Digitalizer: Offline-First Multimodal OCR & Semantic Knowledge Engine

Offline-first visual document extraction engine bridging local multimodal vision models, two-tier persistent disk caching, and hierarchical Notion knowledge synchronization.

---

📌 Executive Summary

Digitizing dense technical manuscripts, mathematical lecture notes, and multi-column research papers typically imposes an unacceptable dilemma: surrender confidential documents to cloud OCR APIs at compounding per-page costs, or accept legacy OCR engines (e.g., Tesseract) that mutilate mathematical notation and structural layout.

Digitalizer resolves this operational bottleneck with an offline-first multimodal pipeline executing 100% locally via Ollama.

High-resolution PDF rasterization feeds local vision models (glm-4v, qwen2.5-vl), isolated behind a two-tier persistent caching architecture that eliminates redundant GPU compute and ingests structured AST blocks directly into Notion.

---

🏛️ Pipeline Architecture

                       DIGITALIZER ENGINE
                               │
               Dense Technical Document (PDF)
                               │
                               ▼
    ┌─────────────────────────────────────────────────────┐
    │          PHASE 1: LOCAL INGESTION & VISION OCR      │
    │  PyMuPDF High-Res Rendering (300 DPI Pixmaps)       │
    │  Local Multimodal Inference via Ollama (GLM-4V)     │
    │  Tier 1 Storage: cache_ocr_grezzo/ (Raw Transcripts)│
    └──────────────────────────┬──────────────────────────┘
                               │
                               ▼
    ┌─────────────────────────────────────────────────────┐
    │     PHASE 2: NORMALIZATION & SEMANTIC CLUSTERING    │
    │  Formula-Preserving Heuristic Case Normalization    │
    │  Regex Lecture & Chapter Boundary Identification    │
    │  Semantic Topic Map via Local Qwen 2.5 (14B)        │
    │  Tier 2 Storage: cache_ocr_revisionato/ (Clean AST) │
    └──────────────────────────┬──────────────────────────┘
                               │
                               ▼
    ┌─────────────────────────────────────────────────────┐
    │     PHASE 3: AST BLOCK PARSING & NOTION SYNC        │
    │  Native Markdown-to-Notion AST Block Generator      │
    │  Payload Chunking (Max 100 Blocks / 2000 Chars)     │
    │  Exponential Backoff Retries with Jitter            │
    └──────────────────────────┬──────────────────────────┘
                               │
                               ▼
    ┌─────────────────────────────────────────────────────┐
    │     PHASE 4: ACADEMIC NOTE ENRICHMENT (OPTIONAL)    │
    │  Professor Prompt Engine (Gemini / Local LLM)       │
    │  Acronym Expansion, Formula LaTeX & Key Takeaways   │
    │  Atomic Block Stream Replacement in Knowledge Base  │
    └─────────────────────────────────────────────────────┘

The offline-first foundation guarantees zero API expenditure during intensive raster extraction while ensuring complete data confidentiality over unpublished intellectual property.

---

🔬 Core Technical Modules

1. Multimodal Vision OCR & Caching Core (ocr_pipeline.py)

• High-Fidelity Rasterization: Employs PyMuPDF to extract raw page vectors and render them as optimized pixmaps, preserving high-frequency text boundaries and dense diagrams.
• Two-Tier Disk Caching: Isolates raw transcriptions in cache_ocr_grezzo/ before feeding them into normalization. Polished, structured outputs are committed to cache_ocr_revisionato/, preventing duplicate vision inference on unchanged source documents.
• Structural Case Normalization: Implements specialized heuristics (_correct_case_python) that correct OCR capitalization artifacts without altering mathematical notation, variable names, or programmatic code snippets.
• Semantic Topic Clustering: Deploys qwen2.5:14b locally to infer macro-thematic boundaries across fragmented lecture slides, grouping loose pages into coherent hierarchical chapters.

2. Native AST Markdown-to-Notion Synchronizer (ocr_pipeline.py)

• Block-Level Translation: Recursively converts raw Markdown into native Notion block schemas (heading_1, heading_2, heading_3, code, callout, bulleted_list_item, equation).
• Hard Limit Chunking: Enforces Notion API specifications, dynamically splitting text buffers exceeding 2,000 characters and batching payloads into strict 100-block atomic operations.
• Fault-Tolerant Transport: Implements exponential backoff with randomized jitter (@retry(retries=3, delay=2.0)), defending ingestion jobs against transient network drops and HTTP 429 rate limits.

3. Academic Knowledge Enrichment Engine (enrich_notes.py)

• Professor-Level Structuring: Connects to high-speed LLM inference engines (Google Gemini or local fallbacks) to expand cryptic lecture shorthand into rigorous academic prose.
• LaTeX Normalization: Detects raw mathematical expressions and converts them into standardized LaTeX equations for native display.
• Atomic Block Replacement: Flushes preliminary OCR placeholder blocks (clear_blocks) and streams enriched conceptual notes with zero database corruption or orphan records.

---

📊 Architectural & Economic Trade-Off Matrix

Capability	Digitalizer (Local Engine)	Cloud Vision APIs (AWS / GCP)	Legacy Rule-Based (Tesseract)
Marginal Compute Cost	$0.00 (Local Hardware)	$1.50–$2.50 per 1k pages	$0.00
Confidentiality / IP Defense	Zero Data Leakage (Air-Gapped)	External cloud transit	Zero Data Leakage
Mathematical / LaTeX Retention	High (Multimodal Context)	Medium (Frequent formatting loss)	Poor (Unreadable character noise)
Topic Clustering & Segmentation	Automated via Local LLM	Manual post-processing required	None
State Caching & Resumability	Native Two-Tier Filesystem	Requires custom external cache	None

---

🛠️ Environment & Production Quickstart

1. Prerequisites & Dependencies

Ensure Python 3.10+ and Ollama are installed and running locally:

# Clone the repository
git clone https://github.com/FRA-0023/Digitalizer.git
cd Digitalizer

# Install required Python packages
pip install pymupdf ollama requests python-dotenv rich google-genai

2. Local Vision & Language Models

Pull the recommended vision and language models via Ollama:

# Pull multimodal vision model
ollama pull glm-4v

# Pull semantic clustering model
ollama pull qwen2.5:14b

3. Environment Configuration

Configure credentials in .env (referencing .gitignore):

NOTION_TOKEN="your_notion_integration_token"
NOTION_DATABASE_ID="your_notion_database_id"
GEMINI_API_KEY="your_gemini_api_key" # Optional, used exclusively for enrich_notes.py

4. End-to-End Execution

Run document ingestion or trigger academic knowledge enrichment:

# Execute visual OCR extraction, two-tier caching, and Notion upload
python ocr_pipeline.py

# Enrich existing Notion notes with academic definitions and LaTeX equations
python enrich_notes.py

---

Author: Francesco Colombini
GitHub Profile · LinkedIn