From bdbf47a341d70cdc67b5c4ba95abaace7e79d88a Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Wed, 27 May 2026 00:11:40 +0400
Subject: [PATCH] Studio: default multimodal RAG embedder to BAAI/BGE-VL-large
 (Qwen3-VL kept as alt)

---
 studio/backend/requirements/extras-no-deps.txt |  7 ++++---
 studio/backend/routes/rag.py                   |  7 +++----
 studio/backend/utils/rag/config.py             | 14 +++++++++++++-
 3 files changed, 20 insertions(+), 8 deletions(-)

diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt
index 952a709137..78eec61cbf 100644
--- a/studio/backend/requirements/extras-no-deps.txt
+++ b/studio/backend/requirements/extras-no-deps.txt
@@ -15,9 +15,10 @@ git+https://github.com/meta-pytorch/OpenEnv.git
 # executorch>=1.0.1               # 41.5 MB - no imports in unsloth/zoo/studio
 torch-c-dlpack-ext
 # Bumped from 5.2.0 to expose the `sentence_transformers.base.modules`
-# package that Qwen3-VL-Embedding-2B (the multimodal RAG embedder)
-# references in its modules.json. Upper bound left at <7 to avoid
-# accidental jumps across a future major rewrite.
+# package that Qwen3-VL-Embedding-2B references in its modules.json
+# (still supported as an alternative multimodal RAG embedder; the
+# current default is BAAI/BGE-VL-large via our in-process adapter).
+# Upper bound at <7 to avoid jumps across a future major rewrite.
 sentence_transformers>=5.3.0,<7
 transformers==4.57.6
 pytorch_tokenizers
diff --git a/studio/backend/routes/rag.py b/studio/backend/routes/rag.py
index 38ef6b14e4..7ad38c3b64 100644
--- a/studio/backend/routes/rag.py
+++ b/studio/backend/routes/rag.py
@@ -441,10 +441,9 @@ def warmup_rag_embedder(
 ) -> dict:
     """Preload the configured default embedder so the first retrieval is warm.
 
-    Called from the frontend when the user enables the RAG pill — moves the
-    cold-load latency (Qwen3-VL-Embedding-2B is ~4 GB) out of the first
-    chat-completion path, where a 30s+ load can race the llama-server
-    prefill timeout.
+    Called from the frontend when the user enables the RAG pill — moves
+    the cold-load latency out of the first chat-completion path, where a
+    multi-second load can race the llama-server prefill timeout.
     """
     from utils.rag.config import resolve_embedder
 
diff --git a/studio/backend/utils/rag/config.py b/studio/backend/utils/rag/config.py
index e394288368..b265b90b51 100644
--- a/studio/backend/utils/rag/config.py
+++ b/studio/backend/utils/rag/config.py
@@ -33,10 +33,22 @@ RAG_EMBEDDING_MODEL: str = (
 
 # Default embedder per (mode, chunking). (multimodal, late) is unsupported
 # and rejected at KB-create time in routes/rag.py.
+#
+# Multimodal default is BAAI/BGE-VL-large (~400 M params, 768-d, ~800 MB
+# bf16) — small, fast, shared text/image space. Loaded via the
+# `_BGEVLAdapter` in core/rag/embeddings.py which bypasses BGE-VL's
+# fragile sentence-transformers shim and pre-truncates text to CLIP's
+# 77-token cap.
+#
+# To switch back to Qwen3-VL-Embedding-2B (2 B params, 2048-d, no CLIP
+# text cap; ~4 GB bf16 / ~1.5 GB 4-bit via FastSentenceTransformer),
+# change the ("multimodal", "standard") entry below — the in-process
+# loader supports both via `model_name.startswith("BAAI/BGE-VL")`
+# routing.
 RAG_EMBEDDER_MATRIX: dict[tuple[str, str], str] = {
     ("text", "standard"): "BAAI/bge-small-en-v1.5",
     ("text", "late"): "nomic-ai/nomic-embed-text-v1.5",
-    ("multimodal", "standard"): "Qwen/Qwen3-VL-Embedding-2B",
+    ("multimodal", "standard"): "BAAI/BGE-VL-large",
 }