Spectacles

Spectacles analyzes your asset library with a vision model, stores structured metadata and a similarity vector for each image, and exposes endpoints + Twig helpers for finding similar images. Visitors can upload an image and get back the closest matches from your library.

Supported providers

Vision and embedding are configured independently — pick whichever combination suits your budget and quality bar.

Provider	Vision	Text embedding	Image embedding	Notes
OpenAI	gpt-4o, gpt-4o-mini, gpt-4.1	text-embedding-3-small / -large	—	Strong default. JSON mode is reliable.
Anthropic	Claude 4.x (Opus / Sonnet / Haiku)	—	—	Highest-quality descriptions. Pair with another embedder.
Google Gemini	gemini-2.5-flash, gemini-2.5-pro	text-embedding-004	—	Fast and inexpensive.
Voyage AI	—	voyage-multimodal-3	voyage-multimodal-3	Best similarity results — embeds the image directly, no description-loss.
Ollama (local)	llava, llama3.2-vision, bakllava	nomic-embed-text, mxbai-embed-large	—	Self-hosted, no API costs, slower.

Recommended pairings:

Best quality: Anthropic Claude (vision) + Voyage voyage-multimodal-3 (embedding)
Best price/perf: OpenAI gpt-4o-mini + OpenAI text-embedding-3-small
Self-hosted: Ollama llava + Ollama nomic-embed-text
Strict similarity, weak descriptions OK: any vision provider + Voyage multimodal

When the embedding provider is multimodal-capable (currently only Voyage), Spectacles will embed the image bytes directly instead of embedding the text description — this captures visual features the description would lose.