AI Analysis

AI Analysis lets you ask questions about your library in plain language — “is there a pattern between resolution and video codec?”, “what have I never watched?”, “what are my ten largest movies?” — and get answers grounded in your own data.

The assistant is read-only. It can inspect and summarize your library through a fixed set of analysis tools, but it can never change, delete, download, or act on anything.

How it works

When you ask a question, the model translates it into calls to a small set of validated, read-only tools (the same query and aggregation engine that powers the dashboard and the Query workspace). Those tools run against your database, and the model narrates the actual numbers they return — it does not invent figures or answer from memory.

• Grounded — every quantitative claim comes from a tool result. Charts and tables shown beneath an answer are the exact data the answer is based on.
• Bounded — the model can only call the analysis tools listed below. It cannot run arbitrary SQL, and there are no tools that modify anything.
• Local-scope — the assistant only knows about your library and your servers’ watch history. It knows nothing about the outside world: global popularity, streaming-service trends, ratings sites, or release news. “What’s popular right now” means most-played on your servers recently, not globally.

Setup

1. Open Settings → AI.

2. Choose a Provider:

   • OpenAI-compatible — works with OpenAI and any endpoint that speaks the OpenAI chat-completions API: Ollama, LM Studio, OpenRouter, Groq, vLLM, LocalAI, and more.
   • Anthropic (Claude) — the native Claude Messages API.

3. Set the API base URL (optional). Leave blank to use the provider default. For a local model, point it at that model’s endpoint — for example http://localhost:11434/v1 for Ollama or http://localhost:1234/v1 for LM Studio.

4. Enter an API key if your provider needs one. Local models usually don’t. The key is stored on your server and masked in every response — it is never shown again.

5. Enter the Model identifier your provider expects (e.g. gpt-4o-mini, claude-sonnet-5, llama3.1). The model must support tool / function calling.

6. Click Test connection, then Save, and turn on Enable AI assistant.

Once enabled, open Tools → AI Analysis to start asking questions.

Tip

Prefer to keep everything on your own hardware? Run a local model with Ollama or LM Studio, pick OpenAI-compatible, and point the base URL at it. No data leaves your machine. Pick a model that supports tool calling (e.g. a recent Llama, Qwen, or Mistral instruct model).

What you can ask

The assistant is strongest at understanding, auditing, and cleaning up your library:

• Overview & composition — “give me an overview”, “what’s the genre breakdown”, “how much space do my 4K movies take”.
• Technical quality & hygiene — “is there a pattern between resolution and video codec”, “how much is still H.264”, “what percentage is HDR”.
• Storage & cleanup — “my 20 largest files”, “which genres use the most space”, “space used by unwatched titles older than two years”.
• Watch behavior — “most popular shows right now”, “what have I never watched”, “not played in a year”, “who watches the most”, “which devices are used most”.
• Anomalies — “duplicate titles across servers”, “items missing external IDs”.
• Reporting — “summarize my library”, “how has it grown over the past year”.

Tools the assistant can use

Tool	What it does
get_library_overview	Counts, storage, top resolutions/codecs/genres, most-played titles (all-time)
get_breakdown	Distribution across one dimension (resolution, codec, genre, year, …)
get_cross_tab	Relationship between two dimensions (e.g. resolution × codec)
get_timeline	How the library changed over time, optionally split by a dimension
search_media	Find and rank specific items by metadata filters
get_watch_trends	Most-played titles within a recent rolling window (“popular right now”)
get_watch_leaderboard	Rank users / devices / platforms by recent plays

Privacy & security

• The assistant is read-only — there are no tools that delete, unmonitor, download, or otherwise change anything, so it cannot take a destructive action even if asked.
• The tools never run model-generated SQL. They take structured, validated arguments and run the same safe queries the rest of the app uses.
• Your library metadata (titles, technical details, watch history) is sent to the provider you configure. Use a local model if you’d rather keep everything on your own hardware.
• The API key is stored on your server and masked in all API responses.
• The chat endpoint is rate-limited and requires an authenticated session.

Caution

The assistant reads your own library and your servers’ watch history — not the outside world. It also can’t do everything: quality-upgrade questions that depend on Sonarr/Radarr metadata, true personalized recommendations, and global/streaming-service trends are out of scope in this version. Double-check anything important before acting on it.