[Plugin] LLM_AI — AI-Powered TV Recommendations & Guide Assistant

[reneboulard 38]

# LLM_AI — Emby LLM Recommendations Plugin (latest version v1.13.4.4)

**Get the plugin here:** [https://github.com/reneboulard/LLM_AI]

** plugin forum page: ** [llm-ai-integration]

## Simply put: the assistant that knows your TV

### The problem

You have an Emby server that records TV. Every night, the same question: *"What should I watch or record tonight?"* Answering it yourself means scrolling through the TV guide (hundreds of shows), remembering what you've already seen, checking what's already recorded, and deciding if a movie is actually worth your time... It's tedious manual sorting — and exactly the kind of task an AI excels at.

### The idea

**LLM_AI** is an assistant built into Emby that **reads the TV guide for you** and tells you what to watch or record. It isn't just a bot making random suggestions: it queries your server (What's on tonight? What do I usually like? What do I already have?), cross-references this with external sources (movie databases, the web, new-release lists), and then **analyzes and proposes** — always providing a clear explanation in plain language.

### Key Features

- **A "Recommendations" page**: Suggested series and movies to record soon, each accompanied by a specific reason.
- **"Watch Tonight"**: A personalized selection for each household member — based on your tastes, shows you started but never finished, or recordings piling up ("You have 6 episodes of this series: time to start!").
- **Tailored Delivery**: Whether it's a playlist, a collection, a tag, favorites, or a directly scheduled recording — each option is handled independently. Everything is tidied up every night. **The plugin never deletes any content.**

### Continuous Learning

- **Learning from feedback**: Once a week, it compares its recommendations with what you actually watched and adjusts its future suggestions.
- **Strategic Memory**: Recommendations, watch events, and "abandoned" shows are journaled. The AI regularly updates a "memory card" of its strategy, which it reviews before making new proposals.
- **You remain in control**: Every suggestion can be rejected ("Forget"), and sensitive actions are disabled by default — you must explicitly opt-in.

### Extra Capabilities

- **AI Chat**: Ask questions in natural language (e.g., "What documentaries are airing this week?").
- **Server Health Reports**: On demand, it inspects Emby (disk space, logs, performance) and generates a report with actionable advice.
- **Metadata Repair**: When a recorded show lacks proper metadata (title not found in catalogs), the plugin searches the web to fill in the missing information, mimicking a manual search.

### The Foundation: Clean Data

The assistant **can only evaluate the data it is given**. However, TV guides and libraries often use different terminology: the guide might say "Kids" while your library says "Enfant"; a rating might be "PG-13" in one place and "14A" in another. For an AI comparing lists, these discrepancies lead to less accurate results.

To solve this, two upstream cleanup plugins are **highly recommended**:

- **GenreCleaner** — Standardizes genre vocabulary across the entire library.
- **Classification Mapper** — Unifies content rating lists.

When installed, LLM_AI automatically leverages their work, translates the TV guide into the same vocabulary, and can even propose missing translations for you to validate with one click. **Clean metadata upstream, relevant recommendations downstream.**

### Accessible to Everyone (No Powerhouse Required)

Many assume AI requires massive hardware. That's not the case here:

- **Free Cloud Models** — Using a **free Ollama account**, you can use base models hosted on their servers. This is the easiest path: no software to install, no complex configuration, and it's perfectly sufficient for the plugin's needs. This is ideal for low-power devices; even on a **Raspberry Pi 4**, the heavy lifting is delegated to the cloud.
- **Local Models (Optional)** — Running the AI **locally** is the fully private option (nothing leaves your network), but it requires specific hardware: **a GPU with at least 8 GB of VRAM**. Below this, the model will run, but responses will be very slow.
- **Hybrid Mode** — The plugin can handle multiple services in parallel. It can use your local model when available and automatically switch to the cloud model as a backup if the local one is unavailable.

In short: the cost can be **zero dollars**. Depending on your hardware, the processing happens either at home (8 GB+ GPU) or for free in the Ollama cloud.

### In a nutshell

**It's a digital butler for your TV: it monitors the guide, learns your tastes, tells you what to watch or record, and cleans up your metadata — all while ensuring it never deletes your content or makes a sensitive decision without your approval.**