superlocalmemory 3.3.11 → 3.3.13

package/src/superlocalmemory/retrieval/reranker.py

@@ -35,7 +35,9 @@ _live_rerankers: set[weakref.ref] = set()
 logger = logging.getLogger(__name__)
 
 _IDLE_TIMEOUT_SECONDS = 120  # 2 min → kill worker
-_SUBPROCESS_RESPONSE_TIMEOUT = 120  # 120s for ONNX cold start
+# V3.3.12: Configurable via SLM_RERANKER_IDLE_TIMEOUT env var
+_IDLE_TIMEOUT_SECONDS = int(os.environ.get("SLM_RERANKER_IDLE_TIMEOUT", _IDLE_TIMEOUT_SECONDS))
+_SUBPROCESS_RESPONSE_TIMEOUT = 180  # V3.3.12: 180s (was 120s) for stressed system respawns
 _WORKER_RECYCLE_AFTER = 500  # Recycle after N requests
 
 
@@ -57,7 +59,7 @@ class CrossEncoderReranker:
 
     def __init__(
         self,
-        model_name: str = "cross-encoder/ms-marco-MiniLM-L-6-v2",
+        model_name: str = "cross-encoder/ms-marco-MiniLM-L-12-v2",
         backend: str = "onnx",
     ) -> None:
         self._model_name = model_name
@@ -129,8 +131,23 @@ class CrossEncoderReranker:
             finally:
                 self._worker_loading = False
 
-        t = threading.Thread(target=_warmup, daemon=True, name="ce-warmup")
-        t.start()
+        self._warmup_thread = threading.Thread(target=_warmup, daemon=True, name="ce-warmup")
+        self._warmup_thread.start()
+
+    def warmup_sync(self, timeout: float = 120.0) -> bool:
+        """Block until reranker model is loaded. Returns True if ready.
+
+        V3.3.12: Critical for benchmarks and first-recall quality.
+        Without this, first 30-60s of recalls get no reranking (-30.7pp).
+        """
+        if self._model_loaded:
+            return True
+        if not self._worker_loading and not self._model_loaded:
+            self._start_background_warmup()
+        t = getattr(self, '_warmup_thread', None)
+        if t is not None:
+            t.join(timeout=timeout)
+        return self._model_loaded
 
     # ------------------------------------------------------------------
     # Worker management (mirrors EmbeddingService pattern)
@@ -304,13 +321,13 @@ class CrossEncoderReranker:
 
         documents = [fact.content for fact, _ in candidates]
 
-        # Short timeout (10s) — model should already be loaded by warmup.
-        # If worker crashed or is still loading, fallback immediately.
+        # V3.3.12: Increased timeout 10s→60s — L-12-v2 needs PyTorch + ONNX load.
+        # Critical: Paper 2 ablation showed -30.7pp without reranking.
         resp = self._send_request({
             "cmd": "rerank",
             "query": query,
             "documents": documents,
-        }, timeout=10.0)
+        }, timeout=60.0)
 
         if resp is None or not resp.get("ok"):
             # Fallback: return by existing score

package/src/superlocalmemory/retrieval/semantic_channel.py

@@ -182,7 +182,7 @@ class SemanticChannel:
             else:
                 sim = cos_sim
 
-            if sim > 0.3:
+            if sim > 0.05:
                 scored.append((fact.fact_id, sim))
 
         scored.sort(key=lambda x: x[1], reverse=True)
@@ -235,7 +235,7 @@ class SemanticChannel:
             else:
                 sim = cos_sim
 
-            if sim > 0.3:
+            if sim > 0.05:
                 scored.append((fact.fact_id, sim))
 
         scored.sort(key=lambda x: x[1], reverse=True)

package/src/superlocalmemory/retrieval/temporal_channel.py

@@ -124,6 +124,16 @@ class TemporalChannel:
         import re
         _PROPER_RE = re.compile(r"\b([A-Z][a-z]+)\b")
         names = [m.group(1) for m in _PROPER_RE.finditer(query)]
+        # Also try title-cased version for lowercase queries
+        if not names:
+            names = [m.group(1) for m in _PROPER_RE.finditer(query.title())]
+        # Filter out common words from title-casing
+        _stop = {"What", "When", "Where", "Who", "Which", "How", "Does", "Did",
+                 "The", "That", "This", "There", "Then", "Have", "Has", "Had",
+                 "About", "After", "Before", "From", "With", "Would", "Could",
+                 "Should", "Will", "Because", "Also", "Just", "Like", "Know",
+                 "Think", "Tell", "Said"}
+        names = [n for n in names if n not in _stop]
         if not names:
             return []
 
@@ -146,7 +156,10 @@ class TemporalChannel:
                 fid = dict(row)["fact_id"]
                 if fid not in seen:
                     seen.add(fid)
-                    results.append((fid, 0.85))  # High base score for entity-temporal
+                    # Rank by position (first events more likely relevant) instead
+                    # of flat 0.85 which loses discrimination
+                    rank_score = 0.85 - len(seen) * 0.02
+                    results.append((fid, max(0.3, rank_score)))
 
         return results
 

package/src/superlocalmemory/storage/schema.py

@@ -442,6 +442,8 @@ CREATE INDEX IF NOT EXISTS idx_edges_target
     ON graph_edges (profile_id, target_id);
 CREATE INDEX IF NOT EXISTS idx_edges_type
     ON graph_edges (profile_id, edge_type);
+CREATE INDEX IF NOT EXISTS idx_edges_exists_check
+    ON graph_edges (profile_id, source_id, target_id, edge_type);
 """