Skip to content

Graph-Native Infrastructure for Context and Accountable AI Systems

fix: address code review findings on backend-delegated search path

KaifAhmad194adcf7ad320
diff --git a/semantica/vector_store/hybrid_search.py b/semantica/vector_store/hybrid_search.py
index edfb8ccd..1f6d2a84 100644
--- a/semantica/vector_store/hybrid_search.py
+++ b/semantica/vector_store/hybrid_search.py
@@ -298,9 +298,13 @@ class HybridSearch:
Returns:
List of search results
"""
- # Handle legacy argument top_k
+ # Handle legacy argument top_k. Must be popped, not just read: it's
+ # later forwarded via **options to VectorStore.search_vectors(),
+ # which itself passes an explicit top_k=k to some backends (e.g.
+ # sqlite, pgvector) — a leftover 'top_k' in options collides with
+ # that and raises "got multiple values for keyword argument".
if "top_k" in options:
- k = options["top_k"]
+ k = options.pop("top_k")
# query_vector is derived from `query` above; drop any stray value
# passed in **options so it doesn't collide with that derivation.
# (**options is a fresh dict per call, so this can't affect the caller.)
@@ -373,12 +377,17 @@ class HybridSearch:
# Normalize to the same result schema the local search
# path produces, since backends don't all return the
- # same set of keys.
+ # same set of keys. Unlike the local path (which always
+ # knows distance = 1.0 - cosine_similarity), backends use
+ # different metrics (L2, inner product, cosine, ...), so
+ # a missing distance is left as None rather than guessed
+ # from score — assuming cosine semantics would produce
+ # nonsensical values for e.g. raw L2 distances.
normalized_results = [
{
"id": result.get("id"),
"score": result.get("score", 0.0),
- "distance": result.get("distance", result.get("score", 0.0)),
+ "distance": result.get("distance"),
"metadata": result.get("metadata", {}),
}
for result in backend_results[:k]
diff --git a/semantica/vector_store/vector_store.py b/semantica/vector_store/vector_store.py
index 1348162a..8fbd48e4 100644
--- a/semantica/vector_store/vector_store.py
+++ b/semantica/vector_store/vector_store.py
@@ -644,8 +644,13 @@ class VectorStore:
return self._backend_store.search(query_vector, top_k=k, **options)
elif hasattr(self._backend_store, 'search_similar'):
return self._backend_store.search_similar(query_vector, k=k, **options)
+ elif hasattr(self._backend_store, 'search_vectors'):
+ # Some stores (QdrantStore, MilvusStore, PineconeStore) name
+ # their count parameter differently (limit vs k), so bind it
+ # positionally rather than guessing the keyword.
+ return self._backend_store.search_vectors(query_vector, k, **options)
else:
- raise NotImplementedError(f"Backend store {type(self._backend_store).__name__} does not have search or search_similar method")
+ raise NotImplementedError(f"Backend store {type(self._backend_store).__name__} does not have search, search_similar, or search_vectors method")
# Use in-memory implementation
tracking_id = self.progress_tracker.start_tracking(