pdf-pro

Overview

PDFs are everywhere in business workflows: invoices, reports, contracts, forms. Working with them programmatically means three distinct jobs — generating PDFs from data, extracting text/structure from existing PDFs, and manipulating pages (merge, split, rotate, watermark). This skill covers the patterns and library choices for each, in any language ecosystem.

When to use

• Generating invoices, reports, or certificates from application data
• Extracting text or tables from PDFs for processing or search
• Merging, splitting, rotating, or watermarking PDF pages
• Filling PDF form fields (AcroForms) programmatically
• Choosing between HTML-to-PDF, native libraries, and template engines

Core concepts

Generation strategies. Three main approaches: (1) HTML/CSS → PDF renderers (great when you already have web templates and want pixel control via CSS); (2) native PDF libraries (precise, code-driven layout — best for complex documents like financial statements); (3) template fill (take a designed PDF and fill fields — best when design is fixed by someone else). Pick by who owns the design and how dynamic the layout is.

PDFs are a presentation format, not a data format. Text extraction is heuristic: PDFs store positioned glyphs, not paragraphs or tables. "Reading order" must be reconstructed, and scanned PDFs are images requiring OCR first. Always validate extraction against the visual document — especially for tables, multi-column layouts, and numbers.

Table extraction is its own problem. Bordered tables extract reasonably via line detection; borderless tables need whitespace analysis. Dedicated table extractors beat generic text extraction. For critical data (financial figures), build a verification step comparing extracted totals against known sums.

Forms (AcroForms). PDFs can carry named form fields; filling them programmatically preserves the original design exactly. Check whether the PDF has real fields (fillable) vs flat text that merely looks like a form — the latter needs overlay text placement instead.

Fonts and encoding. Embed fonts when generating (viewers substitute otherwise, breaking layout for non-Latin scripts). When extracting, know that some PDFs use custom glyph encodings that produce garbage text — a sign to try a different extractor or OCR.

Practical workflow

1. Classify the job: generate / extract / manipulate / fill — each has different best tools.
2. For generation: build an HTML template with print CSS (page size, margins, page-break rules) if the design is web-friendly; use a native library when you need precise control (headers/footers, flowing tables across pages, barcodes).
3. For extraction: try text extraction first; check quality on tables and multi-column pages; fall back to OCR for scanned documents; add a sanity-check layer (row counts, checksums, regex validation of key fields).
4. For manipulation: use a page-level library for merge/split/rotate/ watermark — these are byte-level operations that don't require re-rendering.
5. For forms: list field names first, then fill by name; flatten the form after filling if the result should be non-editable.
6. Test with real documents, not just the happy path — PDFs in the wild are malformed, huge, password-protected, or scanned; handle each case explicitly.

Common pitfalls

• Assuming text extraction preserves layout — it doesn't; columns interleave, tables collapse. Verify, don't assume.
• Generating PDFs without embedded fonts — renders fine on your machine, breaks on the client's (especially CJK, Arabic, or emoji content).
• Hardcoding coordinates for text overlay — fragile across PDF versions; prefer named form fields or anchor-based placement.
• Ignoring PDF/A for archival — if documents must be archivable long-term, generate PDF/A (embedded fonts, no encryption) from the start.
• Loading entire huge PDFs into memory — stream or page-window large documents; a 2GB PDF will OOM a naive process.
• Forgetting accessibility — tagged PDFs matter for screen readers in regulated contexts; HTML-to-PDF pipelines can preserve semantic structure if built with it in mind.