fix: expand KG seed fulltext to key phrase

Author: hammertoe

lib/kg_hybrid_graph_rag.py

@@ -617,6 +617,33 @@ def _retrieve_seed_nodes(
             }
         )
 
+    query_terms = _query_terms(query)
+    if len(query_terms) >= 2:
+        phrase_query = " ".join(query_terms[:2])
+        if phrase_query and phrase_query != query:
+            rows = postgres.execute_query(
+                """
+                SELECT id, type, label, aliases, ts_rank(tsv, plainto_tsquery('english', %s)) as rank
+                FROM kg_nodes
+                WHERE tsv @@ plainto_tsquery('english', %s)
+                ORDER BY rank DESC
+                LIMIT %s
+                """,
+                (phrase_query, phrase_query, seed_k * 2),
+            )
+            for row in rows:
+                rank = float(row[4] or 0.0)
+                fulltext_candidates.append(
+                    {
+                        "id": row[0],
+                        "type": row[1],
+                        "label": row[2],
+                        "aliases": row[3] or [],
+                        "score": rank,
+                        "match_reason": "fulltext_phrase",
+                    }
+                )
+
     # Alias exact match (cheap + good for proper nouns)
     try:
         from lib.id_generators import normalize_label

tests/test_kg_hybrid_graph_rag_unit.py

@@ -165,6 +165,28 @@ def generate_query_embedding(self, _query: str) -> list[float]:
         return [0.0] * 768
 
 
+class _FakePostgresPhraseMatch(_FakePostgres):
+    def execute_query(self, sql: str, params: tuple[Any, ...] | None = None):
+        if "plainto_tsquery" in sql and params:
+            query = params[0]
+            if query == "future barbados":
+                return [
+                    (
+                        "kg_future",
+                        "schema:Organization",
+                        "Future Barbados",
+                        ["Future Barbados"],
+                        0.8,
+                    )
+                ]
+            return []
+
+        if "FROM kg_nodes" in sql and "WHERE id IN" in sql:
+            return [("kg_future", "Future Barbados", "schema:Organization")]
+
+        return super().execute_query(sql, params)
+
+
 class _FakePostgresNameMatch(_FakePostgres):
     def execute_query(self, sql: str, params: tuple[Any, ...] | None = None):
         if "FROM kg_nodes" in sql and "embedding <=>" in sql:
@@ -270,6 +292,26 @@ def test_kg_hybrid_graph_rag_prefers_name_substring_matches() -> None:
     assert "Tameisha Rochester" not in seed_labels
 
 
+def test_kg_hybrid_graph_rag_adds_phrase_fulltext_candidates() -> None:
+    from lib.kg_hybrid_graph_rag import kg_hybrid_graph_rag
+
+    postgres = _FakePostgresPhraseMatch()
+    embedding = _FakeEmbedding()
+
+    out = kg_hybrid_graph_rag(
+        postgres=postgres,
+        embedding_client=embedding,
+        query="Future Barbados recent 2026 budget debates",
+        hops=1,
+        seed_k=5,
+        max_edges=10,
+        max_citations=5,
+    )
+
+    seed_labels = {s["label"] for s in out["seeds"]}
+    assert "Future Barbados" in seed_labels
+
+
 def test_kg_hybrid_graph_rag_respects_bill_citation_limit() -> None:
     from lib.kg_hybrid_graph_rag import kg_hybrid_graph_rag_with_bills