AI reads. The catalog decides.

RepairLoop uses a server-side Gemini model for extraction, source-cited explanation, and AI-assisted repair reports. The configured model name is intentionally not published in this repository. AI is a helper, not the authority.

What AI does

• Reads a redacted battery report or label image and returns structured text fields for the user to review.
• Maps free-text symptoms to controlled tags, with a safety flag for damaged-battery descriptions.
• Extracts device/part/symptom facts from a shop customer request for a human-reviewed quote.
• Explains a catalog-selected part using only the cited catalog/source evidence, with a deterministic fallback.
• Writes an AI-assisted repair report from the confirmed fields, the optional redacted image, and the catalog evidence. The selected part and fitment remain catalog-authoritative.

What AI never does

• Choose compatibility or invent a part recommendation.
• Set price, stock, seller, shipping, or payment status.
• Create catalog rows or authorize/capture a payment.
• Override a no-match or safety stop.

Privacy and limits

Images, free-text symptom/shop text, and the case details used for a source-cited explanation or AI repair report are sent to Gemini after the user checks an AI-consent box; the free tier may use submitted content to improve Google products. Images are re-encoded server-side to strip EXIF/GPS/device metadata before the Gemini request, but the browser preview can retain the local file. Crop and redact serials, service tags, and personal data. Do not upload payment credentials or customer data. AI extraction can be wrong; the user confirms the fields, and manual entry remains the fallback. The catalog and cited sources remain authoritative. The API’s “grounded” explanation check verifies the cited URL, not the semantics of every sentence.

Read the implementation notes in AI_USAGE.md in the public repository, the privacy disclosure, and the terms.