AI Settings

AI Settings

Settings > AI Settings... configures the provider, the model, and how AI picks behave. Only the API key is stored encrypted; the other AI settings are stored as plain configuration values in the settings database.

Provider and model

Field	Description
Enable AI-assisted image selection	Master switch; without it no AI features are available
Provider	OpenRouter, OpenAI, Anthropic, Gemini, GLM, Local, or a custom endpoint
Base URL	Editable API base URL; preset per provider. Custom (OpenAI-compatible) has no preset and requires one
API Key	Provider key; stored encrypted and never logged. Required for OpenRouter, OpenAI, Anthropic, Gemini and GLM; optional for Local and custom endpoints
Model	Model identifier; use Test / Load Models to pick from the provider list. Enter exactly one ID — if you paste several (for example the whole recommended-models list), only the first one is kept and the app tells you when the field was adjusted
Test / Load Models	Verifies the connection and loads the model list from <Base URL>/models
Filter box	Narrows the loaded model list by text
Vision-capable only	Hides models that most likely cannot analyze images (default on)

A Get an OpenRouter API key hyperlink below the API Key field opens the OpenRouter key page in your browser.

Validation rules

When you save with AI Assist enabled:

• The Base URL must be an absolute URL if you entered one (for example https://openrouter.ai/api/v1 or http://localhost:11434/v1).
• Custom providers require both a Base URL and a Model.
• OpenRouter, OpenAI, Anthropic, Gemini, and GLM require an API key.
• Local never requires an API key.
• The Model field must contain a single model ID. Pasted multi-ID text (for example the whole recommended-models list) is reduced to the first ID, and the app shows a "Model Adjusted" message when it changes the field.

The loaded model list is cached per provider and lowercased base URL for 7 days in %LocalAppData%\FindRomCover\ai-models.json. If a connection test fails, the previously cached list stays loaded; after changing the base URL the list is empty until you click Test / Load Models.

Custom request parameters

The Custom request parameters box is an advanced, optional JSON object that is merged into every AI request body. Use it to tune values the app otherwise fixes (temperature, output tokens, penalties) or to pass provider-specific options.

Provider family	Example	Notes
OpenAI, OpenRouter, GLM, Local, custom OpenAI	{"temperature": 0.2, "top_p": 0.9}	Top-level chat-completions fields
Anthropic, custom Anthropic	{"max_tokens": 8192, "temperature": 0.3}	max_tokens is required by the Messages API and defaults to 4096
Gemini	{"generationConfig": {"temperature": 0.2, "topP": 0.9}}	Nested objects are merged, so maxOutputTokens keeps its default

Rules:

• The value must be a JSON object; arrays, strings, and numbers are rejected.
• Objects are merged recursively; arrays replace the default value.
• Set a field to null to remove it: {"temperature": null}.
• These keys are reserved and rejected on save because the app generates them: model, messages, contents, system, systemInstruction, and stream.
• OpenAI reasoning models (o1, o3, o4, gpt-5*) reject temperature and max_tokens; use max_completion_tokens for those models.
• Insert example fills the box with an example for the selected provider.
• Invalid JSON is rejected when saving. If a hand-edited value becomes invalid, it is ignored and logged instead of breaking AI requests.
• The JSON is stored as plain configuration, not encrypted. Do not put secrets in it.

For the exact fields accepted by each provider, see that provider's API documentation.

Limits & behavior

Setting	Default	Range	Description
Timeout (seconds)	90	10–300	How long to wait for the provider before giving up
Max candidates sent	6	1–20	How many top-ranked local images are sent to the model
Candidate similarity (%)	70	0–100	Only local images with filename similarity at or above this value are sent to the AI
Image max size (px)	512	128–2048	Images are downscaled to this maximum dimension before upload
Min cover width (px)	200	0–2048	Downloaded covers narrower than this are rejected as too low resolution (0 disables the check)
Auto-save threshold (%)	80	0–100	Minimum confidence required for automatic saving

Behavior toggles

Toggle	Default	Effect
Automatically save the AI pick when confidence is above the threshold	Off	The AI-selected image is saved and the game leaves the missing list without further clicks
Run AI Pick automatically when a game is selected	Off	Starts the AI analysis right after a local search on the Local Files tab or a Google API search on the Google API tab
Verify images with AI before renaming or saving	Off	Applies to the image folder watcher, manual local saves, manual Google API saves, and batch API downloads. Manual saves show a confirmation prompt when the AI considers the image a mismatch; Batch Fill rejects the download automatically and remembers it in the query history
Skip missing covers already queried in previous sessions	On	Batch Fill will not ask the AI twice for the same cover, even across restarts

Query history

The Clear query history button forgets every cover that was already queried. The number of remembered entries is shown next to the button. See Query History for retention details.

Tuning guidance

Goal	Suggested changes
Lower cost	Reduce Max candidates, raise Candidate similarity, keep the verdict cache enabled
Fewer wrong auto-saves	Raise Auto-save threshold to 90 or disable auto-save
Faster picks	Reduce Image max size to 384 and Max candidates to 4
More thorough picking	Raise Max candidates to 10–15 and lower Candidate similarity to 50
Slow local models	Raise Timeout to 180–300 seconds

Note: The Candidate similarity value is separate from the main window's similarity threshold. The main window threshold controls which local images are listed; the AI value controls which of those are sent to the model.

Where settings live

AI settings are stored in %LocalAppData%\FindRomCover\Settings.dat. Only the API key (AiApiKey) is AES-encrypted; the provider, base URL, model, custom request parameters (AiRequestParameters), timeouts, thresholds, and toggles are plain key/value rows. The provider model list lives in ai-models.json, AI verdicts in ai-cache.json, and skipped covers in QueryHistory.dat. See Data Storage.

Related pages

• Providers & Models
• Picking Covers
• Batch Fill
• Query History