This PR adds support for PygmalionAI's [Aphrodite
Engine](https://github.com/PygmalionAI/aphrodite-engine), based on
vLLM's attention mechanism. At the moment, this PR does not include
support for the API servers, but they will be added in a later PR.
The only dependency as of now is `aphrodite-engine==0.4.2`. We pin the
version to prevent breakage due to changes in the aphrodite-engine
library.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
- **Description:** Introducing an ability to work with the
[YandexGPT](https://cloud.yandex.com/en/services/yandexgpt) embeddings
models.
---------
Co-authored-by: Dmitry Tyumentsev <dmitry.tyumentsev@raftds.com>
- **Description:** Modify community chat model vertexai to handle png
and other image types encoded in base64
- **Dependencies:** added `import re` but no new dependencies.
This addresses a problem where the vertexai method
_parse_chat_history_gemini() was only recognizing image uris in jpeg
format. I made a simple change to cover other extension types.
- **Description:** The Qianfan SDK offers multiple authentication
methods, but in the `QianfanEndpoint` of Langchain, it currently only
supports authentication through AK and SK. In order to accommodate users
who wish to use alternative authentication methods, this pull request
makes AK and SK optional. This change should not impact existing users,
while allowing users to configure other authentication methods as per
the Qianfan SDK documentation.
- **Issue:** /
- **Dependencies:** No
- **Tag maintainer:** No
- **Twitter handle:**
Added Entry ID as a return value inside get_summaries_as_docs
- **Description:** Added the Entry ID as a return, so it's easier to
track the IDs of the papers that are being returned.
With the addition return of the entry ID in functions like
ArxivRetriever, it will be easier to reference the ID of the paper
itself.
- **Description:** Going forward, we have a own API `pip install
gradientai`. Therefore gradually removing the self-build packages in
llamaindex, haystack and langchain.
- **Issue:** None.
- **Dependencies:** `pip install gradientai`
- **Tag maintainer:** @michaelfeil
Very simple change in relation to the issue
https://github.com/langchain-ai/langchain/issues/14550
@baskaryan, @eyurtsev, @hwchase17.
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
**Description:** Added logic for re-calling the YandexGPT API in case of
an error
---------
Co-authored-by: Dmitry Tyumentsev <dmitry.tyumentsev@raftds.com>
Description: A new vector store Jaguar is being added. Class, test
scripts, and documentation is added.
Issue: None -- This is the first PR contributing to LangChain
Dependencies: This depends on "pip install -U jaguardb-http-client"
client http package
Tag maintainer: @baskaryan, @eyurtsev, @hwchase1
Twitter handle: @workbot
---------
Co-authored-by: JY <jyjy@jaguardb>
Co-authored-by: Bagatur <baskaryan@gmail.com>
Addded missed docstrings. Fixed inconsistency in docstrings.
**Note** CC @efriis
There were PR errors on
`langchain_experimental/prompt_injection_identifier/hugging_face_identifier.py`
But, I didn't touch this file in this PR! Can it be some cache problems?
I fixed this error.
- **Description:** added support for chat_history for Google
GenerativeAI (to actually use the `chat` API) plus since Gemini
currently doesn't have a support for SystemMessage, added support for it
only if a user provides additional `convert_system_message_to_human`
flag during model initialization (in this case, SystemMessage would be
prepanded to the first HumanMessage)
- **Issue:** #14710
- **Dependencies:** any dependencies required for this change,
- **Tag maintainer:** for a quicker response, tag the relevant
maintainer (see below),
- **Twitter handle:** lkuligin
---------
Co-authored-by: William FH <13333726+hinthornw@users.noreply.github.com>
- **Description:** This is addition to [my previous
PR](https://github.com/langchain-ai/langchain/pull/13930) with
improvements to flexibility allowing different models and notebook to
use ONNX runtime for faster speed. Since the last PR, [our
model](https://huggingface.co/laiyer/deberta-v3-base-prompt-injection)
got more than 660k downloads, and with the [public
benchmark](https://huggingface.co/spaces/laiyer/prompt-injection-benchmark)
showed much fewer false-positives than the previous one from deepset.
Additionally, on the ONNX runtime, it can be running 3x faster on the
CPU, which might be handy for builders using Langchain.
**Issue:** N/A
- **Dependencies:** N/A
- **Tag maintainer:** N/A
- **Twitter handle:** `@laiyer_ai`
Fixing issue - https://github.com/langchain-ai/langchain/issues/14494 to
avoid Kendra query ValidationException
<!-- Thank you for contributing to LangChain!
Replace this entire comment with:
- **Description:** Update kendra.py to avoid Kendra query
ValidationException,
- **Issue:** the issue
#https://github.com/langchain-ai/langchain/issues/14494,
- **Dependencies:** None,
- **Tag maintainer:** ,
- **Twitter handle:**
If no one reviews your PR within a few days, please @-mention one of
@baskaryan, @eyurtsev, @hwchase17.
-->
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
- **Description:**
- Add a break case to `text_splitter.py::split_text_on_tokens()` to
avoid unwanted item at the end of result.
- Add a testcase to enforce the behavior.
- **Issue:**
- #14649
- #5897
- **Dependencies:** n/a,
---
**Quick illustration of change:**
```
text = "foo bar baz 123"
tokenizer = Tokenizer(
chunk_overlap=3,
tokens_per_chunk=7
)
output = split_text_on_tokens(text=text, tokenizer=tokenizer)
```
output before change: `["foo bar", "bar baz", "baz 123", "123"]`
output after change: `["foo bar", "bar baz", "baz 123"]`
This is technically a breaking change because it'll switch out default
models from `text-davinci-003` to `gpt-3.5-turbo-instruct`, but OpenAI
is shutting off those endpoints on 1/4 anyways.
Feels less disruptive to switch out the default instead.
Gpt-3.5 sometimes calls with empty string arguments instead of `{}`
I'd assume it's because the typescript representation on their backend
makes it a bit ambiguous.
- **Description:** VertexAIEmbeddings performance improvements
- **Twitter handle:** @vladkol
## Improvements
- Dynamic batch size, starting from 250, lowering down to 5. Batch size
varies across regions.
Some regions support larger batches, and it significantly improves
performance.
When running large batches of texts in `us-central1`, performance gain
can be up to 3.5x.
The dynamic batching also makes sure every batch is below 20K token
limit.
- New model parameter `embeddings_type` that translates to `task_type`
parameter of the API. Newer model versions support [different embeddings
task
types](https://cloud.google.com/vertex-ai/docs/generative-ai/embeddings/get-text-embeddings#api_changes_to_models_released_on_or_after_august_2023).
Now that it's supported again for OAI chat models .
Shame this wouldn't include it in the `.invoke()` output though (it's
not included in the message itself). Would need to do a follow-up for
that to be the case
Fixed:
- `_agenerate` return value in the YandexGPT Chat Model
- duplicate line in the documentation
Co-authored-by: Dmitry Tyumentsev <dmitry.tyumentsev@raftds.com>
Builds out a developer documentation section in the docs
- Links it from contributing.md
- Adds an initial guide on how to contribute an integration
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Adds the option for `similarity_score_threshold` when using
`MongoDBAtlasVectorSearch` as a vector store retriever.
Example use:
```
vector_search = MongoDBAtlasVectorSearch.from_documents(...)
qa_retriever = vector_search.as_retriever(
search_type="similarity_score_threshold",
search_kwargs={
"score_threshold": 0.5,
}
)
qa = RetrievalQA.from_chain_type(
llm=OpenAI(),
chain_type="stuff",
retriever=qa_retriever,
)
docs = qa({"query": "..."})
```
I've tested this feature locally, using a MongoDB Atlas Cluster with a
vector search index.
… (#14723)
- **Description:** Minor updates per marketing requests. Namely, name
decisions (AI Foundation Models / AI Playground)
- **Tag maintainer:** @hinthornw
Do want to pass around the PR for a bit and ask a few more marketing
questions before merge, but just want to make sure I'm not working in a
vacuum. No major changes to code functionality intended; the PR should
be for documentation and only minor tweaks.
Note: QA model is a bit borked across staging/prod right now. Relevant
teams have been informed and are looking into it, and I'm placeholdered
the response to that of a working version in the notebook.
Co-authored-by: Vadim Kudlay <32310964+VKudlay@users.noreply.github.com>
