- Remove the client implementation (this breaks backwards compatibility
for existing testers. I could keep the stub in that file if we want, but
not many people are using it yet
- Add SDK as dependency
- Update the 'run_on_dataset' method to be a function that optionally
accepts a client as an argument
- Remove the langchain plus server implementation (you get it for free
with the SDK now)
We could make the SDK optional for now, but the plan is to use w/in the
tracer so it would likely become a hard dependency at some point.
Extract the methods specific to running an LLM or Chain on a dataset to
separate utility functions.
This simplifies the client a bit and lets us separate concerns of LCP
details from running examples (e.g., for evals)
Previously, the client expected a strict 'prompt' or 'messages' format
and wouldn't permit running a chat model or llm on prompts or messages
(respectively).
Since many datasets may want to specify custom key: string , relax this
requirement.
Also, add support for running a chat model on raw prompts and LLM on
chat messages through their respective fallbacks.
# Your PR Title (What it does)
<!--
Thank you for contributing to LangChain! Your PR will appear in our next
release under the title you set. Please make sure it highlights your
valuable contribution.
Replace this with a description of the change, the issue it fixes (if
applicable), and relevant context. List any dependencies required for
this change.
After you're done, someone will review your PR. They may suggest
improvements. If no one reviews your PR within a few days, feel free to
@-mention the same people again, as notifications can get lost.
-->
<!-- Remove if not applicable -->
Fixes # (issue)
## Before submitting
<!-- If you're adding a new integration, include an integration test and
an example notebook showing its use! -->
## Who can review?
Community members can review the PR once tests pass. Tag
maintainers/contributors who might be interested:
<!-- For a quicker response, figure out the right person to tag with @
@hwchase17 - project lead
Tracing / Callbacks
- @agola11
Async
- @agola11
DataLoaders
- @eyurtsev
Models
- @hwchase17
- @agola11
Agents / Tools / Toolkits
- @vowelparrot
VectorStores / Retrievers / Memory
- @dev2049
-->
### Refactor the BaseTracer
- Remove the 'session' abstraction from the BaseTracer
- Rename 'RunV2' object(s) to be called 'Run' objects (Rename previous
Run objects to be RunV1 objects)
- Ditto for sessions: TracerSession*V2 -> TracerSession*
- Remove now deprecated conversion from v1 run objects to v2 run objects
in LangChainTracerV2
- Add conversion from v2 run objects to v1 run objects in V1 tracer
## Change Chain argument in client to accept a chain factory
The `run_over_dataset` functionality seeks to treat each iteration of an
example as an independent trial.
Chains have memory, so it's easier to permit this type of behavior if we
accept a factory method rather than the chain object directly.
There's still corner cases / UX pains people will likely run into, like:
- Caching may cause issues
- if memory is persisted to a shared object (e.g., same redis queue) ,
this could impact what is retrieved
- If we're running the async methods with concurrency using local
models, if someone naively instantiates the chain and loads each time,
it could lead to tons of disk I/O or OOM
### Add on_chat_message_start to callback manager and base tracer
Goal: trace messages directly to permit reloading as chat messages
(store in an integration-agnostic way)
Add an `on_chat_message_start` method. Fall back to `on_llm_start()` for
handlers that don't have it implemented.
Does so in a non-backwards-compat breaking way (for now)
