feat(langchain): add similarity_search() to CoordinodeGraph

Wraps CoordinodeClient.vector_search() with label/property defaults,
returning [{id, node, distance}, ...] sorted by ascending distance.
Guards against empty query_vector to match server validation behaviour.
Adds two integration tests: one seeding a :LCSim node and verifying the
seeded node appears in top-k results, one verifying empty-vector returns [].

Closes #20

Author: polaz

langchain-coordinode/langchain_coordinode/graph.py

@@ -194,6 +194,39 @@ def query(
         # cypher() returns List[Dict[str, Any]] directly — column name → value.
         return self._client.cypher(query, params=params or {})
 
+    def similarity_search(
+        self,
+        query_vector: list[float],
+        k: int = 10,
+        label: str = "Chunk",
+        property: str = "embedding",
+    ) -> list[dict[str, Any]]:
+        """Find nodes whose ``property`` vector is closest to ``query_vector``.
+
+        Wraps ``CoordinodeClient.vector_search()``.  The returned list contains
+        one dict per result with the keys ``node`` (node properties), ``id``
+        (internal integer node ID), and ``distance`` (cosine distance, lower =
+        more similar).
+
+        Args:
+            query_vector: Embedding vector to search for.
+            k: Maximum number of results to return.
+            label: Node label to search (default ``"Chunk"``).
+            property: Embedding property name (default ``"embedding"``).
+
+        Returns:
+            List of result dicts sorted by ascending distance.
+        """
+        if not query_vector:
+            return []
+        results = self._client.vector_search(
+            label=label,
+            property=property,
+            vector=query_vector,
+            top_k=k,
+        )
+        return [{"id": r.node.id, "node": r.node.properties, "distance": r.distance} for r in results]
+
     # ── Lifecycle ─────────────────────────────────────────────────────────
 
     def close(self) -> None:

tests/integration/adapters/test_langchain.py

@@ -133,6 +133,44 @@ def test_add_graph_documents_idempotent(graph, unique_tag):
     assert result[0]["cnt"] == 1
 
 
+# ── similarity_search ─────────────────────────────────────────────────────────
+
+
+def test_similarity_search_returns_results(graph, unique_tag):
+    """similarity_search() returns node dicts with id, node, and distance keys.
+
+    Seeds a :LCSim node with a known embedding, then searches for the closest
+    vector. The seeded node must appear in the top-k results.
+    """
+    # Derive a unique embedding from the test tag (same technique as llama-index
+    # test) to avoid collisions with other :LCSim nodes in the shared DB.
+    seed = list(bytes.fromhex(unique_tag))
+    vec = [float(seed[i % len(seed)]) / 255.0 for i in range(16)]
+
+    try:
+        seed_rows = graph.query(
+            "CREATE (n:LCSim {id: $id, embedding: $vec}) RETURN n AS nid",
+            params={"id": f"lcsim-{unique_tag}", "vec": vec},
+        )
+        seeded_internal_id = seed_rows[0]["nid"]
+
+        results = graph.similarity_search(vec, k=5, label="LCSim", property="embedding")
+
+        assert isinstance(results, list)
+        assert len(results) >= 1
+        assert all("id" in r and "node" in r and "distance" in r for r in results)
+        assert any(r["id"] == seeded_internal_id for r in results)
+        assert results[0]["distance"] >= 0.0
+    finally:
+        graph.query("MATCH (n:LCSim {id: $id}) DELETE n", params={"id": f"lcsim-{unique_tag}"})
+
+
+def test_similarity_search_empty_vector_returns_empty(graph):
+    """similarity_search() with an empty vector list returns an empty list without error."""
+    results = graph.similarity_search([], k=5)
+    assert isinstance(results, list)
+
+
 def test_schema_refreshes_after_add(graph, unique_tag):
     """structured_schema is invalidated and re-fetched after add_graph_documents."""
     graph._schema = None  # force refresh