sboghossian Scanned PDF Handler (Router Entry Point)
What it does
This is the lightweight, router-level entry point invoked when a user uploads a PDF and the system detects it has no usable text layer. Its job is:
- Confirm the document is scanned (not native-text).
- Select the appropriate OCR language mode (English, Arabic, French, or auto-detect).
- Delegate to the full OCR pipeline ([[voice-multimodal-scanned-pdf-handler]]).
- Surface a user-confirmation step when overall OCR confidence is low.
Think of this skill as the gatekeeper that ensures no scanned document silently fails into downstream analysis with garbled text.
Detection logic
A PDF requires scanned handling if:
- Page text token count is zero or near-zero for the first three pages.
- Text extracted is a pattern of gibberish characters consistent with embedded-font issues or pre-baked OCR artifacts.
- The file's creation metadata shows a scanner as the producing application.
- The user explicitly describes the file as scanned ("I scanned this contract", "it's a scan").
If only some pages are scanned and others have native text, process each page type independently and merge the results.
Language modes
| Mode | When to use |
|---|---|
| English | Document is in English only |
| Arabic | Document is in Arabic only |
| French | Document is in French only |
| Auto-detect | Document has mixed language sections (most common for MENA commercial documents) |
Default to auto-detect unless the user specifies a language. Auto-detect applies per-page language identification before routing to the appropriate OCR engine.
User-confirmation fallback
If the overall OCR confidence score for the document falls below 80%:
Show the user a preview of the extracted text from the first page with this message:
"I was able to extract text from this scanned document, but the quality is limited. Here is a sample — if it looks correct, I'll proceed. If the text looks wrong, you may get better results by uploading a higher-resolution scan."
[Show first-page extracted text preview]
[Looks good — proceed] [I'll upload a better scan]
This is preferable to silently passing low-quality OCR text into an analysis skill, which would produce unreliable legal output.
Handoff to full pipeline
Once the scanned-PDF determination is confirmed, delegate entirely to [[voice-multimodal-scanned-pdf-handler]] for:
- Page-by-page processing
- Table detection
- Signature page flagging
- PII redaction before RAG ingestion
- Per-block confidence scoring
This skill does not duplicate that logic — it is the entry point, not the full implementation.
What to tell the user
When a scanned PDF is detected, acknowledge it proactively:
"This looks like a scanned document — I'll run OCR to extract the text before we work with it. This may take a moment for longer documents."
Do not silently process. Users need to know why there is a delay and what is happening.
Related skills
- [[voice-multimodal-scanned-pdf-handler]]
- [[voice-image-of-contract-page-handler]]
- [[voice-multimodal-signature-page-detector]]
- [[voice-dictation-cleanup]]