Merge pull request #1561 from segmentationf4u1t/fix/windows-segfault-minimal

fix: prevent Windows ORT preload crash

bun.lock

@@ -103,6 +103,7 @@
       "dependencies": {
         "@oh-my-pi/pi-ai": "catalog:",
         "fastembed": "catalog:",
+        "onnxruntime-node": "catalog:",
       },
       "devDependencies": {
         "@types/bun": "catalog:",
@@ -278,6 +279,7 @@
     "lucide-react": "^1.16.0",
     "marked": "^18.0.4",
     "markit-ai": "0.5.3",
+    "onnxruntime-node": "1.24.3",
     "openai": "^6.39.0",
     "partial-json": "^0.1.7",
     "postcss": "^8.5.15",

package.json

@@ -58,6 +58,7 @@
       "lucide-react": "^1.16.0",
       "marked": "^18.0.4",
       "markit-ai": "0.5.3",
+      "onnxruntime-node": "1.24.3",
       "openai": "^6.39.0",
       "partial-json": "^0.1.7",
       "postcss": "^8.5.15",

packages/mnemosyne/CHANGELOG.md

@@ -4,6 +4,7 @@
 
 ### Fixed
 
+- Fixed Windows startup crashes by keeping fastembed's older ONNX Runtime binding lazy until local embeddings are used.
 - Fixed a segfault at startup from eagerly loading fastembed: importing the embeddings module pulled in `fastembed`, which eagerly loads the `onnxruntime-node` native addon. The import is now deferred until a local fastembed model is actually initialized, so API-model, disabled-embeddings, and test runtimes never load the native addon.
 
 ## [15.6.0] - 2026-05-30

packages/mnemosyne/package.json

@@ -40,7 +40,8 @@
 	},
 	"dependencies": {
 		"@oh-my-pi/pi-ai": "catalog:",
-		"fastembed": "catalog:"
+		"fastembed": "catalog:",
+		"onnxruntime-node": "catalog:"
 	},
 	"devDependencies": {
 		"@types/bun": "catalog:"

packages/mnemosyne/src/core/embeddings.ts

@@ -1,5 +1,6 @@
 import { mkdirSync } from "node:fs";
-import type { EmbeddingModel } from "fastembed";
+import { createRequire } from "node:module";
+import type { EmbeddingModel, FlagEmbedding } from "fastembed";
 import { getMnemosyneRuntimeOptions, resolveEmbeddingProvider } from "./runtime-options";
 
 export type Vector = number[];
@@ -24,8 +25,14 @@ type LocalModelInitOptions = {
 };
 type LocalModelInitializer = (options: LocalModelInitOptions) => Promise<LocalEmbeddingModel>;
 
+interface FastembedRuntime {
+	EmbeddingModel: typeof EmbeddingModel;
+	FlagEmbedding: typeof FlagEmbedding;
+}
+
 const FASTEMBED_CACHE_DIR = `${process.env.HOME ?? ""}/.hermes/cache/fastembed`;
 const QUERY_CACHE_MAX = 512;
+const sourceRequire = createRequire(import.meta.url);
 
 let providerOverride: EmbeddingProvider | null = null;
 let localModelPromise: Promise<LocalEmbeddingModel> | null = null;
@@ -33,12 +40,14 @@ let localModelInitializer: LocalModelInitializer = defaultLocalModelInitializer;
 let apiCallCount = 0;
 const queryCache = new Map<string, Vector>();
 
-async function defaultLocalModelInitializer(options: LocalModelInitOptions): Promise<LocalEmbeddingModel> {
-	// Lazily pull in fastembed here — its module eagerly imports `onnxruntime-node`,
-	// whose native addon load segfaults in some runtimes. Deferring to the actual
-	// model init keeps API-model, disabled, and test runtimes from ever loading it.
-	const { FlagEmbedding } = await import("fastembed");
-	return FlagEmbedding.init(options) as Promise<LocalEmbeddingModel>;
+function loadFastembedRuntime(): FastembedRuntime {
+	// Preload ORT 1.24 before fastembed's ORT 1.21 binding to avoid Windows DLL reuse crashes.
+	sourceRequire("onnxruntime-node");
+	return sourceRequire("fastembed") as FastembedRuntime;
+}
+
+function defaultLocalModelInitializer(options: LocalModelInitOptions): Promise<LocalEmbeddingModel> {
+	return loadFastembedRuntime().FlagEmbedding.init(options);
 }
 
 function activeEmbeddingOptions() {