders": {
"1": "<b>",
"/1": "</b>",
"_prefix": "<p>",
"_suffix": "</p>"
},
"translated_text": "φ1φ你好φ/1φ 世界。",
"context": "body"
}
]
```
| 字段 | 类型 | 说明 |
|------|------|------|
| `entry_id` | str | **全局唯一ID** (`file_path#element_id`),用于精准追踪。 |
| `file_path` | str | 标识该段落属于哪一章,用于按章分组批处理。 |
| `element_id` | str | HTML DOM 元素的 ID,回填时的锚点。 |
| `original_text` | str | 经过智能抽提和占位符化后的文本,发送给 LLM。 |
| `placeholders` | Dict | 格式映射表,用于还原 HTML 结构。 |
| `translated_text`| str | 翻译结果(含占位符),初始为 null。 |
---
## 4. 预处理与抽提流程 (Extraction)
流程由 `FormatExtractor` 驱动,分为三个阶段:
### Step 1: 结构索引 (Structure Indexing)
* **动作**: 解析 OPF 文件,提取 Spine 和 Metadata。
* **目的**: 建立骨架,确定处理顺序。
### Step 2: 语义识别 (Semantic Detection)
* **动作**: 使用 `HeadingDetector` 分析 DOM 节点。
* **逻辑**:
* 匹配 `h1`-`h6` 正则,区分 `Chapter`(章)与 `Section`(节)。
* 识别 `blockquote` 或特定 class 判定 `Epigraph`(引言)。
* **目的**: 为 LLM 提供差异化的 Prompt(例如翻译标题时不要加句号)。
### Step 3: 智能抽提 (Smart Extraction - `_smart_extract_v3`)
这是核心算法,将 HTML 转换为“纯文本+占位符”。
1. **首尾分离 (Prefix/Suffix Separation)**:
* 将包裹文本的外层标签(如 `<p>`, `div`)剥离到 `_prefix``_suffix`
* **目的**: 极大减少 LLM 输入 Token,且防止 LLM 随意修改外层布局。
2. **首字下沉处理 (Drop Cap Handling)**:
* 检测并合并被 `<span>` 单独包裹的首字母(如 `<span class="drop">O</span>` + `nce``Once`)。
* **目的**: 修复语意割裂,让 LLM 看到完整的单词。
3. **占位符化 (Placeholder Mapping)**:
* 将内联标签(`<a>`, `<em>`)或公式替换为短码 `φIDφ`
* **目的**: 保护 HTML 属性不被“翻译”,降低噪声干扰。
4. **完整性校验 (Integrity Check)**:
* **逻辑**: 抽提后的文本(去占位符)与原始纯文本进行归一化比对,要求覆盖率 **100%**
* **兜底**: 若校验失败(如误删内容),回退到简单模式(只剥离首尾标签)。
---
## 5. 构建与回填流程 (Backfill & Build)
### 5.1 翻译回填
* 根据 `entry_id` 找到对应的 DOM 节点。
* 使用 `FormatRestorer``translated_text` 中的占位符(`φ1φ`)还原为原始 HTML 标签(`<b>`)。
* 根据模式(双语/单语)决定将新节点插入到原文后还是替换原文。
### 5.2 链接修复 (Link Repair)
`BilingualBuilder` 中执行:
1. **ID 补全**: 为缺失 ID 的 TOC 节点自动生成 UUID。
2. **死链检测**: 检查 TOC/Nav 指向的文件是否存在。
3. **模糊修复**: 尝试通过文件名后缀匹配(解决路径前缀变更问题)或特定重定向(如 `c0.xhtml` -> `cover.xhtml`)。
4. **坏死剔除**: 无法修复的死链将从目录中移除。
### 5.3 CSS 样式恢复
`EbookLib` 默认可能会重写 `<head>` 导致样式丢失。
* **逻辑**:
* 收集所有 CSS 资源。
* 在构建每个 HTML Item 时,显式计算 HTML 到 CSS 的**相对路径**。
* 强制调用 `item.add_link(..., rel='stylesheet', type='text/css')` 注入引用。
---
## 6. 常见问题排查
### 6.1 UUID 不匹配
**现象**: `WARNING - Element uuid-xxx not found`.
**原因**: 手动删除了 `book_structure.json` 但保留了 `manifest.json`,导致重新生成的 HTML ID 与清单记录不一致。
**解决**: 清空 `work/BookName` 目录重新运行,或确保两个 JSON 文件版本一致。
### 6.2 样式丢失
**现象**: 打开书面目全非,只有黑白文字。
**检查**: 解压 EPUB,查看 HTML `<head>` 是否有 `<link rel="stylesheet">`。如果没有,检查 `BilingualBuilder` 的 Step 4 逻辑。
### 6.3 翻译错位
**现象**: 译文出现在了错误的位置。
**检查**: 确认 `entry_id` 生成逻辑是否包含文件名,且文件名在处理过程中未被意外修改。
+108
View File
@@ -0,0 +1,108 @@
# Translation Layer Technical Documentation
This document details the architecture of the **Translation Layer**, which operates primarily on `manifest.json`.
> **Core Principle**: The Translation Layer is **decoupled** from the EPUB file format. It reads translatable units from `manifest.json`, processes them using an LLM, and writes translations back to `manifest.json`. It does **not** read or parse the EPUB file directly.
---
## 1. Architecture Overview
### Data Flow
```mermaid
graph LR
A[manifest.json] -->|Load| B(ManifestManager)
B -->|Entries| C{Translator Orchestrator}
C -->|Sample Text| D[BookProfiler]
D -->|Style Guide| C
C -->|Chunks| E[LLMClient]
E -->|Translation| C
C -->|Update| B
B -->|Save| A[manifest.json]
```
1. **Input**: `manifest.json` (Generated by Preprocessing Layer).
2. **Process**:
* **Profiling**: Analyze text samples to generate a `BookProfile` (style, tone, terminology).
* **Translation**: Batch entries into chunks, send to LLM, receive translations.
3. **Output**: `manifest.json` (Updated with `translated_text` fields).
---
## 2. Core Modules (`src/translation/`)
### 2.1 ManifestManager (`manifest_manager.py`)
* **Role**: The interface for the "Source of Truth".
* **Responsibility**:
* Load `manifest.json`.
* Provide list of `ManifestEntry` objects.
* Save updates back to disk.
* **Key Method**: `update_translation(entry_id, translation)`
### 2.2 Translator Engine (`translator_engine.py`)
* **Role**: Orchestrates the translation process.
* **Responsibility**:
* **Filtering**: Identify untranslated entries.
* **Grouping**: Group entries by chapter (file path) to maintain context.
* **Chunking**: Create character-based chunks (default ~5000 chars) that do not cross chapter boundaries.
* **Concurrency**: Manage async workers (default 5 concurrent tasks).
* **Input**: `List[ManifestEntry]`, `BookProfile`.
* **Output**: Updates `ManifestEntry` objects in-place.
### 2.3 LLM Client (`llm_client.py`)
* **Role**: Handles raw communication with the LLM Provider (OpenAI compatible).
* **Responsibility**:
* **Prompt Engineering**: Construct Short-ID based prompts.
* **Rate Limiting**: Control RPM (Requests Per Minute).
* **Retry Logic**: Exponential backoff for API failures.
* **Logging**: Save raw request/response pairs to `work/{book}/chunks/` for debugging.
### 2.4 Book Profiler (`../preprocessing/profiler.py`)
* **Note**: While located in preprocessing, it is often invoked at the start of the translation phase.
* **Role**: Generates a style guide.
* **Mechanism**: Extracts a sample (Intro + Random segments) from `manifest.json` entries and asks the LLM to analyze author style.
---
## 3. Short ID Strategy
To optimize token usage and ensuring mapping accuracy, we use a **Short ID** system for LLM interaction.
**Prompt Format**:
```text
#1: First paragraph text...
#2: Second paragraph text...
```
**Response Format**:
```text
#1: 第一段翻译...
#2: 第二段翻译...
```
The `LLMClient` maintains a mapping of `Short ID (#N)` <-> `Entry ID (File#UUID)` for each chunk lifecycle.
---
## 4. Pipeline Usage
The translation is executed via the standalone pipeline script:
```bash
python pipeline/02_translate.py --input-epub inputs/my_book.epub
```
* **--input-epub**: Uses the filename to locate the `work/` directory.
* **--book-name**: Alternatively, specify the book folder name directly.
### Dependencies
* Environment variables must be set in `.env`:
* `OPENAI_API_KEY`
* `OPENAI_BASE_URL` (Optional)
---
## 5. Development & Debugging
* **Chunk Logs**: Check `.work/{book}/chunks/` to see exactly what was sent to and received from the LLM.
* **Idempotency**: The translation script skips entries that already have `translated_text`. To re-translate, you must manually clear `translated_text` in `manifest.json` or delete the manifest (to restart from preprocessing).