Skip to content

Graph-Native Infrastructure for Context and Accountable AI Systems

fixed qodo reviews

Sameer6305a4dac2342be7
diff --git a/semantica/vector_store/decision_embedding_pipeline.py b/semantica/vector_store/decision_embedding_pipeline.py
index 070614d8..46f61ac7 100644
--- a/semantica/vector_store/decision_embedding_pipeline.py
+++ b/semantica/vector_store/decision_embedding_pipeline.py
@@ -346,9 +346,6 @@ class DecisionEmbeddingPipeline:
# If vector is all zeros (placeholder), we use the score from search_vectors
if not np.any(original_vec):
score = candidate_embeddings["scores"][idx]
- # Simple normalization if score looks like a distance (FAISS)
- if score > 1.0:
- score = 1.0 / (1.0 + score)
sim["semantic_similarity"] = score
# Recompute overall similarity
sim["similarity"] = (actual_weights[0] * score) + (actual_weights[1] * sim["structural_similarity"])
@@ -793,29 +790,69 @@ class DecisionEmbeddingPipeline:
if query_vector is None:
query_vector = np.zeros(self.embedding_dimension)
- results = self.vector_store.search_vectors(query_vector, k=limit, **(filters or {}))
-
embeddings = []
metadata = []
scores = []
- for res in results:
- vector_metadata = res.get("metadata", {})
+ # If we have filters, we might need to fetch more candidates iteratively
+ # if the backend doesn't support native filtering and we rely on post-filtering.
+ current_k = limit
+ max_k = limit * 10 if limit else 100000
+
+ while current_k <= max_k:
+ results = self.vector_store.search_vectors(query_vector, k=current_k, filter=filters)
- # Extract structural embedding if available
- struct_emb = vector_metadata.get("structural_embedding")
- if struct_emb:
- struct_emb = np.array(struct_emb)
- else:
- struct_emb = np.zeros(self.node_embedding_dimension)
+ # Post-filter and collect
+ collected_embeddings = []
+ collected_metadata = []
+ collected_scores = []
+
+ for res in results:
+ # Handle unstandardized backends (e.g. Qdrant uses payload, Pinecone uses metadata)
+ vector_metadata = res.get("metadata") or res.get("payload") or {}
+
+ # Apply post-filtering for backends that ignored the `filter` param
+ if filters:
+ match = True
+ for key, value in filters.items():
+ if key not in vector_metadata or vector_metadata.get(key) != value:
+ match = False
+ break
+ if not match:
+ continue
+
+ # Extract structural embedding if available
+ struct_emb = vector_metadata.get("structural_embedding")
+ if struct_emb:
+ struct_emb = np.array(struct_emb)
+ else:
+ struct_emb = np.zeros(self.node_embedding_dimension)
+
+ # Handle unstandardized backends (e.g. Pinecone might use values)
+ vector = res.get("vector")
+ if vector is None:
+ vector = res.get("values")
+ if vector is None:
+ vector = np.zeros(self.embedding_dimension)
+
+ collected_embeddings.append((vector, struct_emb))
+ collected_metadata.append(vector_metadata)
+
+ if "distance" in res and res["distance"] is not None:
+ collected_scores.append(1.0 / (1.0 + float(res["distance"])))
+ else:
+ collected_scores.append(float(res.get("score", 0.0)))
- vector = res.get("vector")
- if vector is None:
- vector = np.zeros(self.embedding_dimension)
+ # If we fetched enough matching results, or if the backend returned fewer results
+ # than we asked for (meaning we hit the end of the db), we can stop fetching.
+ if len(collected_embeddings) >= limit or len(results) < current_k:
+ embeddings = collected_embeddings[:limit]
+ metadata = collected_metadata[:limit]
+ scores = collected_scores[:limit]
+ break
- embeddings.append((vector, struct_emb))
- metadata.append(vector_metadata)
- scores.append(res.get("score", 0.0))
+ # Otherwise, double the search pool and try again
+ current_k *= 2
return {"embeddings": embeddings, "metadata": metadata, "scores": scores}
@@ -835,8 +872,6 @@ class DecisionEmbeddingPipeline:
# Calculate semantic similarity
if not np.any(sem_emb) and candidate_scores:
similarity = candidate_scores[i]
- if similarity > 1.0:
- similarity = 1.0 / (1.0 + similarity)
else:
similarity = self.hybrid_calculator._calculate_similarity(
query_embedding, sem_emb, "cosine"
diff --git a/semantica/vector_store/faiss_store.py b/semantica/vector_store/faiss_store.py
index 2ebf37e8..cd759886 100644
--- a/semantica/vector_store/faiss_store.py
+++ b/semantica/vector_store/faiss_store.py
@@ -137,11 +137,17 @@ class FAISSSearch:
for i, (dist, idx) in enumerate(zip(distances[0], indices[0])):
if idx < len(self.index.vector_ids):
vector_id = self.index.vector_ids[idx]
+ dist_val = float(dist)
+
+ # Standardize score as similarity (0.0 to 1.0)
+ # while preserving original distance
+ similarity_score = 1.0 / (1.0 + dist_val)
+
results.append(
{
"id": vector_id,
- "score": float(dist),
- "distance": float(dist),
+ "score": similarity_score,
+ "distance": dist_val,
"metadata": self.index.metadata.get(vector_id, {}),
}
)
diff --git a/tests/vector_store/test_decision_embedding_pipeline.py b/tests/vector_store/test_decision_embedding_pipeline.py
index 9a1b5351..15e5f444 100644
--- a/tests/vector_store/test_decision_embedding_pipeline.py
+++ b/tests/vector_store/test_decision_embedding_pipeline.py
@@ -395,6 +395,9 @@ class TestDecisionEmbeddingPipelineEdgeCases:
def test_find_similar_decisions_real_faiss_backend(self):
"""Regression test for FAISS backend crashing due to .vectors access (issue #839)"""
+ # Skip if faiss is not installed
+ pytest.importorskip("faiss")
+
try:
from semantica.vector_store import VectorStore
except ImportError:
@@ -402,6 +405,9 @@ class TestDecisionEmbeddingPipelineEdgeCases:
# Create a real VectorStore with FAISS backend
vs = VectorStore(backend="faiss", config={"dimension": 384})
+ # Force fallback random embeddings of the exact dimension (384) to avoid sentence-transformers
+ # dependency or default 128-d mismatches.
+ vs.embedder = None
vs.initialize_decision_pipeline()
# Store a couple of dummy decisions to ensure index has data
@@ -425,6 +431,7 @@ class TestDecisionEmbeddingPipelineEdgeCases:
pytest.skip("VectorStore not available")
vs = VectorStore(backend="inmemory", config={"dimension": 384})
+ vs.embedder = None
vs.initialize_decision_pipeline()
vs.store_decision(scenario="Apple product launch", category="tech")