feat(core): attempt graceful fastembed import error handling, add tod…

…o to try fastembed-web as fallback later (#1871)

This PR adds error handling for if the default embedder import fails. I
also added a TODO for trying to fallback to
[`fastembed-web`](https://www.npmjs.com/package/fastembed-web) later -
right now I can't try it cause our vercel deploys are failing for other
reasons

---------

Co-authored-by: Your Name <you@example.com>

.changeset/nasty-adults-push.md

@@ -0,0 +1,5 @@
+---
+'@mastra/core': patch
+---
+
+Add fastembed import error handling

packages/core/src/vector/fastembed.ts

@@ -15,20 +15,67 @@ function getModelCachePath() {
   return cachedPath;
 }
 
+function unbundleableImport(name: string) {
+  const nonStaticallyAnalyzableName = `${name}?d=${Date.now()}`; // +? date to prevent esbuild from seeing this as statically analyzable and bundling this as a regular import.
+  return import(nonStaticallyAnalyzableName.split(`?`)[0]!); // remove the date to prevent module not found errors in cloud
+}
+
+let useFastEmbedWebFallback = false;
+
 // Shared function to generate embeddings using fastembed
 async function generateEmbeddings(values: string[], modelType: 'BGESmallENV15' | 'BGEBaseENV15') {
   try {
-    // NOTE: This dynamic import janky dance is due to some runtimes not having support for the native bindings used by onnxruntime-node
-    // I made a fork of fastembed-js and published it as https://www.npmjs.com/package/fastembed-web
-    // The fork uses onnxruntime-web instead of the node package. Theoretically this fork has greater compatibility while also being less performant.
-    // If we need greater compat for this embedder we can do a dynamic import for fastembed-js and if the import fails we can do a second dynamic import for fastembed-web and try that
-    // I haven't done this yet because I got fastembed-js to work locally and in Mastra Cloud.
-    // If we need this to also work places like Netlify and Vercel and it's not working we will likely need to try using fastembed-web.
-    //
     // Dynamically import fastembed only when this function is called
     // this is to avoid importing fastembed in runtimes that don't support its native bindings
-    const fastEmbedImportPath = 'fastembed?d=' + Date.now(); // +? date to prevent esbuild from seeing this as statically analyzable and bundling this as a regular import.
-    const { EmbeddingModel, FlagEmbedding } = await import(fastEmbedImportPath.split(`?`)[0]!); // remove the date to prevent module not found errors in cloud
+    let mod: any;
+    const importErrors: Error[] = [];
+
+    if (useFastEmbedWebFallback) {
+      // TODO: see below
+      // mod = await unbundleableImport('fastembed-web');
+    } else {
+      try {
+        mod = await unbundleableImport('fastembed');
+      } catch (e) {
+        if (e instanceof Error) {
+          importErrors.push(e);
+        } else {
+          throw e;
+        }
+        // TODO: once vercel deploys are fixed, install fastembed-web and try it as a fallback
+        // try {
+        //   mod = await unbundleableImport('fastembed-web');
+        //   useFastEmbedWebFallback = true;
+        // } catch (e) {
+        //   if (e instanceof Error) {
+        //     importErrors.push(e);
+        //   } else {
+        //     throw e;
+        //   }
+        // }
+      }
+    }
+
+    if (!mod) {
+      throw new Error(`${importErrors.map(e => e.message).join(`\n`)}
+
+This runtime does not support fastembed-js, which is the default embedder in Mastra. 
+Scroll up to read import errors. These errors mean you can't use the default Mastra embedder on this hosting platform.
+You can either use Mastra Cloud which supports the default embedder, or you can configure an alternate provider.
+
+For example if you're using Memory:
+
+import { openai } from "@ai-sdk/openai";
+
+const memory = new Memory({
+  embedder: openai.embedding("text-embedding-3-small"), // <- doesn't have to be openai
+})
+
+Visit https://sdk.vercel.ai/docs/foundations/overview#embedding-models to find an alternate embedding provider
+`);
+    }
+
+    const { FlagEmbedding, EmbeddingModel } = mod;
 
     const model = await FlagEmbedding.init({
       model: EmbeddingModel[modelType],
@@ -42,7 +89,7 @@ async function generateEmbeddings(values: string[], modelType: 'BGESmallENV15' |
     for await (const result of embeddings) {
       // result is an array of embeddings, one for each text in the batch
       // We convert each Float32Array embedding to a regular number array
-      // @ts-ignore ? Date.now() import breaks types
+      // @ts-ignore unbundleableImport fn breaks types
       allResults.push(...result.map(embedding => Array.from(embedding)));
     }