diff --git a/generations/library_rag/mcp_tools/retrieval_tools.py b/generations/library_rag/mcp_tools/retrieval_tools.py
index 3660b5a..f7bd618 100644
--- a/generations/library_rag/mcp_tools/retrieval_tools.py
+++ b/generations/library_rag/mcp_tools/retrieval_tools.py
@@ -69,9 +69,39 @@ from mcp_tools.logging_config import (
     log_weaviate_query,
 )
 
+# GPU embedder for BGE-M3 vectorization (replaces text2vec-transformers)
+from memory.core import get_embedder
+
 # Logger for this module - uses structured logging
 logger = get_tool_logger("retrieval")
 
+# =============================================================================
+# GPU Embedder Singleton (BGE-M3)
+# =============================================================================
+
+_embedder = None
+
+
+def get_gpu_embedder():
+    """Get or create GPU embedder singleton for BGE-M3 vectorization.
+
+    Returns the shared GPU embedding service instance. The embedder uses
+    BAAI/bge-m3 model (1024 dimensions) for semantic vectorization.
+
+    Returns:
+        GPUEmbeddingService instance.
+
+    Note:
+        This singleton pattern ensures the model is loaded only once,
+        avoiding repeated GPU memory allocation.
+    """
+    global _embedder
+    if _embedder is None:
+        logger.info("Initializing GPU embedder (BGE-M3) for retrieval...")
+        _embedder = get_embedder()
+        logger.info(f"GPU embedder ready: {_embedder.model_name}")
+    return _embedder
+
 
 # =============================================================================
 # Canonical Reference Extraction
@@ -472,10 +502,14 @@ async def search_chunks_handler(input_data: SearchChunksInput) -> SearchChunksOu
                     )
                     filters = (filters & lang_f) if filters else lang_f
 
-                # Perform near_text query with timing
+                # Vectorize query with GPU embedder (BGE-M3)
+                embedder = get_gpu_embedder()
+                query_vector = embedder.embed_single(input_data.query)
+
+                # Perform near_vector query with timing
                 query_start = time.perf_counter()
-                result = chunks.query.near_text(
-                    query=input_data.query,
+                result = chunks.query.near_vector(
+                    near_vector=query_vector.tolist(),
                     limit=input_data.limit,
                     filters=filters,
                     return_metadata=wvq.MetadataQuery(distance=True),
@@ -484,7 +518,7 @@ async def search_chunks_handler(input_data: SearchChunksInput) -> SearchChunksOu
 
                 # Log Weaviate query
                 log_weaviate_query(
-                    operation="near_text",
+                    operation="near_vector",
                     collection="Chunk",
                     filters={
                         "author": input_data.author_filter,
@@ -619,10 +653,14 @@ async def search_summaries_handler(
                     )
                     filters = (filters & max_filter) if filters else max_filter
 
-                # Perform near_text query with timing
+                # Vectorize query with GPU embedder (BGE-M3)
+                embedder = get_gpu_embedder()
+                query_vector = embedder.embed_single(input_data.query)
+
+                # Perform near_vector query with timing
                 query_start = time.perf_counter()
-                result = summaries.query.near_text(
-                    query=input_data.query,
+                result = summaries.query.near_vector(
+                    near_vector=query_vector.tolist(),
                     limit=input_data.limit,
                     filters=filters,
                     return_metadata=wvq.MetadataQuery(distance=True),
@@ -631,7 +669,7 @@ async def search_summaries_handler(
 
                 # Log Weaviate query
                 log_weaviate_query(
-                    operation="near_text",
+                    operation="near_vector",
                     collection="Summary",
                     filters={
                         "min_level": input_data.min_level,