OPP - Omni Pre-Processor

Document content extraction for DOCX, PPTX, PDF, XLSX, CSV, JSON, XML, HTML, EPUB, EML, MSG, and Image (OCR).

Features

• Multi-format extraction - DOCX, PPTX, PDF, XLSX, CSV, JSON, XML, HTML, EPUB, EML, MSG, Image, IPYNB, YouTube URL
• Inline formatting tracking - Bold, italic, underline, strikethrough preserved in XLIFF as <bx>/<ex> tags for downstream formatting restoration
• Image OCR - Tesseract and RapidOCR with graceful fallback
• Email extraction - EML (RFC 822) and MSG (Outlook) with attachment recursion
• Audio/Video transcription - Whisper-based ASR
• Format auto-detection - Magic bytes detection (extension not required)
• Resource management - MD5 deduplication, UUID naming for images
• Floating image extraction - Distinguishes wp:anchor (floating) from wp:inline drawings and carries wp:positionH/wp:positionV offsets in EMU units through images.json for downstream ORF reinjection
• Pipeline orchestrator - detect → extract → manage → report
• CLI interface - Full command-line with batch support
• Output formats - Markdown and XLIFF 1.2/2.0
• Manifest metadata - JSON manifest with source info, extraction stats, and image data
• Skeleton preservation - Original DOCX/PPTX ZIP structure preserved for downstream XLIFF→DOCX/PPTX backfill

Prerequisites

• Python >= 3.13 — required by all Omni Suite components (OPP, OL, ORF). Verify with python3 --version.

Installation

# Core package
pip install -e .

# With office/data formats (XLSX, CSV, JSON, XML)
pip install -e ".[office]"

# With email and OCR (EML, MSG, Tesseract, RapidOCR)
pip install -e ".[email]"

可选依赖

除 office、email 外，以下可选依赖组支持特定输入格式：

依赖组	命令	用途	体积
[ocr]	pip install -e ".[ocr]"	图片 OCR（RapidOCR）	~100MB
[youtube]	pip install -e ".[youtube]"	YouTube .url 自动检测	~50MB
[audio]	pip install -e ".[audio]"	音频转录（Whisper）	~2GB

• PDF 输入生成 XLIFF 已被正确拦截——此前 format_type == "PDF" 因大小写匹配错误从未生效，现已修复为 "pdf"。
• .url 文件现在被自动检测为 YouTube 源（读取首行 URL 并匹配 YouTube 域名）。

Quick Start

Python API

from opp import DOCXExtractor, PDFExtractor, PPTXExtractor
from opp.detector import detect_format
from opp.pipeline import OPPPipeline

# Direct extraction
extractor = DOCXExtractor()
result = extractor.extract("document.docx")
print(result.content)

# Auto-detection
fmt, confidence = detect_format("document.docx")
print(f"Format: {fmt.value}, Confidence: {confidence}")

# Full pipeline
pipeline = OPPPipeline(resource_storage_dir="./resources")
result = pipeline.process_file("document.docx")
print(f"Extracted: {len(result.content)} chars, {result.images_stored} images")

CLI

# Extract to Markdown
opp --target-format=md document.docx

# Extract to XLIFF for translation
opp --target-format=xlf --source-lang=en --target-lang=zh document.docx

# Generate both MD and XLIFF
opp --target-format=both --source-lang=en --target-lang=zh document.docx

# Custom output directory
opp --target-format=md --output-dir ./output document.docx

# Image OCR
opp --ocr-engine tesseract scan.png

# Batch processing
opp --batch file1.docx file2.pdf file3.pptx

Note: Windows .bat helper scripts are no longer provided. Use the CLI directly — see examples above.

Supports drag-drop of files and folders. Logs saved to logs/.

Output Files

Every extraction produces a manifest.json and optionally a skeleton.zip:

output_dir/
├── document.md              # Extracted Markdown
├── document.xlf             # Extracted XLIFF (translation-ready)
├── document_manifest.json   # Metadata about source and extraction
└── document.skeleton.zip    # Original DOCX/PPTX ZIP (for backfill)

manifest.json

Records source file info, extraction outputs, and resources:

{
  "manifest_version": "1.0",
  "generated_at": "2026-05-22T14:30:00Z",
  "tool": "OPP",
  "tool_version": "0.2.0",
  "source": {
    "file_path": "/path/to/spec.docx",
    "original_filename": "spec.docx",
    "format": "DOCX",
    "file_size_bytes": 45824,
    "file_hash_md5": "a1b2c3d4e5f6..."
  },
  "extraction": {
    "source_lang": "en",
    "target_lang": "zh",
    "outputs": {
      "markdown": { "path": "spec.md", "paragraph_count": 150, "table_count": 3 },
      "xliff": { "path": "spec.xlf", "trans_unit_count": 42 }
    },
    "images": [
      { "mime_type": "image/png", "width": 800, "height": 600, "data_size_bytes": 24580 }
    ],
    "warnings": []
  },
  "resources": { "storage_dir": "resources", "image_count": 5 },
  "skeleton": {
    "path": "spec.skeleton.zip",
    "format": "ZIP",
    "key_files": ["word/document.xml", "word/styles.xml", "[Content_Types].xml"]
  }
}

skeleton.zip

Preserves the original OOXML ZIP structure for DOCX/PPTX files. This enables downstream ORF tools to perform XLIFF→DOCX/PPTX backfill by replacing content in the preserved skeleton.

Format	Key Files Preserved
DOCX	word/document.xml, word/styles.xml, word/numbering.xml, word/settings.xml, [Content_Types].xml
PPTX	All files under ppt/ prefix (slides, layouts, media)

Project Structure

src/opp/
├── detector.py           # Format auto-detection
├── extractors/           # Document extractors
│   ├── docx.py
│   ├── pptx.py
│   ├── pdf.py
│   ├── xlsx.py
│   ├── csv.py
│   ├── json.py
│   ├── xml.py
│   ├── email.py
│   └── image_ocr.py
├── channels/             # Output formatters
│   ├── table_channel.py   # DataFrame → Markdown table
│   └── keyvalue_channel.py # dict → XLIFF
├── xliff/                # XLIFF 1.2/2.0 generator
├── pipeline.py           # OPPPipeline orchestrator
├── resource_manager.py   # Image deduplication
└── cli.py               # Command-line interface

Architecture

                     ┌─────────────────────────────────────────┐
                     │              OPPPipeline                  │
                     │  detect_format() → Extractor → Report   │
                     └─────────────────────────────────────────┘

┌──────────┐    ┌───────────┐    ┌────────────────┐    ┌──────────────┐
│ detector │───▶│ extractors│───▶│resource_manager│───▶│error_handler │
│  magic   │    │  DOCX/...  │    │  MD5 + UUID    │    │ HTML/text    │
└──────────┘    └───────────┘    └────────────────┘    └──────────────┘

Development

pip install -e ".[dev]"
pytest tests/ -v --cov=src/opp --cov-report=term-missing

Test Coverage

Module	Tests
detector	13
resource_manager	18
error_handler	18
integration	25
cli	18
e2e	52
xliff	40+
extractors	140+
inline formatting	53
manifest generation	6
skeleton preservation	6
Total	544+

Batch Testing

Test files available in batch_test/ covering all formats.

opp --target-format=both --source-lang=en --target-lang=zh --output-dir=output batch_test/

Pipeline — Omni Localization Suite

OPP is Step 1 of the Omni Localization Suite pipeline:

┌────────────────────────────────────────────────────────────────────────┐
│                     OMNI LOCALIZATION SUITE                             │
│                                                                        │
│  ┌─────────────┐    ┌─────────────┐    ┌─────────────┐               │
│  │     OPP     │───▶│     OL      │───▶│     ORF      │               │
│  │  (提取)     │    │   (翻译)    │    │   (回写)    │               │
│  └─────────────┘    └─────────────┘    └─────────────┘               │
│                                                                        │
│  Step 1: OPP        Step 2: OL            Step 3: ORF                  │
│  Extract →          Translate →           Backfill →                  │
│  MD + XLIFF +       MD + XLIFF            DOCX/PPTX                   │
│  skeleton.zip                                                    │
└────────────────────────────────────────────────────────────────────────┘

Complete Workflow

# Step 1: OPP - Extract document to MD/XLIFF + skeleton.zip
opp --target-format=both --source-lang=en --target-lang=zh document.docx
# Output: document.md, document.xlf, document_manifest.json, document.skeleton.zip

# Step 2: OL - Translate to target language
ol translate-md document.md -s en -t zh -o translated/

# Step 3: ORF - Backfill translated content to target format
orf apply-xliff document.docx --xliff translated/document.xlf --output result.docx

Related Projects

• OL (Omni-Localizer) - NEXT STEP after OPP. Translates MD/XLIFF produced by OPP.
• ORF (Omni-Re-Formatter) - Backfills translated content to DOCX/PPTX/EPUB.

For AI Agents

OPP outputs standardized artifacts for downstream processing:

Artifact	Description	Used By
{name}.md	Markdown with YAML frontmatter (source_lang, target_lang)	OL (translate-md)
{name}.xlf	XLIFF 1.2/2.0 with <bx>/<ex> inline tags	OL (translate-xliff)
{name}_manifest.json	Metadata: source info, output paths, resources	ORF (manifest parser)
{name}.skeleton.zip	Original DOCX/PPTX ZIP structure	ORF (XLIFF→DOCX backfill)

MCP Tools Available: extract_document, batch_extract, detect_format_tool, generate_markdown, generate_xliff, save_skeleton, ping

Floating Image Metadata

For DOCX inputs that contain anchored (floating) images, OPP emits an is_floating: true flag alongside wp_anchor_h and wp_anchor_v fields in images.json — both expressed in EMU (English Metric Units, 914400 EMU = 1 inch). The paragraph_index for floating drawings is None because they are not anchored to a w:p element. ORF consumes these fields to reinject <wp:anchor> blocks with <wp:positionH>/<wp:positionV> when backfilling the translated DOCX, preserving the original page layout. Inline images keep the previous JSON shape (no is_floating key) so downstream consumers that don't care about floating layout remain unaffected.

MCP Server (Agent-Facing)

The OPP MCP server provides document extraction capabilities to AI agents via the Model Context Protocol. AI assistants can use these tools to process documents without needing to understand OPP's internal architecture.

Why Use the MCP Server?

• Agent integration - Connect OPP to any MCP-compatible AI assistant
• stdio transport - Communication over standard input/output for security
• 7 extraction tools - Cover all major document formats
• Path security - Directory allowlist prevents unauthorized file access

Installation

# Install OPP with MCP server support
pip install -e ".[mcp]"

Quick Start

Start the server manually:

python -m opp.mcp.server

Auto-start with uvx:

uvx opp-mcp-server

Auto-start with npx:

npx opp-mcp-server

Hermes Configuration

Add OPP to your Hermes agent configuration:

agents:
  my-agent:
    tools:
      - name: opp
        type: code
        config:
          server_command: uvx opp-mcp-server
          allowed_directories:
            - /path/to/documents
            - /path/to/output

Available Tools

Tool	Description
extract_document	Extract content from a single document file. Supports DOCX, PPTX, PDF, XLSX, CSV, JSON, XML, HTML, EPUB, EML, MSG, and images. Returns markdown or structured content.
batch_extract	Process multiple files in one request. Takes an array of file paths and processes them sequentially. Returns extraction results for each file.
detect_format	Identify the file format of a document using magic bytes detection. Works regardless of file extension. Returns format name and confidence score.
generate_markdown	Convert a document to markdown format. Specify source and target languages for proper text processing.
generate_xliff	Convert a document to XLIFF format for translation workflows. Requires source-lang and target-lang parameters.
save_skeleton	Save the skeleton ZIP for an extracted document, preserving original OOXML structure (required by ORF apply-xliff).
ping	Health check endpoint. Returns server version and status.

Security

The MCP server enforces path validation to prevent unauthorized file access.

Allowlist configuration:

# Via environment variable
export OPP_MCP_ALLOWED_DIRS="/allowed/documents,/allowed/output"

# Via configuration file
# Create opp_mcp_config.yaml with:
# security:
#   allowed_directories:
#     - /path/to/documents
#     - /path/to/output

**Configuration file** (`opp_mcp_config.yaml`):

```yaml
security:
  allowed_directories:
    - <your-documents-dir>
    - <your-output-dir>
    - ./documents

server:
  host: localhost
  port: 8765

extraction:
  default_target_format: md
  ocr_engine: tesseract

Environment Variables

Variable	Scope	Description	Default
OPP_MCP_ALLOWED_DIRS	MCP	Colon/semicolon-separated allowlist of directories the MCP server can read (e.g. /docs:/tmp/out). Required for any opp mcp tool call to succeed.	(none)
OPP_ALLOWED_DIRECTORIES	CLI only	Comma-separated allowlist for the CLI's --resource-dir guard (src/opp/cli.py:496). NOT read by the MCP server.	(none)
OPP_MCP_MAX_FILE_SIZE	MCP	Max input file size in bytes	104857600 (100 MB)
OPP_MCP_TIMEOUT	MCP	Per-tool request timeout in seconds	300
OPP_MCP_HOST	MCP	Bind host	127.0.0.1
OPP_MCP_PORT	MCP	Bind port	8766
OMNI_METRICS_DIR	MCP	Prometheus metrics directory	/tmp/omni-metrics
MCP_SHARED_SECRET	MCP	Shared-secret auth (Phase A4)	(none — auth disabled)
OPP_RESOURCE_STORAGE_DIR	CLI	Directory for extracted images	./resources
OPP_OCR_ENGINE	CLI	OCR engine	tesseract
OPP_OCR_LANG	CLI	OCR language	eng
OPP_LOG_LEVEL	CLI	Log level	INFO
OMNI_LOG_FORMAT	CLI/MCP	console (default) or json	console
OMNI_TEST_FAKE_LLM=1	CLI	Mock LLM responses (hermetic testing)	unset