fix(cassandra): stop forwarding batch_size into Cassandra.__init__ - #14573

Summary

Fixes #6255.

CassandraVectorStoreComponent.build_vector_store() ingests documents through langchain_community.vectorstores.Cassandra.from_documents(..., batch_size=self.batch_size, ...). That classmethod forwards every keyword argument it doesn't explicitly consume — including batch_size — straight into Cassandra.__init__(), which never accepted a batch_size parameter (it's only accepted by add_texts/add_documents).

Because the component's Batch Size input defaults to 16, this isn't an edge case — any flow that connects an "Ingest Data" step to the Cassandra vector store hits this immediately:

TypeError: Cassandra.__init__() got an unexpected keyword argument 'batch_size'

Reproduced locally against the exact langchain_community version pinned in this repo (no live Cassandra cluster needed — the mismatch is raised at Python's argument-binding step, before any network call):

Cassandra.from_documents(
    documents=[Document(page_content="hello")],
    embedding=FakeEmbeddings(),
    table_name="t", keyspace="k", ttl_seconds=None,
    batch_size=16, body_index_options=None,
)
# TypeError: Cassandra.__init__() got an unexpected keyword argument 'batch_size'

Fix

Construct the Cassandra store explicitly (no batch_size in the constructor call) and pass batch_size to add_documents() afterward, which is where the parameter actually belongs:

 if documents:
     self.log(f"Adding {len(documents)} documents to the Vector Store.")
-    table = Cassandra.from_documents(
-        documents=documents,
+    table = Cassandra(
         embedding=self.embedding,
         table_name=self.table_name,
         keyspace=self.keyspace,
         ttl_seconds=self.ttl_seconds or None,
-        batch_size=self.batch_size,
         body_index_options=body_index_options,
+        setup_mode=setup_mode,
     )
+    table.add_documents(documents, batch_size=self.batch_size)
 else:
     ...

As a side effect, this also fixes setup_mode being silently dropped whenever documents were present: the old from_documents(...) call in that branch never passed setup_mode at all, so the table was always created with the library's SYNC default regardless of what the user picked in the UI. The explicit constructor call now passes setup_mode through in both branches, matching the existing (already-correct) no-documents branch.

No other call sites reference this component's build_vector_store internals, so the change is isolated to src/lfx/src/lfx/components/cassandra/cassandra.py.

Test plan

• Added src/backend/tests/unit/components/vectorstores/test_cassandra_vector_store_component.py:
  • batch_size never reaches Cassandra.__init__, only add_documents
  • setup_mode reaches the constructor even when documents are present (the secondary bug)
  • the no-documents branch is unaffected
  • end-to-end check against the real langchain_community.vectorstores.Cassandra signature confirms the original TypeError no longer occurs
• uv run pytest src/backend/tests/unit/components/vectorstores/test_cassandra_vector_store_component.py -v → 4 passed
• uv run pytest src/backend/tests/unit/components/vectorstores/ → no new failures (pre-existing, unrelated failures on Chroma/FAISS tests reproduce identically on main without this change — local-file-access sandboxing issue, unrelated to Cassandra)
• ruff check / ruff format --check on changed files → clean
• Manually reproduced the original crash against the real langchain_community.vectorstores.Cassandra class and confirmed the fix moves past the TypeError (fails later only on the expected "no live cluster" error in this sandboxed environment)

Summary by CodeRabbit

• Bug Fixes

  • Fixed Cassandra vector store setup so document ingestion no longer causes constructor errors related to batch sizing.
  • Preserved configuration options, including setup mode, embeddings, table and keyspace settings, TTL, body indexing, and batch size.
  • Maintained existing behavior when no documents are provided.

• Tests

  • Added coverage for Cassandra setup, document ingestion, batch-size handling, and no-document scenarios.