Ask the images.
Describe a street, or upload a photo. Every one of Amsterdam's 99,100 street views is matched on its image and an open model's observations.
Images load from the City's server (Gemeente Amsterdam, CC BY 4.0). Observations are model output (Qwen3.5-4B), not measurements. Search: EmbeddingGemma 2 on a CPU. Uploaded photos are not stored.
Nothing is generated: every answer is a real view.
Street views
Four per panorama of the City's 2025 drive, each read for nineteen features by an open model. How, in WP3 →
Image and words together
Each view's image and the model's observations are embedded together by EmbeddingGemma 2, as 256 numbers.
Closest views
Your words or photo are embedded the same way and compared with every view; the twelve closest are shown, with links to the City's images.
Similarity is not a measurement: a high score means the view resembles the description as the model saw it. Footways and park paths are rarely in the images, because the camera rides on the road.