fix: clamp overly aggressive edge_rank_threshold to preserve evidence

Author: hammertoe

lib/kg_agent_loop.py

@@ -621,7 +621,7 @@ def _tool_declarations(self) -> list[types.Tool]:
                 "Hybrid Graph-RAG: vector/fulltext seed search over kg_nodes, then expand kg_edges N hops "
                 "and return a compact subgraph plus provenance citations with youtube timecoded URLs. "
                 "Also retrieves bill excerpt citations for legislation questions. "
-                "Use edge_rank_threshold to filter low-quality edges (0.05 recommended after normalization)."
+                "Use edge_rank_threshold to filter low-quality edges (recommended: 0.001 or omit for no filtering)."
             ),
             parameters_json_schema=tool_schema,
         )

lib/kg_hybrid_graph_rag.py

@@ -914,9 +914,13 @@ def kg_hybrid_graph_rag(
         max_citations: Maximum citations to return (default: 12)
         edge_rank_threshold: Optional threshold to filter edges by edge_rank_score.
             Only returns edges with score >= threshold.
-            Recommended: 0.05 after normalization, or 0.00001 before normalization.
+            Recommended: 0.001 or omit for no filtering.
             None means no threshold filtering (default: None)
     """
+    # Guard against overly aggressive thresholds that filter out most edges
+    if edge_rank_threshold is not None and edge_rank_threshold >= 0.05:
+        edge_rank_threshold = None
+
     query = (query or "").strip()
     if not query:
         return {