davebb/linear-coding-agent
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

fix: Replace text2vec-transformers with BGE-M3 in MCP retrieval tools

Browse Source 
The text2vec-transformers Docker service was removed in Jan 2026,
but retrieval_tools.py still used near_text() which requires it.
Now uses GPU embedder (BGE-M3) with near_vector() like flask_app.py.

Changes:
- Add GPU embedder singleton (get_gpu_embedder)
- search_chunks_handler: near_text → near_vector + BGE-M3
- search_summaries_handler: near_text → near_vector + BGE-M3

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
David Blanc Brioir
2026-01-30 00:34:37 +01:00
parent 698e20f292
commit 376d77ddfa
1 changed files with 46 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
generations/library_rag/mcp_tools/retrieval_tools.py
View File

@@ -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,