Iris — Vision

Iris is Synth’s image and video understanding subsystem. When Synth receives a picture or a video, Iris describes its content so the reply can take the visual into account.

Iris is invoked automatically by the media dispatcher for image/* and video/* MIME types, so it does not register a normal model-facing action — the vision_describe capability is wired through the dispatcher. Its public API is await iris_plugin.describe_media(file_path, mime_type, prompt, engine_name, model). Results are stored in a content-addressed media cache to avoid re-analyzing the same file.

The active engine is selected with ACTIVE_IRIS_ENGINE, chosen from the Iris engine registry (core/iris_registry.py).

Configuration

Key

Purpose

ACTIVE_IRIS_ENGINE

Active vision engine (default selenium-llm-engine).