PRC PDF Extraction

Extracts text from topline survey PDFs using AI OCR providers and surfaces the content as web pages, Markdown, and plain text for researchers and LLMs.

Overview

This plugin reads topline-type entries from a parent post's reportMaterials meta, runs AI OCR on the attached PDFs, and persists each result as a pdf_extraction child post. Extracted content is available at /{parent-slug}/extraction (Markdown with YAML frontmatter) and /{parent-slug}/text (plain text). Extraction jobs are queued asynchronously via Action Scheduler and can also be triggered from WP-CLI or the block editor via REST API. The plugin integrates with prc-markdown-for-agents to serve standardized Markdown responses with proper headers for LLM discovery.

Dependencies

• Upstream: prc-platform-core (required), prc-markdown-for-agents (optional — Markdown endpoint falls back to a legacy template when absent), Action Scheduler (required for async processing)
• External APIs: PRC_PLATFORM_ANTHROPIC_API_KEY for Claude (primary provider), PRC_PLATFORM_GOOGLE_API_KEY for Gemini (fallback provider)
• Downstream: Nothing depends on this plugin directly; consumers access extractions through public URL endpoints or via the pdf_extraction post type.

Architecture

Extractions follow a request → queue → process → persist flow:

1. An editor triggers extraction from the block editor (REST) or WP-CLI.
2. Action_Scheduler_Handler::schedule() enqueues an async job.
3. On the job callback, Extraction_Service resolves the PDF (local uploads dir or remote download), calls OCR_Orchestrator::extract_text(), and saves the result as a pdf_extraction CPT post.
4. OCR_Orchestrator tries registered providers sorted by priority. Claude (priority 4) runs first; Gemini (priority 5) is the fallback. Each provider returns an OCR_Response with plain text, Markdown, and Gutenberg block HTML.
5. Public endpoints (/extraction, /text) are intercepted early in parse_request by Rewrite_Rules, which delegates Markdown serving to prc-markdown-for-agents when available.
6. Content_Discovery adds <link rel="alternate"> tags, JSON-LD DigitalDocument structured data, /llms.txt, and robots.txt Allow rules to make extractions discoverable by crawlers and LLMs.

The OCR layer uses a layered namespace under PRC\Platform\PDF_Extraction\OCR:

Key Files

Hooks & Filters

The plugin also filters prc_markdown_for_agents_frontmatter to inject OCR metadata (provider, confidence, extraction date, source URL, parent article link) into the YAML frontmatter of Markdown responses.

REST API

Both endpoints require edit_post capability on the target post.

WP-CLI Commands

# Process a single post's topline PDF
wp prc pdf-extraction process --post_id=123

# Force reprocess (overwrites existing extraction)
wp prc pdf-extraction process --post_id=123 --force

# Test with a local PDF file without saving to the database
wp prc pdf-extraction test-file --file=/path/to/test.pdf

# Preview bulk scheduling across all posts with reportMaterials (dry-run is default)
wp prc pdf-extraction bulk-process

# Enqueue bulk extraction jobs
wp prc pdf-extraction bulk-process --dry-run=false

# Resume a bulk run after interruption from a specific post ID
wp prc pdf-extraction bulk-process --dry-run=false --start-id=12345

# List configured OCR providers and their availability
wp prc pdf-extraction list-providers

# Check extraction status for a post
wp prc pdf-extraction list-extractions --post_id=123

# Validate an existing extraction post
wp prc pdf-extraction validate --post_id=456

# Estimate OCR cost before running
wp prc pdf-extraction estimate-cost --post_id=123

Provider-specific options

# Force a specific OCR provider (bypasses orchestrator priority)
wp prc pdf-extraction process --post_id=123 --provider=claude
wp prc pdf-extraction process --post_id=123 --provider=gemini

# Override the Gemini model for a single test run
wp prc pdf-extraction test-file --file=/path/to/test.pdf --provider=gemini --model=gemini-3-flash-preview

# Show Markdown output instead of plain text in test-file
wp prc pdf-extraction test-file --file=/path/to/test.pdf --show-markdown

Stored Post Meta

Each pdf_extraction post stores the following meta (all exposed via REST):

Troubleshooting

Extraction fails with "file not found (may have expired)"

Symptom: Claude provider throws an extraction error referencing an expired file. Logs show Claude API: file not found (may have expired).

Cause: Claude's Files API caches uploaded PDFs for approximately 30 days. The file_id is stored on the WP attachment as _claude_file_id. When the file expires on Anthropic's end, subsequent calls referencing the cached file_id return a 404.

Fix: The provider automatically retries once by deleting the stale _claude_file_id meta and re-uploading the file. If you need to force a manual reset, delete the _claude_file_id post meta from the attachment, then trigger a new extraction.