1 files changed, 44 insertions, 0 deletions
diff --git a/toolkit/components/ml/docs/index.md b/toolkit/components/ml/docs/index.md
new file mode 100644
index 0000000000..1b2015456b
--- /dev/null
+++ b/toolkit/components/ml/docs/index.md
@@ -0,0 +1,44 @@
+# Machine Learning
+
+This component is an experimental machine learning local inference engine. Currently there is no inference engine actually integrated yet.
+
+Here is an example of the API:
+
+```js
+// The engine process manages the life cycle of the engine. It runs in its own process.
+// Models can consume large amounts of memory, and this helps encapsulate it at the
+// operating system level.
+const EngineProcess = ChromeUtils.importESModule("chrome://global/content/ml/EngineProcess.sys.mjs");
+
+// The MLEngineParent is a JSActor that can communicate with the engine process.
+const mlEngineParent = await EngineProcess.getMLEngineParent();
+
+
+/**
+ * When implementing a model, there should be a class that provides a `getModel` function
+ * that is responsible for providing the `ArrayBuffer` of the model. Typically this
+ * download is managed by RemoteSettings.
+ */
+class SummarizerModel {
+  /**
+   * @returns {ArrayBuffer}
+   */
+  static getModel() { ... }
+}
+
+// An engine can be created using a unique name for the engine, and the function
+// to get the model. This class handles the life cycle of the engine.
+const summarizer = mlEngineParent.getEngine(
+  "summarizer",
+  SummarizerModel.getModel
+);
+
+// In order to run the model, use the `run` method. This will initiate the engine if
+// it is needed, and return the result. The messaging to the engine process happens
+// through a MessagePort.
+const result = await summarizer.run("A sentence that can be summarized.")
+
+// The engine can be explicitly terminated, or it will be destroyed through an idle
+// timeout when not in use, as the memory requirements for models can be quite large.
+summarizer.terminate();
+```