Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

perf: reuse query embeddings in hierarchical retriever (#93)

mildred522f555d6a34b52
diff --git a/openviking/retrieve/hierarchical_retriever.py b/openviking/retrieve/hierarchical_retriever.py
index a0a54b2f1..689f0b2c9 100644
--- a/openviking/retrieve/hierarchical_retriever.py
+++ b/openviking/retrieve/hierarchical_retriever.py
@@ -108,13 +108,30 @@ class HierarchicalRetriever:
final_metadata_filter = {"op": "and", "conds": filters_to_merge}
+ if not await self.storage.collection_exists(collection):
+ logger.warning(f"[RecursiveSearch] Collection {collection} does not exist")
+ return QueryResult(
+ query=query,
+ matched_contexts=[],
+ searched_directories=[],
+ )
+
+ # Generate query vectors once to avoid duplicate embedding calls
+ query_vector = None
+ sparse_query_vector = None
+ if self.embedder:
+ result: EmbedResult = self.embedder.embed(query.query)
+ query_vector = result.dense_vector
+ sparse_query_vector = result.sparse_vector
+
# Step 1: Determine starting directories based on context_type
root_uris = self._get_root_uris_for_type(query.context_type)
# Step 2: Global vector search to supplement starting points
global_results = await self._global_vector_search(
- query=query.query,
collection=collection,
+ query_vector=query_vector,
+ sparse_query_vector=sparse_query_vector,
limit=self.GLOBAL_SEARCH_TOPK,
filter=final_metadata_filter,
)
@@ -125,8 +142,10 @@ class HierarchicalRetriever:
# Step 4: Recursive search
candidates = await self._recursive_search(
query=query.query,
- starting_points=starting_points,
collection=collection,
+ query_vector=query_vector,
+ sparse_query_vector=sparse_query_vector,
+ starting_points=starting_points,
limit=limit,
mode=mode,
threshold=effective_threshold,
@@ -145,21 +164,16 @@ class HierarchicalRetriever:
async def _global_vector_search(
self,
- query: str,
collection: str,
+ query_vector: Optional[List[float]],
+ sparse_query_vector: Optional[Dict[str, float]],
limit: int,
filter: Optional[Dict[str, Any]] = None,
) -> List[Dict[str, Any]]:
"""Global vector search to locate initial directories."""
- if not self.embedder:
- return []
- if not await self.storage.collection_exists(collection):
- return []
- result: EmbedResult = self.embedder.embed(query)
- query_vector = result.dense_vector
if not query_vector:
return []
- sparse_query_vector = result.sparse_vector or {}
+ sparse_query_vector = sparse_query_vector or {}
global_filter = {
"op": "and",
@@ -215,8 +229,10 @@ class HierarchicalRetriever:
async def _recursive_search(
self,
query: str,
- starting_points: List[Tuple[str, float]],
collection: str,
+ query_vector: Optional[List[float]],
+ sparse_query_vector: Optional[Dict[str, float]],
+ starting_points: List[Tuple[str, float]],
limit: int,
mode: str,
threshold: Optional[float] = None,
@@ -249,18 +265,7 @@ class HierarchicalRetriever:
return base_filter
return {"op": "and", "conds": [base_filter, extra_filter]}
- # Generate query vectors
- query_vector = None
- sparse_query_vector = None
-
- if self.embedder:
- result: EmbedResult = self.embedder.embed(query)
- query_vector = result.dense_vector
- sparse_query_vector = result.sparse_vector
-
- if not await self.storage.collection_exists(collection):
- logger.warning(f"[RecursiveSearch] Collection {collection} does not exist")
- return []
+ sparse_query_vector = sparse_query_vector or None
collected: List[Dict[str, Any]] = [] # Collected results (directories and leaves)
dir_queue: List[tuple] = [] # Priority queue: (-score, uri)