Linguistic Intelligence Architecture

Overview: Local & Private NLP #

Unlike cloud-based AI tools that upload user manuscripts to third-party servers, FleshNote's linguistic intelligence operates 100% locally on the author's machine. The architecture combines three core libraries:

• spaCy / HuSpaCy: Powers Named Entity Recognition (NER), Part-of-Speech (POS) tagging, and Dependency Parsing across English, Hungarian, and Polish.
• NLTK WordNet: Provides rich synonym lookup and Open Multilingual WordNet (OMW) synsets.
• Hunspell (via phunspell): Provides local spellchecking with automatic whitelisting of worldbuilding entities.

Dynamic AppData Model Loader (`nlp_manager.py`) #

Bundling neural language models would bloat the desktop installer beyond 1.2 GB. Instead, FleshNote ships as a lightweight ~80 MB binary and downloads required language models on demand into the user's OS profile:

Execution Context Branching

• Development Mode: Executes an isolated pip install --target {models_dir} subprocess.
• PyInstaller Frozen Production: Because pip is not available inside the compiled binary, nlp_manager.py directly downloads the .whl or .tar.gz archive via Python's urllib, extracts it in a temporary folder, and builds the package structure natively in AppData/Models while streaming DOWNLOAD_PROGRESS: {pct} events to Electron.

NLP Subsystem Deep Dives #

Explore specific components of the NLP and linguistic intelligence architecture: