Cassandra support for chat history using CassIO library

[hemidactylus]

Overview

This PR aims at building on #4378, expanding the capabilities and building on top of the cassIO library to interface with the database (as opposed to using the core drivers directly).

Usage of cassIO (a library abstracting Cassandra access for ML/GenAI-specific purposes) is already established since #6426 was merged, so no new dependencies are introduced.

In the same spirit, we try to uniform the interface for using Cassandra instances throughout LangChain: all our appreciation of the work by @jj701 notwithstanding, who paved the way for this incremental work (thank you!), we identified a few reasons for changing the way a CassandraChatMessageHistory is instantiated. Advocating a syntax change is something we don't take lighthearted way, so we add some explanations about this below.

Additionally, this PR expands on integration testing, enables use of Cassandra's native Time-to-Live (TTL) features and improves the phrasing around the notebook example and the short "integrations" documentation paragraph.

We would kindly request @hwchase to review (since this is an elaboration and proposed improvement of #4378 who had the same reviewer).

About the init breaking changes

There are many options when creating the Cluster object, and new ones might be added at any time. Choosing some of them and exposing them as __init__ parameters CassandraChatMessageHistory will prove to be insufficient for at least some users.

On the other hand, working through kwargs or adding a long, long list of arguments to __init__ is not a desirable option either. For this reason, (as done in #6426), we propose that whoever instantiates the Chat Message History class provide a Cassandra Session object, ready to use. This also enables easier injection of mocks and usage of Cassandra-compatible connections (such as those to the cloud database DataStax Astra DB, obtained with a different set of init parameters than contact_points and port).

We feel that a breaking change might still be acceptable since LangChain is at 0.*. However, while maintaining that the approach we propose will be more flexible in the future, room could be made for a "compatibility layer" that respects the current init method. Honestly, we would to that only if there are strong reasons for it, as that would entail an additional maintenance burden.

Other changes

We propose to remove the keyspace creation from the class code for two reasons: first, production Cassandra instances often employ RBAC so that the database user reading/writing from tables does not necessarily (and generally shouldn't) have permission to create keyspaces, and second that programmatic keyspace creation is not a best practice (it should be done more or less manually, with extra care about schema mismatched among nodes, etc). Removing this (usually unnecessary) operation from the __init__ path would also improve initialization performance (shorter time).

We suggest, likewise, to remove the __del__ method (which would close the database connection), for the following reason: it is the recommended best practice to create a single Cassandra Session object throughout an application (it is a resource-heavy object capable to handle concurrency internally), so in case Cassandra is used in other ways by the app there is the risk of truncating the connection for all usages when the history instance is destroyed. Moreover, the Session object, in typical applications, is best left to garbage-collect itself automatically.

As mentioned above, we defer the actual database I/O to the cassIO library, which is designed to encode practices optimized for LLM applications (among other) without the need to expose LangChain developers to the internals of CQL (Cassandra Query Language). CassIO is already employed by the LangChain's Vector Store support for Cassandra.

We added a few more connection options in the companion notebook example (most notably, Astra DB) to encourage usage by anyone who cannot run their own Cassandra cluster.

We surface the ttl_seconds option for automatic handling of an expiration time to chat history messages, a likely useful feature given that very old messages generally may lose their importance.

We elaborated a bit more on the integration testing (Time-to-live, separation of "session ids", ...).

Remarks from linter & co.

We reinstated cassio as a dependency both in the "optional" group and in the "integration testing" group of pyproject.toml. This might not be the right thing do to, in which case the author of this PR offer his apologies (lack of confidence with Poetry - happy to be pointed in the right direction, though!).

During linter tests, we were hit by some errors which appear unrelated to the code in the PR. We left them here and report on them here for awareness:

langchain/vectorstores/mongodb_atlas.py:137: error: Argument 1 to "insert_many" of "Collection" has incompatible type "List[Dict[str, Sequence[object]]]"; expected "Iterable[Union[MongoDBDocumentType, RawBSONDocument]]"  [arg-type]
langchain/vectorstores/mongodb_atlas.py:186: error: Argument 1 to "aggregate" of "Collection" has incompatible type "List[object]"; expected "Sequence[Mapping[str, Any]]"  [arg-type]

langchain/vectorstores/qdrant.py:16: error: Name "grpc" is not defined  [name-defined]
langchain/vectorstores/qdrant.py:19: error: Name "grpc" is not defined  [name-defined]
langchain/vectorstores/qdrant.py:20: error: Name "grpc" is not defined  [name-defined]
langchain/vectorstores/qdrant.py:22: error: Name "grpc" is not defined  [name-defined]
langchain/vectorstores/qdrant.py:23: error: Name "grpc" is not defined  [name-defined]

In the same spirit, we observe that to even get import langchain run, it seems that a pip install bs4 is missing from the minimal package installation path.

Thank you!

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)			Jun 29, 2023 4:58pm

[hemidactylus]

Trying to fix linter/build errors I see only here in the PR discussion.
Locally when I make lint only a Python 3.10 run starts, is there a way to run the same checks (incl. "build") which I see are then triggered in the PR? I'm not very familiar with some of the tools involved, sorry.

[jbellis]

@rlancemartin could you take a look at this Cassandra extension?

[hemidactylus]

Thank you for taking action @rlancemartin !
However, I am afraid an unresolved <<<<<<< HEAD resulted in the poetry.lock (from the force-push) which is what is making all tests above fail.

[rlancemartin]

Thank you for taking action @rlancemartin ! However, I am afraid an unresolved <<<<<<< HEAD resulted in the poetry.lock (from the force-push) which is what is making all tests above fail.

Ya, re-generating it now from the updated pyproject.toml :)

Just discussed a bit w/ @dev2049 and typical workflow when the pyproject.toml is updated is to run:

poetry lock --no-update

Which will re-generate the lock file (without updating the dependencies to their latest versions).