Add AstraDBGraphVectorStore testing #75
Conversation
This adapts the CassandraGraphVectorStore to work with AstraDB. Fortunately, not schema changes are required to support the graph retrieval interface, so this should be a drop-in replacement for AstraDBVectorStore in cases where the graph retrieval functions are needed.
…s integration tests
| _documents: Iterable[Document] | ||
| if ids is not None: | ||
| _documents = [document.copy() for document in documents] | ||
| for _doc_id, _document in zip(ids, _documents): |
There was a problem hiding this comment.
There should be some check that ids and documents have the same size.
There was a problem hiding this comment.
Closing since ids is going away from this method.
Side note for your amusement. This check would be not entirely trivial if one does not want to list(...) the iterables. I had started to implement something like this before removing it all.
But the real caveat is that zip(ite1, ite2) treats the two iterables differently (i.e. it's implemented straigthforwardly). It can consume one item too much from ite1 which goes nowhere. The first snippet here prints "4", the second one gives a StopIteration:
r1 = (i for i in range(4))
r2 = (i for i in range(5))
for p in zip(r1, r2):
print(p)
next(r2)
r1 = (i for i in range(4))
r2 = (i for i in range(5))
for p in zip(r2, r1):
print(p)
next(r2)
(almost obvious in hindsight, but still)
… deprecation; adjust testing
|
CI is having lots of troubles (hitting DB limits and whatnot; I'll have a PR to simplify, rationalize and optimize the tests and the way they use DB resources). |
Adding integration+unit test for AstraDBGraphVectorStore.
Doing so uncovered unexpected behaviour w.r.t the "id in document vs. ids passed as standalone" in the from_documents implementation for both AstraDBGraphVectorStore and AstraDBVectorStore, which this PR also fixes (and tests).
Incidentally the "filter to metadata" utility method was made public to allow the Graph layer to defer to its VectorStore delegate instead of reimplementing it.
This PR also addresses a comment to #67 that was unattended, inadvertently bringing a critical bug into main.