From b3df311be7ab6f98ec07d26282645e553eb6ccfb Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Azade=20=F0=9F=90=90?= Date: Fri, 30 Jan 2026 16:06:27 +0000 Subject: [PATCH] feat(memory): use QAT embedding model as default Switch default local embedding model from embeddinggemma-300M to embeddinggemma-300m-qat (Quantization Aware Training variant). QAT models are trained with quantization in mind, which typically yields better quality embeddings compared to post-training quantization. --- src/memory/embeddings.ts | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/src/memory/embeddings.ts b/src/memory/embeddings.ts index 993fe8124..5b04d7f4f 100644 --- a/src/memory/embeddings.ts +++ b/src/memory/embeddings.ts @@ -43,7 +43,8 @@ export type EmbeddingProviderOptions = { }; }; -const DEFAULT_LOCAL_MODEL = "hf:ggml-org/embeddinggemma-300M-GGUF/embeddinggemma-300M-Q8_0.gguf"; +const DEFAULT_LOCAL_MODEL = + "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf"; function canAutoSelectLocal(options: EmbeddingProviderOptions): boolean { const modelPath = options.local?.modelPath?.trim();