Previously the code was able to only handle a single level of nesting
for subgraphs in mermaid. This change adds support for arbitrary nesting
of subgraphs.
**Description:**
LLM will stop generating text even in the middle of a sentence if
`finish_reason` is `length` (for OpenAI) or `stop_reason` is
`max_tokens` (for Anthropic).
To obtain longer outputs from LLM, we should call the message generation
API multiple times and merge the results into the text to circumvent the
API's output token limit.
The extra line breaks forced by the `merge_message_runs` function when
seamlessly merging messages can be annoying, so I added the option to
specify the chunk separator.
**Issue:**
No corresponding issues.
**Dependencies:**
No dependencies required.
**Twitter handle:**
@hanama_chem
https://x.com/hanama_chem
---------
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
[langchain_core] Fix UnionType type var replacement
- Added types.UnionType to typing.Union mapping
Type replacement cause `TypeError: 'type' object is not subscriptable`
if any of union type comes as function `_py_38_safe_origin` return
`types.UnionType` instead of `typing.Union`
```python
>>> from types import UnionType
>>> from typing import Union, get_origin
>>> type_ = get_origin(str | None)
>>> type_
<class 'types.UnionType'>
>>> UnionType[(str, None)]
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: 'type' object is not subscriptable
>>> Union[(str, None)]
typing.Optional[str]
```
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
This PR deprecates the beta upsert APIs in vectorstore.
We'll introduce them in a V2 abstraction instead to keep the existing
vectorstore implementations lighter weight.
The main problem with the existing APIs is that it's a bit more
challenging to
implement the correct behavior w/ respect to IDs since ID can be present
in
both the function signature and as an optional attribute on the document
object.
But VectorStores that pass the standard tests should have implemented
the semantics properly!
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
This PR gets rid `root_validators(allow_reuse=True)` logic used in
EdenAI Tool in preparation for pydantic 2 upgrade.
- add another test to secret_from_env_factory
Add a utility that can be used as a default factory
The goal will be to start migrating from of the pydantic models to use
`from_env` as a default factory if possible.
```python
from pydantic import Field, BaseModel
from langchain_core.utils import from_env
class Foo(BaseModel):
name: str = Field(default_factory=from_env('HELLO'))
```
- **Description:** This includes Pydantic field metadata in
`_create_subset_model_v2` so that it gets included in the final
serialized form that get sent out.
- **Issue:** #25031
- **Dependencies:** n/a
- **Twitter handle:** @gramliu
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
This PR adds a minimal document indexer abstraction.
The goal of this abstraction is to allow developers to create custom
retrievers that also have a standard indexing API and allow updating the
document content in them.
The abstraction comes with a test suite that can verify that the indexer
implements the correct semantics.
This is an iteration over a previous PRs
(https://github.com/langchain-ai/langchain/pull/24364). The main
difference is that we're sub-classing from BaseRetriever in this
iteration and as so have consolidated the sync and async interfaces.
The main problem with the current design is that runt time search
configuration has to be specified at init rather than provided at run
time.
We will likely resolve this issue in one of the two ways:
(1) Define a method (`get_retriever`) that will allow creating a
retriever at run time with a specific configuration.. If we do this, we
will likely break the subclass on BaseRetriever
(2) Generalize base retriever so it can support structured queries
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
This PR introduces a module with some helper utilities for the pydantic
1 -> 2 migration.
They're meant to be used in the following way:
1) Use the utility code to get unit tests pass without requiring
modification to the unit tests
2) (If desired) upgrade the unit tests to match pydantic 2 output
3) (If desired) stop using the utility code
Currently, this module contains a way to map `schema()` generated by
pydantic 2 to (mostly) match the output from pydantic v1.
Add compatibility for pydantic 2 for a utility function.
This will help push some small changes to master, so they don't have to
be kept track of on a separate branch.
supports following UX
```python
class SubTool(TypedDict):
"""Subtool docstring"""
args: Annotated[Dict[str, Any], {}, "this does bar"]
class Tool(TypedDict):
"""Docstring
Args:
arg1: foo
"""
arg1: str
arg2: Union[int, str]
arg3: Optional[List[SubTool]]
arg4: Annotated[Literal["bar", "baz"], ..., "this does foo"]
arg5: Annotated[Optional[float], None]
```
- can parse google style docstring
- can use Annotated to specify default value (second arg)
- can use Annotated to specify arg description (third arg)
- can have nested complex types
Anthropic models (including via Bedrock and other cloud platforms)
accept a status/is_error attribute on tool messages/results
(specifically in `tool_result` content blocks for Anthropic API). Adding
a ToolMessage.status attribute so that users can set this attribute when
using those models
This PR proposes to create a rate limiter in the chat model directly,
and would replace: https://github.com/langchain-ai/langchain/pull/21992
It resolves most of the constraints that the Runnable rate limiter
introduced:
1. It's not annoying to apply the rate limiter to existing code; i.e.,
possible to roll out the change at the location where the model is
instantiated,
rather than at every location where the model is used! (Which is
necessary
if the model is used in different ways in a given application.)
2. batch rate limiting is enforced properly
3. the rate limiter works correctly with streaming
4. the rate limiter is aware of the cache
5. The rate limiter can take into account information about the inputs
into the
model (we can add optional inputs to it down-the road together with
outputs!)
The only downside is that information will not be properly reflected in
tracing
as we don't have any metadata evens about a rate limiter. So the total
time
spent on a model invocation will be:
* time spent waiting for the rate limiter
* time spend on the actual model request
## Example
```python
from langchain_core.rate_limiters import InMemoryRateLimiter
from langchain_groq import ChatGroq
groq = ChatGroq(rate_limiter=InMemoryRateLimiter(check_every_n_seconds=1))
groq.invoke('hello')
```
### Description
* support asynchronous in InMemoryVectorStore
* since embeddings might be possible to call asynchronously, ensure that
both asynchronous and synchronous functions operate correctly.
This PR introduces the following Runnables:
1. BaseRateLimiter: an abstraction for specifying a time based rate
limiter as a Runnable
2. InMemoryRateLimiter: Provides an in-memory implementation of a rate
limiter
## Example
```python
from langchain_core.runnables import InMemoryRateLimiter, RunnableLambda
from datetime import datetime
foo = InMemoryRateLimiter(requests_per_second=0.5)
def meow(x):
print(datetime.now().strftime("%H:%M:%S.%f"))
return x
chain = foo | meow
for _ in range(10):
print(chain.invoke('hello'))
```
Produces:
```
17:12:07.530151
hello
17:12:09.537932
hello
17:12:11.548375
hello
17:12:13.558383
hello
17:12:15.568348
hello
17:12:17.578171
hello
17:12:19.587508
hello
17:12:21.597877
hello
17:12:23.607707
hello
17:12:25.617978
hello
```
## Interface
The rate limiter uses the following interface for acquiring a token:
```python
class BaseRateLimiter(Runnable[Input, Output], abc.ABC):
@abc.abstractmethod
def acquire(self, *, blocking: bool = True) -> bool:
"""Attempt to acquire the necessary tokens for the rate limiter.```
```
The flag `blocking` has been added to the abstraction to allow
supporting streaming (which is easier if blocking=False).
## Limitations
- The rate limiter is not designed to work across different processes.
It is an in-memory rate limiter, but it is thread safe.
- The rate limiter only supports time-based rate limiting. It does not
take into account the size of the request or any other factors.
- The current implementation does not handle streaming inputs well and
will consume all inputs even if the rate limit has been reached. Better
support for streaming inputs will be added in the future.
- When the rate limiter is combined with another runnable via a
RunnableSequence, usage of .batch() or .abatch() will only respect the
average rate limit. There will be bursty behavior as .batch() and
.abatch() wait for each step to complete before starting the next step.
One way to mitigate this is to use batch_as_completed() or
abatch_as_completed().
## Bursty behavior in `batch` and `abatch`
When the rate limiter is combined with another runnable via a
RunnableSequence, usage of .batch() or .abatch() will only respect the
average rate limit. There will be bursty behavior as .batch() and
.abatch() wait for each step to complete before starting the next step.
This becomes a problem if users are using `batch` and `abatch` with many
inputs (e.g., 100). In this case, there will be a burst of 100 inputs
into the batch of the rate limited runnable.
1. Using a RunnableBinding
The API would look like:
```python
from langchain_core.runnables import InMemoryRateLimiter, RunnableLambda
rate_limiter = InMemoryRateLimiter(requests_per_second=0.5)
def meow(x):
return x
rate_limited_meow = RunnableLambda(meow).with_rate_limiter(rate_limiter)
```
2. Another option is to add some init option to RunnableSequence that
changes `.batch()` to be depth first (e.g., by delegating to
`batch_as_completed`)
```python
RunnableSequence(first=rate_limiter, last=model, how='batch-depth-first')
```
Pros: Does not require Runnable Binding
Cons: Feels over-complicated
