Stabilize core correctness and sync project plan state

src/atocore/retrieval/embeddings.py

@@ -1,8 +1,8 @@
 """Embedding model management."""
 
+import atocore.config as _config
 from sentence_transformers import SentenceTransformer
 
-from atocore.config import settings
 from atocore.observability.logger import get_logger
 
 log = get_logger("embeddings")
@@ -14,9 +14,9 @@ def get_model() -> SentenceTransformer:
     """Load and cache the embedding model."""
     global _model
     if _model is None:
-        log.info("loading_embedding_model", model=settings.embedding_model)
-        _model = SentenceTransformer(settings.embedding_model)
-        log.info("embedding_model_loaded", model=settings.embedding_model)
+        log.info("loading_embedding_model", model=_config.settings.embedding_model)
+        _model = SentenceTransformer(_config.settings.embedding_model)
+        log.info("embedding_model_loaded", model=_config.settings.embedding_model)
     return _model
 
 

src/atocore/retrieval/retriever.py

@@ -3,7 +3,8 @@
 import time
 from dataclasses import dataclass
 
-from atocore.config import settings
+import atocore.config as _config
+from atocore.models.database import get_connection
 from atocore.observability.logger import get_logger
 from atocore.retrieval.embeddings import embed_query
 from atocore.retrieval.vector_store import get_vector_store
@@ -29,7 +30,7 @@ def retrieve(
     filter_tags: list[str] | None = None,
 ) -> list[ChunkResult]:
     """Retrieve the most relevant chunks for a query."""
-    top_k = top_k or settings.context_top_k
+    top_k = top_k or _config.settings.context_top_k
     start = time.time()
 
     query_embedding = embed_query(query)
@@ -59,7 +60,10 @@ def retrieve(
 
     chunks = []
     if results and results["ids"] and results["ids"][0]:
+        existing_ids = _existing_chunk_ids(results["ids"][0])
         for i, chunk_id in enumerate(results["ids"][0]):
+            if chunk_id not in existing_ids:
+                continue
             # ChromaDB returns distances (lower = more similar for cosine)
             # Convert to similarity score (1 - distance)
             distance = results["distances"][0][i] if results["distances"] else 0
@@ -90,3 +94,17 @@ def retrieve(
     )
 
     return chunks
+
+
+def _existing_chunk_ids(chunk_ids: list[str]) -> set[str]:
+    """Filter out stale vector entries whose chunk rows no longer exist."""
+    if not chunk_ids:
+        return set()
+
+    placeholders = ", ".join("?" for _ in chunk_ids)
+    with get_connection() as conn:
+        rows = conn.execute(
+            f"SELECT id FROM source_chunks WHERE id IN ({placeholders})",
+            chunk_ids,
+        ).fetchall()
+    return {row["id"] for row in rows}

src/atocore/retrieval/vector_store.py

@@ -2,7 +2,7 @@
 
 import chromadb
 
-from atocore.config import settings
+import atocore.config as _config
 from atocore.observability.logger import get_logger
 from atocore.retrieval.embeddings import embed_texts
 
@@ -17,13 +17,13 @@ class VectorStore:
     """Wrapper around ChromaDB for chunk storage and retrieval."""
 
     def __init__(self) -> None:
-        settings.chroma_path.mkdir(parents=True, exist_ok=True)
-        self._client = chromadb.PersistentClient(path=str(settings.chroma_path))
+        _config.settings.chroma_path.mkdir(parents=True, exist_ok=True)
+        self._client = chromadb.PersistentClient(path=str(_config.settings.chroma_path))
         self._collection = self._client.get_or_create_collection(
             name=COLLECTION_NAME,
             metadata={"hnsw:space": "cosine"},
         )
-        log.info("vector_store_initialized", path=str(settings.chroma_path))
+        log.info("vector_store_initialized", path=str(_config.settings.chroma_path))
 
     def add(
         self,