AI ranking
Let a model re-rank candidates and write titles, hooks and hashtags.
AI ranking is optional. Without it, shardreel ranks moments from audio, visual and speech signals alone, and titles come from the words spoken near the peak. With it, a model sees the candidate windows and picks the best, tightens the cut, and writes a title, hook and hashtags.
If the model call fails for any reason, the run finishes with signal ranking and logs LLM step failed (...). A model problem never loses a run.
Providers
| Provider | Key | Endpoint |
|---|---|---|
| Gemini | GEMINI_API_KEY | Google's OpenAI-compatible endpoint |
| OpenRouter | OPENROUTER_API_KEY | https://openrouter.ai/api/v1 |
| LM Studio | none | http://localhost:1234/v1 |
| Ollama | none | http://localhost:11434/v1 |
Set a base URL override in Settings to use any other OpenAI-compatible server.
Set it up
Add a key in Settings, or in the env file.
Choose the provider in the job form under AI ranking. To make it the default, set Default provider in Settings and press Save.
Leave the model empty to use the provider default, or set a model name. On OpenRouter, use the full vendor and name form.
OpenRouter model names
OpenRouter names include the vendor, such as google/gemini-3.8-flash, which is also the built-in default. If you set your own model, use the full vendor and name form, or the call fails and the run falls back to signal ranking.
The provider is chosen in the UI. A key in the env file does not select it. A job started with the wrong provider reports set the Gemini API key in Settings, or the matching OpenRouter message.
Choosing a model
Each video is one request holding a few thousand tokens of text, plus a few more thousand if vision is on. Cost is small even for a premium model, so favour quality over the cheapest tier. A current Flash-class model works well: for example gemini-3.8-flash on Gemini directly, or google/gemini-3.8-flash on OpenRouter. Model names and prices change, so check your provider's list.
Vision
Turn on Vision to send three small frames of each candidate along with its speech. It helps for silent or low-commentary gameplay, where the audio alone says little. The model must accept images.
Niche context
The Niche context text tells the model what game and audience the clips are for. The default targets GTA 6 and gaming TikTok and Shorts. Edit it to be specific, for example the game mode and the audience age range. This tends to improve picks more than switching models.
What the model can and cannot change
The model returns a start and end for each pick, and shardreel checks them against guard rails:
- The start must fall between 10 seconds before the candidate's start and 1 second before the peak.
- The end must fall between 2.5 seconds after the peak and 10 seconds after the candidate's end.
- Anything outside those limits is reset to the candidate's own bound.
- Lengths are then clamped to your min and max.
Picks that overlap by more than 40 percent of the shorter clip are dropped, and the list is cut to your clip count.