Ofline AI features for FoliCon

[DineshSolanki]

1. AI-Powered Folder Name Normalizer / Query Extractor

Addresses: Issues #186, #185, user pain with unicode chars/year suffixes/Plex naming

Currently, FoliCon strips some characters but uses the raw folder name as a search query. An LLM or a small fine-tuned NLP model could:

• Parse Chernobyl (2019) {tmdb-87108} → extract title, year, and TMDB ID with 100% confidence
• Normalize Star Wars꞉ The Clone Wars → Star Wars: The Clone Wars
• Handle patterns like The.Dark.Knight.2008.BluRay, Avatar [2009], Mirzapur S01
• Infer the media type (movie vs TV vs anime vs game) from folder name patterns alone

Implementation idea: A small ICandidateExtractor service backed by a local NLP model (e.g., Microsoft.ML, OnnxRuntime) or an optional cloud LLM call to parse folder names before TMDB query.

---

2. Smart Auto-Match / Confidence Scoring

Addresses: The core "skip ambiguous" problem — users with 220+ folders can't review every match

After getting TMDB results, apply a confidence score to auto-accept the best match instead of always prompting. Signals:

• Year in folder name vs. TMDB release year match
• Folder name similarity score (fuzzy/embedding match) to result title
• TMDB popularity score
• Media type alignment (TV folder structure vs. movie)

FolderName: "Breaking Bad (2008)"
Match: "Breaking Bad" (2008, TV) → Confidence: 97% → Auto-accept
Match: "Breaking Bad" (2021, Short) → Confidence: 31% → Prompt user

Implementation: Could use FuzzySharp for string similarity + rule-based year/type signals, no cloud required.

---

3. AI-Based Media Type Detection

Addresses: The Auto search mode being unreliable and prompting on obvious cases

Currently the user manually selects Movie/TV/Game/Automatic. An ML classifier trained on folder naming conventions could predict the media type automatically:

• Season 1, S01E01, Complete Series → TV Show
• [1080p], (2022), no season markers → Movie
• Game keywords, platform tags → Game
• Japanese names with episode counts → Anime

---

4. Generative AI Fallback — AI-Generated Placeholder Icons

Addresses: Cases with zero TMDB/DeviantArt results, obscure anime/games

When no artwork is found anywhere, use a generative model to create a styled placeholder:

• Call a local diffusion model or an optional cloud API (DALL·E, Stability AI)
• Generate a stylized icon from the title text using the app's overlay style
• Could use a simple text-to-icon pipeline using WPF rendering + font glyphs as a free, offline fallback

---

UX & Workflow Enhancements

1. Smart Subfolder Pattern Suggestions

Addresses: Issue #221 — users struggle writing exclusion regex

Replace the raw regex input with an AI-assisted UI:

• User types "exclude subtitle folders" → AI suggests ^(subs?|subtitles?)$
• Show plain-English descriptions of existing regex patterns
• "I want to skip folders named Season X" → ^Season \d+$

Implementation: Simple LLM call (Copilot Chat plugin, OpenAI, or local Ollama).

---

2. Natural Language Query Interface

A power-user feature: Allow users to describe what to process in plain text:

"Process all anime folders that don't have icons yet, skip seasons, use Liaher style"

Parse this into the appropriate FoliCon settings (SearchMode = Anime, SubfolderExclusion = Season*, IconStyle = Liaher, SkipIfIconExists = true).

---