add: onnxruntime instructions

Author: kallebysantos

ext/ai/README.md

@@ -5,9 +5,6 @@ features for the `Supabase.ai` namespace.
 
 ## Model Execution Engine
 
-`Supabase.ai` uses [onnxruntime](https://onnxruntime.ai/) as internal model
-execution engine, backend by [ort pyke](https://ort.pyke.io/) rust bindings.
-
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" srcset="/assets/docs/ai/onnx-backend-dark.svg">
@@ -16,10 +13,66 @@ execution engine, backend by [ort pyke](https://ort.pyke.io/) rust bindings.
   </picture>
 </p>
 
+`Supabase.ai` uses [onnxruntime](https://onnxruntime.ai/) as internal model
+execution engine, backend by [ort pyke](https://ort.pyke.io/) rust bindings.
+
 Following there's specific documentation for both "lands":
 
 <details>
   <summary>Javascript/Frontend</summary>
+
+The **onnxruntime** API is available from `globalThis` and shares similar specs of [onnxruntime-common](https://github.com/microsoft/onnxruntime/tree/main/js/common).
+
+The available items are:
+
+- `Tensor`: represent a basic tensor with specified dimensions and data type. -- "The AI input/output"
+- `InferenceSession`: represent the inner model session. -- "The AI model itself"
+
+### Usage
+
+It can be used from the exported `globalThis[Symbol.for("onnxruntime")]` --
+but manipulating it directly is not trivial, so in the future you may use the [Inference API #501](https://github.com/supabase/edge-runtime/pull/501) for a more user friendly API.
+
+```typescript
+const { InferenceSession, Tensor } = globalThis[Symbol.for("onnxruntime")];
+
+// 'create()' supports an url string buffer or the binary data
+const modelUrlBuffer = new TextEncoder().encode("https://huggingface.co/Supabase/gte-small/resolve/main/onnx/model_quantized.onnx");
+const session = await InferenceSession.create(modelUrlBuffer);
+
+// Example only, in real 'feature-extraction' tensors must be created from the tokenizer step.
+const inputs = {
+   input_ids: new Tensor('float32', [1, 2, 3...], [1, 384]),
+   attention_mask: new Tensor('float32', [...], [1, 384]),
+   token_types_ids: new Tensor('float32', [...], [1, 384])
+};
+
+const { last_hidden_state } = await session.run(inputs);
+console.log(last_hidden_state);
+```
+
+### Third party libs
+
+Originaly this backend was created to implicit integrate with [transformers.js](https://github.com/huggingface/transformers.js/). This way users can still consuming a high-level lib at same time they benefits of all Supabase's Model Execution Engine features, like model optimization and caching. For further information pleas check the [PR #436](https://github.com/supabase/edge-runtime/pull/436)
+
+> [!WARNING]
+> At this moment users need to explicit target `device: 'auto'` to enable the platform compatibility.
+
+```typescript
+import { env, pipeline } from 'https://cdn.jsdelivr.net/npm/@huggingface/[email protected]';
+
+// Broswer cache is now supported for `onnx` models
+env.useBrowserCache = true;
+env.allowLocalModels = false;
+
+const pipe = await pipeline('feature-extraction', 'supabase/gte-small', { device: 'auto' });
+
+const output = await pipe("This embed will be generated from rust land", {
+  pooling: 'mean',
+  normalize: true
+});
+```
+
 </details>
 
 <details>