sboghossian Import: PDF Processing (Anthropic)
What it does
This import adapter migrates a PDF document-processing skill originally built for Anthropic's Claude API into the mini-claude-for-legal standard format. PDF is the dominant format for legal documents globally: court filings, notarised instruments, regulatory submissions, signed contracts, government-issued licences, and certified translations all arrive as PDFs.
Unlike DOCX, PDFs may or may not have a selectable text layer. Scanned documents require OCR; encrypted PDFs require a password; form PDFs have interactive fields; certified documents may have a digital signature that must be preserved. The Anthropic-native skill may have used Claude's native PDF vision capability, a dedicated extraction library, or both.
Import config
| Field | Source mapping | Default if absent |
|---|---|---|
extraction_mode |
Legacy mode |
text_first_ocr_fallback |
ocr_engine |
Legacy ocr |
auto (system default) |
tables |
Legacy extract_tables boolean |
true |
form_fields |
Legacy extract_forms boolean |
true |
annotations |
Legacy extract_annotations boolean |
true |
digital_signature_check |
Legacy check_signature boolean |
true |
language |
Legacy lang |
auto-detect |
rtl_support |
Legacy rtl boolean |
true (for MENA) |
chunk_size |
Legacy chunk_tokens |
2000 |
overlap |
Legacy overlap_tokens |
200 |
output_format |
Legacy format |
markdown |
Dry-run preview
IMPORT PREVIEW — pdf-processing-anthropic
Source shape : Anthropic PDF extraction config
Mode : text_first_ocr_fallback
OCR : auto
Tables : extracted
Form fields : extracted
Annotations : extracted
Digital signature: checked (not verified cryptographically — visual only)
RTL : enabled (Arabic/Hebrew support)
Chunk : 2000 tokens / 200 overlap
Output : markdown
Extraction pipeline (post-import)
Detect PDF type:
- Native (selectable text) → direct text extraction
- Scanned image → OCR pipeline
- Mixed (some pages scanned, some native) → page-by-page detection
- Encrypted → request password; log attempt; never store password
Page-by-page extraction: maintain page-number metadata for every extracted element; legal documents frequently reference "page X, clause Y".
Table extraction: detect table boundaries; serialize to markdown pipe tables; flag tables with merged cells as
[COMPLEX TABLE — manual verification recommended].Form field extraction: identify interactive PDF form fields; extract field names and values; flag unsigned or blank required fields.
Annotation extraction: extract highlighted text, comments, and sticky notes; attribute to author if metadata available; append as footnote block.
Digital signature check: detect presence of digital signature (visual check only — not cryptographic validation); flag signed documents separately.
RTL handling: detect right-to-left text (Arabic, Hebrew); preserve paragraph ordering; ensure column order in tables reflects RTL layout.
Chunking: split by token count with overlap; preserve clause boundaries where possible.
Legal document types and special handling
| Document type | Special requirement |
|---|---|
| Court filing (Lebanon / UAE) | Page/paragraph numbering must be preserved; filing stamps and seals should be flagged |
| Notarised document | Notary signature block and certification text must be extracted intact |
| Certified translation | Translator certification page must be preserved as a separate block |
| Government licence / certificate | Issue date, expiry date, and licence number should be extracted as structured metadata |
| Board resolution / POA | Signatory block and witness/notary endorsement are legally significant |
| Regulatory submission | Submission reference number and date should be extracted as metadata |
Arabic / bilingual PDF considerations
PDFs from MENA jurisdictions frequently mix Arabic (RTL) and English (LTR) text:
- Set
rtl_support: trueto handle RTL paragraphs - In bilingual contracts, Arabic is often the prevailing language (UAE and KSA government contracts); flag and note the prevailing language
- OCR quality for Arabic varies by font and scan quality; flag low-confidence passages
Failure modes
| Error | Likely cause | Resolution |
|---|---|---|
no_text_layer |
Scanned PDF without OCR | Switch to OCR mode; warn user of quality risk |
password_protected |
Encrypted PDF | Prompt for password; never log |
rtl_reversal |
RTL paragraphs extracted as LTR | Ensure rtl_support: true |
table_parse_fail |
Complex merged-cell tables | Flag for manual review |
signature_not_found |
Digital signature present but not detected | Flag as [SIGNATURE STATUS UNKNOWN] |
arabic_ocr_low_quality |
Poor scan of Arabic text | Warn user; recommend re-scan at 300 DPI minimum |
Related skills
- [[import-docx-processing-anthropic]]
- [[import-pptx-processing-anthropic]]
- [[import-contract-review-anthropic]]
- [[multimodal-document-ingestion]]
- [[review-contract-generic]]