See your library through new eyes.

OpenAI (GPT-4o family + text-embedding-3), Anthropic Claude (vision), Google Gemini (vision + text-embedding-004), Voyage AI (multimodal embeddings), and Ollama for self-hosted LLaVA / nomic-embed-text. Vision and embedding are configured independently, so any vision provider can be paired with any embedding provider.

A text embedding turns the AI-generated description into a vector. A multimodal embedding (Voyage voyage-multimodal-3) embeds the raw image bytes directly, capturing texture, composition, and visual style the description would lose. Spectacles uses multimodal automatically when the configured provider supports it.

No. Spectacles ships with a pure-PHP cosine-similarity scan backend that works on any database Craft supports — fine for libraries up to a few thousand images. If you are on Postgres with the vector extension installed, Spectacles auto-detects it and uses pgvector for production scale.

No. The public search endpoint runs vision + embedding on the uploaded image in memory and discards it as soon as the search returns. Only your own asset library is persisted to the metadata table.

Indexing an image costs one vision call (a few cents on gpt-4o-mini) plus one embedding call (fractions of a cent). Visitor searches cost the same per query. Use the “Cheapest hosted” preset (Gemini) for budget deployments, or “Self-hosted” with Ollama for zero API cost.

Different models produce vectors of different dimensions, so Spectacles only compares vectors of matching shape. After switching, click Re-index all images in the settings screen to regenerate the embeddings under the new provider.

Yes. The settings screen lets you check the specific volumes Spectacles should auto-analyze. Leave it empty to analyze every image asset.

By default POST /spectacles/search is open so visitors can use it on the front end, but you can disable public search entirely in the settings. A configurable max upload size prevents abuse.

Yes. The Spectacles panel in the asset edit sidebar shows an Analyze now button for any image that has not been analyzed yet, and re-running analysis is a single queue job push.

In a dedicated spectacles_imagemetadata table that stores the description, tags, objects, colors, the raw provider response, the embedding vector as JSON, and which provider/model produced it.