pdf-conversion-router – skill Tài liệu & Office cho agent AI

PDF Conversion Router

Route every PDF conversion through a short analysis step before choosing tools or CLI flags.

The goal is not "extract the most text". The goal is:

• preserve structure
• preserve attachment between labels and values
• choose the most faithful output shape
• avoid noisy defaults when a better route exists

When to Use

• The user wants a PDF converted into another format.
• The requested output is .md, .html, .txt, .json, .docx, or structured notes.
• The PDF may be scanned, OCR-heavy, table-heavy, slide-based, medical, academic, or multi-column.

Core Rule

Never start with one fixed default pipeline.

Always:

1. classify the PDF

2. classify the target output

3. choose the strongest route for that combination

4. validate the result on representative sections

5. if needed, retry with better settings before delivering

Heuristics are starting points, not guarantees.

Do not promote one flag combination into a universal default just because it worked well on one PDF.

Prefer document-specific evidence over habit.

Primary Engine Rule

Respect a user-selected engine and local-only/privacy constraints. Before conversion, inspect the installed version’s help for supported formats and flags; this catalog does not install an engine. Java/Python and optional OCR/backend dependencies must be available. Hybrid/full modes can send document pages to a backend: inspect its endpoint, retention, cost and authorization before enabling it, especially for medical or confidential files.

Use opendataloader-pdf as the primary conversion engine for every PDF conversion task by default.

This skill should assume:

• opendataloader-pdf is always the first conversion attempt
• other tools are used to classify, validate, OCR, inspect, or support cleanup
• other extractors are not the default replacement for the main conversion route

Use other tools only for one of these reasons:

• quick classification of the PDF
• OCR preprocessing before conversion
• validation against layout-preserving text
• manual repair when the generated output is still noisy
• fallback only if opendataloader-pdf cannot produce a usable result

Step 1: Classify the Source PDF

Identify the document class as quickly as possible:

• Native digital PDF with selectable text
• OCR PDF with noisy text
• Image-only/scanned PDF
• Slide deck / presentation export
• Medical or lab report
• Table-heavy business/finance document
• Narrative report / letter / article
• Mixed layout document with diagrams, tables, and prose

Useful fast checks:

pdfinfo input.pdf
pdftotext -layout input.pdf -

If text is missing or very poor, treat OCR as required.

Document-Type Heuristics

Use these as default starting points:

• medical / lab report

markdown-with-html + --table-method cluster + --image-output off

• slide deck / PowerPoint export

markdown-with-html + --image-output off

add --table-method cluster only if the default route under-structures important tabular content

if tables are visually obvious but missing or badly fused, treat this as a detection problem, not a Markdown formatting problem

if the selected route already reconstructs a real table but clips leading characters at column boundaries, treat that as a boundary-splitting defect, not a missing-table failure

• narrative / article / letter

start with markdown or text

use markdown-with-html only if structure clearly matters

• table-heavy business / finance PDF

start with markdown-with-html

add --table-method cluster when rows or columns flatten

• scanned / image-heavy PDF

OCR first, then convert with opendataloader-pdf

• mixed-layout PDF

prefer markdown-with-html

validate one easy section and one hard section before accepting output

Step 2: Choose the Output Shape

Pick the output that best matches the document and the user's goal.

• markdown-with-html

Use by default when the user wants Markdown and fidelity matters.

Prefer this for tables, medical reports, slides, mixed-layout PDFs, and anything likely to break in pure Markdown.

• markdown

Use only when clean plain Markdown matters more than layout fidelity.

• html

Use when visual structure matters more than LLM readability.

• text

Use for quick linear extraction, narrative documents, or when structure is unimportant.

• json

Use when downstream machine processing matters more than human readability.

• docx

Treat DOCX as a target requiring a separate verified reconstruction tool if the installed engine does not support it; renaming extracted Markdown or HTML does not create a DOCX.

Step 3: Choose the Extraction Route

For OpenDataLoader CLI

Use OpenDataLoader as the default route.

Preferred defaults:

• For Markdown output with fidelity priority:

-f markdown-with-html

• For medical PDFs:

add --table-method cluster

• For table-heavy PDFs:

add --table-method cluster

• For slide decks:

start without --table-method cluster

add it only after a structure check shows meaningful improvement

if a pseudo-table is already collapsed inside one detected row, changing only the Markdown flavor usually will not fix it

if the active engine build recovers the pseudo-table structure, prefer fixing residual boundary artifacts before escalating to hybrid/full mode

• For conversions where images are not requested:

add --image-output off

• For slide decks, medical reports, and structure-sensitive PDFs:

prefer validating both the command success and the actual rendered structure

• For referts/reports where exact values matter:

validate key sections after conversion instead of trusting first pass

For medical or lab PDFs

Default route:

opendataloader-pdf input.pdf -o output -f markdown-with-html --table-method cluster --image-output off

Then verify:

• main table headers
• attachment of value, unit, and reference range
• legends/comments separated from result rows

If a clinical table is flattened, compare against pdftotext -layout before accepting output.

For slide decks

Prefer:

opendataloader-pdf input.pdf -o output -f markdown-with-html --image-output off

Then check for:

• repeated footers
• page numbers
• diagram pseudo-tables
• orphan symbols and chart labels

If CLI output is still poor, do a cleanup pass tuned for slides instead of assuming the raw extract is final.

If the slide contains obvious table-like blocks that are not detected as tables at all, prefer a same-engine retry with a stronger route such as hybrid/full mode before jumping to unrelated extractors.

If the slide now produces a real table, validate the first column and header boundaries before assuming the table is fully correct.

For scanned PDFs

If the text layer is poor or absent:

• run OCR first
• then convert the OCR'd PDF with opendataloader-pdf

Prefer conservative reconstruction over aggressive guessing.

Step 4: Validation Gates

Before claiming success, inspect the output for the patterns most likely to break.

For medical PDFs:

• values attached to correct exam names
• units and reference ranges not merged into neighbors
• comments not merged into rows

For slides:

• bullets normalized
• footers/page numbers removed when they are noise
• diagrams not causing crashes
• remaining tables readable enough to follow
• first column labels not losing their first character at inferred column boundaries
• pseudo-table recovery not breaking row grouping or spilling labels into the next column

For table-heavy documents:

• no catastrophic row flattening
• headers preserved
• repeated empty separator rows minimized
• sparse or single-column tables not accidentally collapsed into prose
• table bodies not fused into a single HTML or Markdown row containing many logical records

For every document class:

• check the first representative section, not just the top of the file
• check one complex section, not only a simple section
• prefer document-level confidence over success on page 1

Red Flags

Treat these as signals that the current output is not ready:

• table rows flattened into long prose lines
• table header looks correct but the entire body is fused into one row with multi-value cells
• labels detached from values
• units or reference ranges drifting into adjacent rows
• repeated page footers or page numbers
• pseudo-tables with mostly empty cells
• legitimate sparse tables collapsed into paragraphs
• single-column tables flattened because they looked "too simple"
• stray symbols, bullets, or OCR fragments
• good command exit code but visibly poor structure
• page 1 looks fine but a later complex section is broken
• switching from markdown to markdown-with-html improves wrapping but does not restore missing row boundaries
• a pseudo-table is now emitted as a table, but key labels are clipped at the left edge of cells

Never Trust Page 1

Do not accept a conversion just because the top o

… (nội dung đầy đủ có trong gói tải về)