DocumentToMarkdown

Convert scientific documents into structured, searchable markdown using AI-powered vision analysis.

Built for pharmaceutical research — handles conference posters, patent filings, presentation slides, and corporate presentations (PPTX + PDF) out of the box.

What It Does

Pipeline	Input	Output
Poster	Conference poster PDFs	Structured markdown with figures, sections, and metadata
Patent	Patent filing PDFs (WO/EP/US)	Claims, chemical structures, SMILES, executive summaries
Talk	Slide-based presentation PDFs	Slide-by-slide extraction with narrative summaries
Presentation	PPTX + PDF presentations	Native text extraction, per-slide image descriptions, action items, chemical structures
Paper	Scientific research paper PDFs	Full-text extraction with sections, figures, and summary
CI	Competitive intelligence PDFs + PPTXs	Company/asset structured markdown, deal & regulatory event extraction
Outlook	Outlook email folders	Threaded conversations, processed attachments, persistent contact book

Each pipeline produces a self-contained .md file with YAML frontmatter, making the output easy to search, filter, and integrate into knowledge bases.

Quick Start

First time? See the full setup guide for step-by-step installation from scratch (Miniconda, packages, credentials, troubleshooting).

1. Setup Environment

conda create -n ds_env python=3.11 -y
conda activate ds_env

conda install -c conda-forge pdfplumber pymupdf pillow -y
conda install pandas openpyxl -y
conda install -c conda-forge tesseract pytesseract -y
pip install anthropic

2. Set Credentials

[Environment]::SetEnvironmentVariable("ANTHROPIC_AUTH_TOKEN", "your-token", "User")
[Environment]::SetEnvironmentVariable("ANTHROPIC_BASE_URL", "your-proxy-url", "User")

Restart your terminal after setting these.

3. Run a Pipeline

Posters:

# with metadata
python poster_pipeline.py --input "path/to/poster_pdfs" --metadata "abstracts.xlsx"
# without metadata
python poster_pipeline.py --input "path/to/poster_pdfs"
# standardized filenames
python poster_pipeline.py --input "path/to/poster_pdfs" --naming standardized --conference AACR --year 2026

Patents:

python patent_pipeline.py --input "path/to/patent_pdfs"
# or a single file:
python patent_pipeline.py --single "path/to/WO2024123456.pdf"
# recursive scan with detailed naming:
python patent_pipeline.py --input "path/to/patent_pdfs" --recursive --naming detailed

Talks:

python talk_pipeline.py --input "path/to/talk_pdfs" --metadata "abstracts.xlsx"
# or a single file:
python talk_pipeline.py --single "path/to/talk.pdf"

Presentations:

python presentation_pipeline.py --input "path/to/presentations"
# text-only (no API calls):
python presentation_pipeline.py --input "path/to/presentations" --no-vision
# single file with dated naming:
python presentation_pipeline.py --single "path/to/file.pptx" --naming dated

Papers:

python paper_pipeline.py --input "path/to/papers"
# single file:
python paper_pipeline.py --single "path/to/paper.pdf"

Competitive Intelligence:

python ci_pipeline.py --input "path/to/ci_docs"
# custom output location:
python ci_pipeline.py --input "path/to/ci_docs" --output "output_ci"

Outlook Emails:

python outlook_pipeline.py --folder "Inbox/CI Reports"
# limit to recent emails, skip attachments:
python outlook_pipeline.py --folder "Inbox" --limit 50 --no-attachments
# reprocess everything:
python outlook_pipeline.py --folder "Inbox/Projects" --no-skip

Common Flags

All pipelines share a consistent CLI interface:

Flag	Description
--input	Input folder containing documents (aliases: --sharepoint, --talks)
--output	Output directory for markdown files
--single	Process a single file instead of a folder
--recursive	Recursively search subfolders for files
--no-skip	Reprocess files that already exist (default: skip existing)
--naming	Output filename scheme (options vary per pipeline)
--verbose	Enable debug logging

Each pipeline also has specialized flags — see the pipeline-specific documentation for full options:

• Poster Pipeline flags — --metadata, --force-ocr, --conference, --col-*
• Patent Pipeline flags — --no-vision, --claims-only, --budget, --ocr-engine
• Talk Pipeline flags — --metadata
• Presentation Pipeline flags — --no-vision
• Outlook Pipeline flags — --folder, --no-attachments, --limit

Output Structure

Each pipeline generates markdown files with:

• YAML frontmatter — structured metadata (dates, authors, scores, classifications)
• Executive summary — AI-generated overview of the document
• Full content — sections, claims, or slides extracted from the PDF
• Figures & visuals — descriptions of charts, structures, and diagrams
• Quality score — automated 0–10 quality assessment

Example output location:

output/
├── poster_1234.md
├── poster_1235.md
└── quality_log.txt

output_patents/
├── patent_WO2024123456A1.md
└── quality_log.txt

output_talks/
├── talk_04_ED03_Bunne_toward_virtual_patients.md
└── processing_log.txt

output_presentations/
├── presentation_ru_onc_operations_update_darmstadt.md
└── presentation_caris_discovery_non_con_apr.md

output_outlook/
├── processed_state.json
├── processing_log.tsv
├── people/                         # global contact book
│   ├── carsten_schweer.yml
│   └── alice_smith.yml
└── Inbox/                          # mirrors Outlook folder hierarchy
    └── CI Reports/
        ├── thread_project_alpha_update/
        │   ├── thread.md
        │   └── attachment_report.md
        └── thread_meeting_notes_q2/
            └── thread.md

Pipeline Comparison

At a Glance

	Poster	Patent	Talk	Presentation	Outlook
Input	Single-page poster PDF	Multi-page patent PDF (50–300+ pages)	Multi-slide presentation PDF (screenshots)	PPTX or PDF slide decks	Outlook email folder
Output	Sections (Methods, Results, Conclusions)	Patent sections + claims + chemical data	Slide-by-slide content + narrative summary	Slide content + action items + metrics	Threaded conversations + processed attachments
Text extraction	Native PDF + OCR + Vision AI	Native PDF + Tesseract OCR (scanned) + selective Vision AI for garbled pages	OCR + Vision AI only (no extractable text)	Native PPTX (python-pptx) or PyMuPDF + pdfplumber tables + Vision AI fallback	Outlook COM (plain text / HTML conversion)
Metadata source	Excel spreadsheet (optional)	Extracted from the PDF itself	Excel spreadsheet (optional)	Extracted from the file itself	Extracted from email headers
Quality gate	Yes — skips FAIR/POOR	No — all patents saved	No — all talks saved	No — all presentations saved	No — all threads saved

Performance & Cost

	Poster	Patent	Talk	Presentation	Outlook
Processing time	1.5–4 min	2–3 min (text+vision), 3–5 min (scanned+Tesseract), ~15s text-only	1.5–4 min	<1s text-only, 30–90s with vision, 2–4 min image-heavy	~1s per email (text only), +pipeline time for attachments
API calls per doc	4 + N figures	10–48 (text-native), 8–35 (scanned+Tesseract)	4–9 (scales with slide count)	1–3 (smart gating) + 5–15 (image enrichment), 0 text-only	0 (email body), varies for attachments
Token usage per doc	~30K–60K	~40K–120K	~18K–35K	~8K–20K (text-only), ~50K–120K (image-heavy)	0 (email body), varies for attachments
Vision AI pages	All pages (mandatory)	~10% of pages (selective)	All slides (mandatory)	Smart gating: global analysis + per-slide image enrichment (requires PowerPoint for PPTX)	None (attachments only)
Text-only mode	No	Yes (--no-vision)	No	Yes (--no-vision)	Yes (default for email body)
Claims-only mode	No	Yes (--claims-only, ~5s)	No	No	No
Concurrency	Up to 5 figure workers	Up to 3 batch workers	Up to 5 batch workers	Sequential	Sequential

Unique Capabilities

Capability	Poster	Patent	Talk	Presentation	Outlook
Two-stage figure analysis	x
Chemical structure extraction (SMILES)		x		x
Claims dependency tree		x
Semantic classification (target, mechanism, modality)		x
Hybrid OCR (Tesseract + Vision AI)		x
Text quality scoring & auto-repair		x
OCR pre-pass as RAG context	x		x
Abstract matching from metadata	x		x
Native PPTX text extraction				x
Per-slide image enrichment (auto-describes figures/screenshots)				x
Smart Vision AI gating (skip when not needed)				x
Language detection (EN/DE) + English output				x
Classification detection (3-signal)				x
Action items & metrics extraction				x
Conditional summary (content-type aware)				x
Conversation threading (ConversationID + subject)					x
Incremental processing (EntryID state tracking)					x
Attachment routing to sub-pipelines					x
HTML-to-markdown email body conversion					x
Standardized naming schemes	x	x		x
Executive summary	x	x	x	x
Quality scoring	x	x	x	x

When to Use Which

• Poster — single-page conference posters with figures, methods, and results sections
• Patent — multi-page patent filings (WIPO, EPO, USPTO) with claims, chemical structures, and experimental data
• Talk — slide-based presentations captured as PDF screenshots (no extractable text)
• Presentation — PPTX or PDF slide decks with native text (corporate meetings, scientific presentations, agendas); supports English and German content
• Outlook — email conversations from any Outlook folder; automatically processes PDF/PPTX/DOCX/XLSX attachments through the appropriate pipeline

How It Works

1. Text Extraction — native text via PyMuPDF/pdfplumber/python-pptx, with OCR fallback (Tesseract) for posters/talks/scanned patents
2. Page Rendering — pages rendered to images using PyMuPDF (PDFs) or PowerPoint COM (PPTX) when Vision AI is needed
3. Vision AI — Claude analyzes page images for figures, chemical structures, and garbled text (smart gating skips this when not needed)
4. Structuring — extracted content is parsed into logical sections
5. Summarization — AI generates executive summaries and key findings (always in English)
6. Quality Scoring — automated scoring flags documents that may need manual review

Documentation

Guide	Description
Setup Guide	Full installation from scratch — Miniconda, packages, credentials, troubleshooting
Poster Pipeline	Deep dive — architecture, processing stages, figure analysis, quality scoring
Patent Pipeline	Deep dive — claims parsing, SMILES extraction, text quality repair, semantic classification
Talk Pipeline	Deep dive — slide extraction, OCR pre-pass, summary generation, batch processing
Presentation Pipeline	Deep dive — PPTX/PDF extraction, classification detection, action items, naming schemes
Outlook Pipeline	Deep dive — email threading, incremental processing, attachment routing

Requirements

• Python 3.11+ (via Conda)
• Tesseract OCR (poster/talk/patent pipelines — required for poster/talk, optional for patent scanned PDFs)
• python-pptx, PyMuPDF, pdfplumber (presentation/paper pipelines)
• python-docx (outlook pipeline — DOCX attachment extraction)
• pyyaml (outlook pipeline — contact book YAML read/write; usually pre-installed with conda)
• Microsoft PowerPoint (optional — enables Vision AI for PPTX slide rendering)
• Microsoft Outlook (outlook pipeline — must be running; no OAuth/admin required)
• Anthropic API access (Claude Sonnet 4.6)
• Windows (uses Windows Registry for credential loading; Outlook COM requires Windows)

Note: paper_pipeline.py and ci_pipeline.py do not yet have dedicated README files — see inline --help for their CLI options.