Metadata repair

When a torrent's tracker title contains unusual formatting, non-standard naming, or non-Latin characters, the built-in release parser may produce low-quality metadata. The AI metadata repair feature sends these uncertain parses to a language model for re-interpretation and merges the corrected fields back.

When repair is triggered

Every torrent goes through the release parser during ingest. After parsing, a confidence heuristic checks the result:

• Title is too short, too long, or contains non-Latin characters → low confidence
• Title ends with the release group name (likely duplicated) → low confidence
• Type is missing or Unknown → low confidence
• Movie type but no plausible year → low confidence
• TV type but no season, episode, or date → low confidence

If confidence is low, the torrent is queued on the AI Repair worker. If it's ok, the parsed data is accepted as-is and the pipeline continues.

How repair works

1. The low-confidence release data is stripped of null or empty fields and sent as JSON to the LLM.
2. The LLM returns a corrected JSON payload matching the release schema (title, group, year, season, episode, resolution, flags, type, etc.).
3. The original and repaired data are merged — the AI's output overrides the original for any populated field.
4. The torrent is updated (isDirty: false, repairedAt timestamp set) and re-indexed in Meilisearch.
5. The torrent is re-queued for enrichment so metadata providers see the corrected title.

Providers

Configure the provider, model, and other settings in Dashboard → Settings under the AI Repair section.

Configuration

OpenRouter

Set OPENROUTER_API_KEY in your .env file, then choose a model from the settings page:

OPENROUTER_API_KEY=sk-or-v1-xxxx

Suggested models: deepseek/deepseek-v4-flash, google/gemini-2.5-flash, qwen/qwen3.5-flash-02-23, anthropic/claude-3.5-haiku.

Ollama

Point Minato at your Ollama instance:

# Default URL (Ollama in the same Docker network)
http://ollama:11434

Make sure the model you select is pulled on your Ollama instance first. Use the Test button in settings to verify connectivity and check the Ollama version.

Suggested models: qwen3.5:4b, qwen3.5:9b, ministral-3:8b, gemma3:4b.

Reasoning mode

A toggle in the AI Repair settings controls whether the model uses extended reasoning (thinking tokens):

• Ollama — passes think: true to the Ollama API when enabled
• OpenRouter — explicitly sets reasoning: { effort: "none" } when disabled, uses model defaults when enabled

Reasoning can improve accuracy on heavily mangled titles at the cost of slower response times.

Worker

The AI Repair worker runs with a concurrency of 5 parallel jobs. Repaired torrents are re-queued for enrichment with a 1-second delay to avoid overwhelming upstream metadata providers.

The AI Repair queue is visible in Dashboard → Queues alongside Ingest, Enrich, and Housekeeper.