* ChatAnyscale was missing coercion to SecretStr for anyscale api key
* The model inherits from ChatOpenAI so it should not force the openai
api key to be secret str until openai model has the same changes
https://github.com/langchain-ai/langchain/issues/12841
Qdrant was incorrectly calculating the cosine similarity and returning
`0.0` for the best match, instead of `1.0`. Internally Qdrant returns a
cosine score from `-1.0` (worst match) to `1.0` (best match), and the
current formula reflects it.
Possibility to pass on_artifacts to a conversation. It can be then
achieved by adding this way:
```python
result = agent.run(
input=message.text,
metadata={
"on_artifact": CALLBACK_FUNCTION
},
)
```
Previously we treated trace_on_chain_group as a command to always start
tracing. This is unintuitive (makes the function do 2 things), and makes
it harder to toggle tracing
When you use a MultiQuery it might be useful to use the original query
as well as the newly generated ones to maximise the changes to retriever
the correct document. I haven't created an issue, it seems a very small
and easy thing.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
- **Description:** Correct number of elements in config list in
`batch()` and `abatch()` of `BaseLLM` in case `max_concurrency` is not
None.
- **Issue:** #12643
- **Twitter handle:** @akionux
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Zep now has the ability to search over chat history summaries. This PR
adds support for doing so. More here: https://blog.getzep.com/zep-v0-17/
@baskaryan @eyurtsev
…s present
<!-- Thank you for contributing to LangChain!
Replace this entire comment with:
- **Description:** a description of the change,
- **Issue:** the issue # it fixes (if applicable),
- **Dependencies:** any dependencies required for this change,
- **Tag maintainer:** for a quicker response, tag the relevant
maintainer (see below),
- **Twitter handle:** we announce bigger features on Twitter. If your PR
gets announced, and you'd like a mention, we'll gladly shout you out!
Please make sure your PR is passing linting and testing before
submitting. Run `make format`, `make lint` and `make test` to check this
locally.
See contribution guidelines for more information on how to write/run
tests, lint, etc:
https://github.com/langchain-ai/langchain/blob/master/.github/CONTRIBUTING.md
If you're adding a new integration, please include:
1. a test for the integration, preferably unit tests that do not rely on
network access,
2. an example notebook showing its use. It lives in `docs/extras`
directory.
If no one reviews your PR within a few days, please @-mention one of
@baskaryan, @eyurtsev, @hwchase17.
-->
### Enabling `device_map` in HuggingFacePipeline
For multi-gpu settings with large models, the
[accelerate](https://huggingface.co/docs/accelerate/usage_guides/big_modeling#using--accelerate)
library provides the `device_map` parameter to automatically distribute
the model across GPUs / disk.
The [Transformers
pipeline](3520e37e86/src/transformers/pipelines/__init__.py (L543))
enables users to specify `device` (or) `device_map`, and handles cases
(with warnings) when both are specified.
However, Langchain's HuggingFacePipeline only supports specifying
`device` when calling transformers which limits large models and
multi-gpu use-cases.
Additionally, the [default
value](8bd3ce59cd/libs/langchain/langchain/llms/huggingface_pipeline.py (L72))
of `device` is initialized to `-1` , which is incompatible with the
transformers pipeline when `device_map` is specified.
This PR addresses the addition of `device_map` as a parameter , and
solves the incompatibility of `device = -1` when `device_map` is also
specified.
An additional test has been added for this feature.
Additionally, some existing tests no longer work since
1. `max_new_tokens` has to be specified under `pipeline_kwargs` and not
`model_kwargs`
2. The GPT2 tokenizer raises a `ValueError: Pipeline with tokenizer
without pad_token cannot do batching`, since the `tokenizer.pad_token`
is `None` ([related
issue](https://github.com/huggingface/transformers/issues/19853) on the
transformers repo).
This PR handles fixing these tests as well.
Co-authored-by: Praveen Venkateswaran <praveen.venkateswaran@ibm.com>
[The python
spec](https://docs.python.org/3/reference/datamodel.html#object.__getattr__)
requires that `__getattr__` throw `AttributeError` for missing
attributes but there are several places throwing `ImportError` in the
current code base. This causes a specific problem with `hasattr` since
it calls `__getattr__` then looks only for `AttributeError` exceptions.
At present, calling `hasattr` on any of these modules will raise an
unexpected exception that most code will not handle as `hasattr`
throwing exceptions is not expected.
In our case this is triggered by an exception tracker (Airbrake) that
attempts to collect the version of all installed modules with code that
looks like: `if hasattr(mod, "__version__"):`. With `HEAD` this is
causing our exception tracker to fail on all exceptions.
I only changed instances of unknown attributes raising `ImportError` and
left instances of known attributes raising `ImportError`. It feels a
little weird but doesn't seem to break anything.
- **Description:** Use all Google search results data in SerpApi.com
wrapper instead of the first one only
- **Tag maintainer:** @hwchase17
_P.S. `libs/langchain/tests/integration_tests/utilities/test_serpapi.py`
are not executed during the `make test`._
<!-- Thank you for contributing to LangChain!
Replace this entire comment with:
- **Description:** a description of the change,
- **Issue:** the issue # it fixes (if applicable),
- **Dependencies:** any dependencies required for this change,
- **Tag maintainer:** for a quicker response, tag the relevant
maintainer (see below),
- **Twitter handle:** we announce bigger features on Twitter. If your PR
gets announced, and you'd like a mention, we'll gladly shout you out!
Please make sure your PR is passing linting and testing before
submitting. Run `make format`, `make lint` and `make test` to check this
locally.
See contribution guidelines for more information on how to write/run
tests, lint, etc:
https://github.com/langchain-ai/langchain/blob/master/.github/CONTRIBUTING.md
If you're adding a new integration, please include:
1. a test for the integration, preferably unit tests that do not rely on
network access,
2. an example notebook showing its use. It lives in `docs/extras`
directory.
If no one reviews your PR within a few days, please @-mention one of
@baskaryan, @eyurtsev, @hwchase17.
-->
It was passing in message instead of generation
<!-- Thank you for contributing to LangChain!
Replace this entire comment with:
- **Description:** a description of the change,
- **Issue:** the issue # it fixes (if applicable),
- **Dependencies:** any dependencies required for this change,
- **Tag maintainer:** for a quicker response, tag the relevant
maintainer (see below),
- **Twitter handle:** we announce bigger features on Twitter. If your PR
gets announced, and you'd like a mention, we'll gladly shout you out!
Please make sure your PR is passing linting and testing before
submitting. Run `make format`, `make lint` and `make test` to check this
locally.
See contribution guidelines for more information on how to write/run
tests, lint, etc:
https://github.com/langchain-ai/langchain/blob/master/.github/CONTRIBUTING.md
If you're adding a new integration, please include:
1. a test for the integration, preferably unit tests that do not rely on
network access,
2. an example notebook showing its use. It lives in `docs/extras`
directory.
If no one reviews your PR within a few days, please @-mention one of
@baskaryan, @eyurtsev, @hwchase17.
-->
* Restrict the chain to specific domains by default
* This is a breaking change, but it will fail loudly upon object
instantiation -- so there should be no silent errors for users
* Resolves CVE-2023-32786
* This is an opt-in feature, so users should be aware of risks if using
jinja2.
* Regardless we'll add sandboxing by default to jinja2 templates -- this
sandboxing is a best effort basis.
* Best strategy is still to make sure that jinja2 templates are only
loaded from trusted sources.
**Description:** Update `langchain.document_loaders.pdf.PyPDFLoader` to
store url in metadata (instead of a temporary file path) if user
provides a web path to a pdf
- **Issue:** Related to #7034; the reporter on that issue submitted a PR
updating `PyMuPDFParser` for this behavior, but it has unresolved merge
issues as of 20 Oct 2023 #7077
- In addition to `PyPDFLoader` and `PyMuPDFParser`, these other classes
in `langchain.document_loaders.pdf` exhibit similar behavior and could
benefit from an update: `PyPDFium2Loader`, `PDFMinerLoader`,
`PDFMinerPDFasHTMLLoader`, `PDFPlumberLoader` (I'm happy to contribute
to some/all of that, including assisting with `PyMuPDFParser`, if my
work is agreeable)
- The root cause is that the underlying pdf parser classes, e.g.
`langchain.document_loaders.parsers.pdf.PyPDFParser`, never receive
information about the url; the parsers receive a
`langchain.document_loaders.blob_loaders.blob`, which contains the pdf
contents and local file path, but not the url
- This update passes the web path directly to the parser since it's
minimally invasive and doesn't require further changes to maintain
existing behavior for local files... bigger picture, I'd consider
extending `blob` so that extra information like this can be
communicated, but that has much bigger implications on the codebase
which I think warrants maintainer input
- **Dependencies:** None
```python
# old behavior
>>> from langchain.document_loaders import PyPDFLoader
>>> loader = PyPDFLoader('https://arxiv.org/pdf/1706.03762.pdf')
>>> docs = loader.load()
>>> docs[0].metadata
{'source': '/var/folders/w2/zx77z1cs01s1thx5dhshkd58h3jtrv/T/tmpfgrorsi5/tmp.pdf', 'page': 0}
# new behavior
>>> from langchain.document_loaders import PyPDFLoader
>>> loader = PyPDFLoader('https://arxiv.org/pdf/1706.03762.pdf')
>>> docs = loader.load()
>>> docs[0].metadata
{'source': 'https://arxiv.org/pdf/1706.03762.pdf', 'page': 0}
```
- **Description:** #12273 's suggestion PR
Like other PDFLoader, loading pdf per each page and giving page
metadata.
- **Issue:** #12273
- **Twitter handle:** @blue0_0hope
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
This will allow you create the schema beforehand. The check was failing
and preventing importing into existing classes.
<!-- Thank you for contributing to LangChain!
Replace this entire comment with:
- **Description:** a description of the change,
- **Issue:** the issue # it fixes (if applicable),
- **Dependencies:** any dependencies required for this change,
- **Tag maintainer:** for a quicker response, tag the relevant
maintainer (see below),
- **Twitter handle:** we announce bigger features on Twitter. If your PR
gets announced, and you'd like a mention, we'll gladly shout you out!
Please make sure your PR is passing linting and testing before
submitting. Run `make format`, `make lint` and `make test` to check this
locally.
See contribution guidelines for more information on how to write/run
tests, lint, etc:
https://github.com/langchain-ai/langchain/blob/master/.github/CONTRIBUTING.md
If you're adding a new integration, please include:
1. a test for the integration, preferably unit tests that do not rely on
network access,
2. an example notebook showing its use. It lives in `docs/extras`
directory.
If no one reviews your PR within a few days, please @-mention one of
@baskaryan, @eyurtsev, @hwchase17.
-->
This input key was missed in the last update PR:
https://github.com/langchain-ai/langchain/pull/7391
The input/output formats are intended to be like this:
```
{"inputs": [<prompt>]}
{"outputs": [<output_text>]}
```
Best to review one commit at a time, since two of the commits are 100%
autogenerated changes from running `ruff format`:
- Install and use `ruff format` instead of black for code formatting.
- Output of `ruff format .` in the `langchain` package.
- Use `ruff format` in experimental package.
- Format changes in experimental package by `ruff format`.
- Manual formatting fixes to make `ruff .` pass.
I always take 20-30 seconds to re-discover where the
`convert_to_openai_function` wrapper lives in our codebase. Chat
langchain [has no
clue](https://smith.langchain.com/public/3989d687-18c7-4108-958e-96e88803da86/r)
what to do either. There's the older `create_openai_fn_chain` , but we
haven't been recommending it in LCEL. The example we show in the
[cookbook](https://python.langchain.com/docs/expression_language/how_to/binding#attaching-openai-functions)
is really verbose.
General function calling should be as simple as possible to do, so this
seems a bit more ergonomic to me (feel free to disagree). Another option
would be to directly coerce directly in the class's init (or when
calling invoke), if provided. I'm not 100% set against that. That
approach may be too easy but not simple. This PR feels like a decent
compromise between simple and easy.
```
from enum import Enum
from typing import Optional
from pydantic import BaseModel, Field
class Category(str, Enum):
"""The category of the issue."""
bug = "bug"
nit = "nit"
improvement = "improvement"
other = "other"
class IssueClassification(BaseModel):
"""Classify an issue."""
category: Category
other_description: Optional[str] = Field(
description="If classified as 'other', the suggested other category"
)
from langchain.chat_models import ChatOpenAI
llm = ChatOpenAI().bind_functions([IssueClassification])
llm.invoke("This PR adds a convenience wrapper to the bind argument")
# AIMessage(content='', additional_kwargs={'function_call': {'name': 'IssueClassification', 'arguments': '{\n "category": "improvement"\n}'}})
```
- Prefer lambda type annotations over inferred dict schema
- For sequences that start with RunnableAssign infer seq input type as
"input type of 2nd item in sequence - output type of runnable assign"
