From 02ba4b4fc2cc4fea9729fe124250faa7056f2f9e Mon Sep 17 00:00:00 2001
From: Michael Gandal <mgandal@gmail.com>
Date: Wed, 22 Jul 2026 15:38:30 -0400
Subject: [PATCH] dims: thread Matryoshka dimensions for Qwen3-Embedding on
 Ollama (#1072)

Qwen3-Embedding family on Ollama supports Matryoshka truncation via the
'dimensions' field on /v1/embeddings. Without this passthrough, gbrain
ignores user-selected reduced dims and the provider returns its native
size, causing dim-mismatch errors against brains configured for narrower
widths (e.g. existing 1536-dim brains).

Matches by bare name 'qwen3-embedding' or any tag variant
'qwen3-embedding:0.6b' / ':4b' / ':8b'.

Native dims: 0.6B=1024, 4B=2560, 8B=4096. All MRL-truncatable.

5 new tests; full AI suite 137/137 green.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Time Attakc <89218912+time-attack@users.noreply.github.com>
---
 src/core/ai/dims.ts                |  9 +++++++
 test/ai/recipe-ollama-dims.test.ts | 41 ++++++++++++++++++++++++++++++
 2 files changed, 50 insertions(+)
 create mode 100644 test/ai/recipe-ollama-dims.test.ts

diff --git a/src/core/ai/dims.ts b/src/core/ai/dims.ts
index 7961516a5..b88a4e175 100644
--- a/src/core/ai/dims.ts
+++ b/src/core/ai/dims.ts
@@ -263,6 +263,15 @@ export function dimsProviderOptions(
       if (modelId === 'text-embedding-v3' || modelId === 'embedding-3') {
         return { openaiCompatible: { dimensions: dims } };
       }
+      // Qwen3-Embedding family on Ollama (and any other openai-compatible
+      // provider serving it) supports Matryoshka truncation via `dimensions`.
+      // Native sizes: 0.6B=1024, 4B=2560, 8B=4096. Without `dimensions`,
+      // Ollama returns the native size and brains configured for narrower
+      // widths hard-fail with a dim-mismatch error. Pattern match the bare
+      // model name + any `:tag` (e.g. `qwen3-embedding:4b`, `qwen3-embedding:0.6b`).
+      if (modelId === 'qwen3-embedding' || modelId.startsWith('qwen3-embedding:')) {
+        return { openaiCompatible: { dimensions: dims } };
+      }
       // MiniMax embo-01 takes a `type: 'db' | 'query'` field for asymmetric
       // retrieval. Today still hardcoded to 'db' for back-compat — opting
       // into the new inputType seam is a follow-up (see plan's deferred
diff --git a/test/ai/recipe-ollama-dims.test.ts b/test/ai/recipe-ollama-dims.test.ts
new file mode 100644
index 000000000..4f1d86ce8
--- /dev/null
+++ b/test/ai/recipe-ollama-dims.test.ts
@@ -0,0 +1,41 @@
+/**
+ * Ollama Matryoshka dims passthrough.
+ *
+ * Several embedding models served via Ollama (Qwen3-Embedding family) support
+ * Matryoshka truncation through the `dimensions` field on /v1/embeddings.
+ * Without this passthrough, gbrain ignores user-selected reduced dims and the
+ * provider returns its native size, causing dim-mismatch failures against
+ * brains configured for smaller widths.
+ */
+
+import { describe, expect, test } from 'bun:test';
+import { dimsProviderOptions } from '../../src/core/ai/dims.ts';
+
+describe('dims: ollama Matryoshka models', () => {
+  test('qwen3-embedding:4b threads dimensions=1536', () => {
+    expect(dimsProviderOptions('openai-compatible', 'qwen3-embedding:4b', 1536))
+      .toEqual({ openaiCompatible: { dimensions: 1536 } });
+  });
+
+  test('qwen3-embedding:0.6b threads dimensions=512', () => {
+    expect(dimsProviderOptions('openai-compatible', 'qwen3-embedding:0.6b', 512))
+      .toEqual({ openaiCompatible: { dimensions: 512 } });
+  });
+
+  test('qwen3-embedding:8b threads dimensions=2048', () => {
+    expect(dimsProviderOptions('openai-compatible', 'qwen3-embedding:8b', 2048))
+      .toEqual({ openaiCompatible: { dimensions: 2048 } });
+  });
+
+  test('bare qwen3-embedding (no quant tag) also recognized', () => {
+    expect(dimsProviderOptions('openai-compatible', 'qwen3-embedding', 1024))
+      .toEqual({ openaiCompatible: { dimensions: 1024 } });
+  });
+
+  test('unrelated openai-compat model returns undefined (regression guard)', () => {
+    expect(dimsProviderOptions('openai-compatible', 'nomic-embed-text', 768))
+      .toBeUndefined();
+    expect(dimsProviderOptions('openai-compatible', 'mxbai-embed-large', 1024))
+      .toBeUndefined();
+  });
+});