Todo:
- [x] Connection options (cloud, localhost url, es_connection) support
- [x] Logging support
- [x] Customisable field support
- [x] Distance Similarity support
- [x] Metadata support
- [x] Metadata Filter support
- [x] Retrieval Strategies
- [x] Approx
- [x] Approx with Hybrid
- [x] Exact
- [x] Custom
- [x] ELSER (excluding hybrid as we are working on RRF support)
- [x] integration tests
- [x] Documentation
👋 this is a contribution to improve Elasticsearch integration with
Langchain. Its based loosely on the changes that are in master but with
some notable changes:
## Package name & design improvements
The import name is now `ElasticsearchStore`, to aid discoverability of
the VectorStore.
```py
## Before
from langchain.vectorstores.elastic_vector_search import ElasticVectorSearch, ElasticKnnSearch
## Now
from langchain.vectorstores.elasticsearch import ElasticsearchStore
```
## Retrieval Strategy support
Before we had a number of classes, depending on the strategy you wanted.
`ElasticKnnSearch` for approx, `ElasticVectorSearch` for exact / brute
force.
With `ElasticsearchStore` we have retrieval strategies:
### Approx Example
Default strategy for the vast majority of developers who use
Elasticsearch will be inferring the embeddings from outside of
Elasticsearch. Uses KNN functionality of _search.
```py
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts,
FakeEmbeddings(),
es_url="http://localhost:9200",
index_name="sample-index"
)
output = docsearch.similarity_search("foo", k=1)
```
### Approx, with hybrid
Developers who want to search, using both the embedding and the text
bm25 match. Its simple to enable.
```py
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts,
FakeEmbeddings(),
es_url="http://localhost:9200",
index_name="sample-index",
strategy=ElasticsearchStore.ApproxRetrievalStrategy(hybrid=True)
)
output = docsearch.similarity_search("foo", k=1)
```
### Approx, with `query_model_id`
Developers who want to infer within Elasticsearch, using the model
loaded in the ml node.
This relies on the developer to setup the pipeline and index if they
wish to embed the text in Elasticsearch. Example of this in the test.
```py
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts,
FakeEmbeddings(),
es_url="http://localhost:9200",
index_name="sample-index",
strategy=ElasticsearchStore.ApproxRetrievalStrategy(
query_model_id="sentence-transformers__all-minilm-l6-v2"
),
)
output = docsearch.similarity_search("foo", k=1)
```
### I want to provide my own custom Elasticsearch Query
You might want to have more control over the query, to perform
multi-phase retrieval such as LTR, linearly boosting on document
parameters like recently updated or geo-distance. You can do this with
`custom_query_fn`
```py
def my_custom_query(query_body: dict, query: str) -> dict:
return {"query": {"match": {"text": {"query": "bar"}}}}
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts, FakeEmbeddings(), **elasticsearch_connection, index_name=index_name
)
docsearch.similarity_search("foo", k=1, custom_query=my_custom_query)
```
### Exact Example
Developers who have a small dataset in Elasticsearch, dont want the cost
of indexing the dims vs tradeoff on cost at query time. Uses
script_score.
```py
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts,
FakeEmbeddings(),
es_url="http://localhost:9200",
index_name="sample-index",
strategy=ElasticsearchStore.ExactRetrievalStrategy(),
)
output = docsearch.similarity_search("foo", k=1)
```
### ELSER Example
Elastic provides its own sparse vector model called ELSER. With these
changes, its really easy to use. The vector store creates a pipeline and
index thats setup for ELSER. All the developer needs to do is configure,
ingest and query via langchain tooling.
```py
texts = ["foo", "bar", "baz"]
docsearch = ElasticsearchStore.from_texts(
texts,
FakeEmbeddings(),
es_url="http://localhost:9200",
index_name="sample-index",
strategy=ElasticsearchStore.SparseVectorStrategy(),
)
output = docsearch.similarity_search("foo", k=1)
```
## Architecture
In future, we can introduce new strategies and allow us to not break bwc
as we evolve the index / query strategy.
## Credit
On release, could you credit @elastic and @phoey1 please? Thank you!
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Updated prompts for the MultiOn toolkit for better functionality
- Non-blocking but good to have it merged to improve the overall
performance for the toolkit
@hinthornw @hwchase17
---------
Co-authored-by: Naman Garg <ngarg3@binghamton.edu>
Add ability to track langchain usage for Rockset. Rockset's new python
client allows setting this. To prevent old clients from failing, it
ignore if setting throws exception (we can't track old versions)
Tested locally with old and new Rockset python client
cc @baskaryan
2 things:
- Implement the private method rather than the public one so callbacks
are handled properly
- Add search_kwargs (Open to not adding this if we are trying to
deprecate this UX but seems like as a user i'd assume similar args to
the vector store retriever. In fact some may assume this implements the
same interface but I'm not dealing with that here)
-
First of a few PRs to add full compatibility to both pydantic v1 and v2.
This PR creates pydantic v1 namespace and adds it to sys.modules.
Upcoming changes:
1. Handle `openapi-schema-pydantic = "^1.2"` and dependent chains/tools
2. bump dependencies to versions that are cross compatible for pydantic
or remove them (see below)
3. Add tests to github workflows to test with pydantic v1 and v2
**Dependencies**
From a quick look (could be wrong since was done manually)
**dependencies pinning pydantic below 2** (some of these can be bumped
to newer versions are provide cross-compatible code)
anthropic
bentoml
confection
fastapi
langsmith
octoai-sdk
openapi-schema-pydantic
qdrant-client
spacy
steamship
thinc
zep-python
Unpinned
marqo (*)
nomic (*)
xinference(*)
## Description:
Sets default values for `client` and `model` attributes in the
BaseOpenAI class to fix Pylance Typing issue.
- Issue: #9182.
- Twitter handle: @evanmschultz
Adds [DeepSparse](https://github.com/neuralmagic/deepsparse) as an LLM
backend. DeepSparse supports running various open-source sparsified
models hosted on [SparseZoo](https://sparsezoo.neuralmagic.com/) for
performance gains on CPUs.
Twitter handles: @mgoin_ @neuralmagic
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
# Added SmartGPT workflow by providing SmartLLM wrapper around LLMs
Edit:
As @hwchase17 suggested, this should be a chain, not an LLM. I have
adapted the PR.
It is used like this:
```
from langchain.prompts import PromptTemplate
from langchain.chains import SmartLLMChain
from langchain.chat_models import ChatOpenAI
hard_question = "I have a 12 liter jug and a 6 liter jug. I want to measure 6 liters. How do I do it?"
hard_question_prompt = PromptTemplate.from_template(hard_question)
llm = ChatOpenAI(model_name="gpt-4")
prompt = PromptTemplate.from_template(hard_question)
chain = SmartLLMChain(llm=llm, prompt=prompt, verbose=True)
chain.run({})
```
Original text:
Added SmartLLM wrapper around LLMs to allow for SmartGPT workflow (as in
https://youtu.be/wVzuvf9D9BU). SmartLLM can be used wherever LLM can be
used. E.g:
```
smart_llm = SmartLLM(llm=OpenAI())
smart_llm("What would be a good company name for a company that makes colorful socks?")
```
or
```
smart_llm = SmartLLM(llm=OpenAI())
prompt = PromptTemplate(
input_variables=["product"],
template="What is a good name for a company that makes {product}?",
)
chain = LLMChain(llm=smart_llm, prompt=prompt)
chain.run("colorful socks")
```
SmartGPT consists of 3 steps:
1. Ideate - generate n possible solutions ("ideas") to user prompt
2. Critique - find flaws in every idea & select best one
3. Resolve - improve upon best idea & return it
Fixes#4463
## Who can review?
Community members can review the PR once tests pass. Tag
maintainers/contributors who might be interested:
- @hwchase17
- @agola11
Twitter: [@UmerHAdil](https://twitter.com/@UmerHAdil) | Discord:
RicChilligerDude#7589
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
# Ensure deployment_id is set to provided deployment, required for Azure
OpenAI.
---------
Co-authored-by: Lucas Pickup <lupickup@microsoft.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
This commit adds the LangChain utility which allows for the real-time
retrieval of cryptocurrency exchange prices. With LangChain, users can
easily access up-to-date pricing information by running the command
".run(from_currency, to_currency)". This new feature provides a
convenient way to stay informed on the latest exchange rates and make
informed decisions when trading crypto.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Description: Adds the ArcGISLoader class to
`langchain.document_loaders`
- Allows users to load data from ArcGIS Online, Portal, and similar
- Users can authenticate with `arcgis.gis.GIS` or retrieve public data
anonymously
- Uses the `arcgis.features.FeatureLayer` class to retrieve the data
- Defines the most relevant keywords arguments and accepts `**kwargs`
- Dependencies: Using this class requires `arcgis` and, optionally,
`bs4.BeautifulSoup`.
Tagging maintainers:
- DataLoaders / VectorStores / Retrievers: @rlancemartin, @eyurtsev
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Formatted docstrings from different formats to consistent format, lile:
>Loads processed docs from Docugami.
"Load from `Docugami`."
>Loader that uses Unstructured to load HTML files.
"Load `HTML` files using `Unstructured`."
>Load documents from a directory.
"Load from a directory."
- `Load` - no `Loads`
- DocumentLoader always loads Documents, so no more
"documents/docs/texts/ etc"
- integrated systems and APIs enclosed in backticks,
As stated in the title the SVM retriever discarded the metadata of
passed in docs. This code fixes that. I also added one unit test that
should test that.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Description: Added a new use case category called "Web Scraping", and
a tutorial to scrape websites using OpenAI Functions Extraction chain to
the docs.
- Tag maintainer:@baskaryan @hwchase17 ,
- Twitter handle: https://www.linkedin.com/in/haiphunghiem/ (I'm on
LinkedIn mostly)
---------
Co-authored-by: Lance Martin <lance@langchain.dev>
This change updates the central utility class to recognize a Redis
cluster server after connection and returns an new cluster aware Redis
client. The "normal" Redis client would not be able to talk to a cluster
node because keys might be stored on other shards of the Redis cluster
and therefor not readable or writable.
With this patch clients do not need to know what Redis server it is,
they just connect though the same API calls for standalone and cluster
server.
There are no dependencies added due to this MR.
Remark - with current redis-py client library (4.6.0) a cluster cannot
be used as VectorStore. It can be used for other use-cases. There is a
bug / missing feature(?) in the Redis client breaking the VectorStore
implementation. I opened an issue at the client library too
(redis/redis-py#2888) to fix this. As soon as this is fixed in
`redis-py` library it should be usable there too.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
This PR introduces [Label Studio](https://labelstud.io/) integration
with LangChain via `LabelStudioCallbackHandler`:
- sending data to the Label Studio instance
- labeling dataset for supervised LLM finetuning
- rating model responses
- tracking and displaying chat history
- support for custom data labeling workflow
### Example
```
chat_llm = ChatOpenAI(callbacks=[LabelStudioCallbackHandler(mode="chat")])
chat_llm([
SystemMessage(content="Always use emojis in your responses."),
HumanMessage(content="Hey AI, how's your day going?"),
AIMessage(content="🤖 I don't have feelings, but I'm running smoothly! How can I help you today?"),
HumanMessage(content="I'm feeling a bit down. Any advice?"),
AIMessage(content="🤗 I'm sorry to hear that. Remember, it's okay to seek help or talk to someone if you need to. 💬"),
HumanMessage(content="Can you tell me a joke to lighten the mood?"),
AIMessage(content="Of course! 🎭 Why did the scarecrow win an award? Because he was outstanding in his field! 🌾"),
HumanMessage(content="Haha, that was a good one! Thanks for cheering me up."),
AIMessage(content="Always here to help! 😊 If you need anything else, just let me know."),
HumanMessage(content="Will do! By the way, can you recommend a good movie?"),
])
```
<img width="906" alt="image"
src="https://github.com/langchain-ai/langchain/assets/6087484/0a1cf559-0bd3-4250-ad96-6e71dbb1d2f3">
### Dependencies
- [label-studio](https://pypi.org/project/label-studio/)
- [label-studio-sdk](https://pypi.org/project/label-studio-sdk/)
https://twitter.com/labelstudiohq
---------
Co-authored-by: nik <nik@heartex.net>
As of the recent PR at #9043, after some testing we've realised that the
default values were not being used for `api_key` and `api_url`. Besides
that, the default for `api_key` was set to `argilla.apikey`, but since
the default values are intended for people using the Argilla Quickstart
(easy to run and setup), the defaults should be instead `owner.apikey`
if using Argilla 1.11.0 or higher, or `admin.apikey` if using a lower
version of Argilla.
Additionally, we've removed the f-string replacements from the
docstrings.
---------
Co-authored-by: Gabriel Martin <gabriel@argilla.io>
This MR corrects the IndexError arising in prep_prompts method when no
documents are returned from a similarity search.
Fixes#1733
Co-authored-by: Sam Groenjes <sam.groenjes@darkwolfsolutions.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
### Description:
`ConversationBufferTokenMemory` should have a simple way of returning
the conversation messages as a string.
Previously to complete this, you would only have the option to return
memory as an array through the buffer method and call
`get_buffer_string` by importing it from `langchain.schema`, or use the
`load_memory_variables` method and key into `self.memory_key`.
### Maintainer
@hwchase17
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Now that we accept any runnable or arbitrary function to evaluate, we
don't always look up the input keys. If an evaluator requires
references, we should try to infer if there's one key present. We only
have delayed validation here but it's better than nothing
- **Description**: [BagelDB](bageldb.ai) a collaborative vector
database. Integrated the bageldb PyPi package with langchain with
related tests and code.
- **Issue**: Not applicable.
- **Dependencies**: `betabageldb` PyPi package.
- **Tag maintainer**: @rlancemartin, @eyurtsev, @baskaryan
- **Twitter handle**: bageldb_ai (https://twitter.com/BagelDB_ai)
We ran `make format`, `make lint` and `make test` locally.
Followed the contribution guideline thoroughly
https://github.com/hwchase17/langchain/blob/master/.github/CONTRIBUTING.md
---------
Co-authored-by: Towhid1 <nurulaktertowhid@gmail.com>
Description: updated BabyAGI examples and experimental to append the
iteration to the result id to fix error storing data to vectorstore.
Issue: 7445
Dependencies: no
Tag maintainer: @eyurtsev
This fix worked for me locally. Happy to take some feedback and iterate
on a better solution. I was considering appending a uuid instead but
didn't want to over complicate the example.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Add convenience methods to `ConversationBufferMemory` and
`ConversationBufferWindowMemory` to get buffer either as messages or as
string.
Helps when `return_messages` is set to `True` but you want access to the
messages as a string, and vice versa.
@hwchase17
One use case: Using a `MultiPromptRouter` where `default_chain` is
`ConversationChain`, but destination chains are `LLMChains`. Injecting
chat memory into prompts for destination chains prints a stringified
`List[Messages]` in the prompt, which creates a lot of noise. These
convenience methods allow caller to choose either as needed.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Description: Due to some issue on the test, this is a separate PR with
the test for #8502
Tag maintainer: @rlancemartin
---------
Co-authored-by: Lance Martin <lance@langchain.dev>
Co-authored-by: Bagatur <baskaryan@gmail.com>
Current regex only extracts agent's action between '` ``` ``` `', this
commit will extract action between both '` ```json ``` `' and '` ``` ```
`'
This is very similar to #7511
Co-authored-by: zjl <junlinzhou@yzbigdata.com>
## Description
This PR adds the `aembed_query` and `aembed_documents` async methods for
improving the embeddings generation for large documents. The
implementation uses asyncio tasks and gather to achieve concurrency as
there is no bedrock async API in boto3.
### Maintainers
@agola11
@aarora79
### Open questions
To avoid throttling from the Bedrock API, should there be an option to
limit the concurrency of the calls?
I was initially confused weather to use create_vectorstore_agent or
create_vectorstore_router_agent due to lack of documentation so I
created a simple documentation for each of the function about their
different usecase.
Replace this comment with:
- Description: Added the doc_strings in create_vectorstore_agent and
create_vectorstore_router_agent to point out the difference in their
usecase
- Tag maintainer: @rlancemartin, @eyurtsev
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Hi @agola11, or whoever is reviewing this PR 😄
## What's in this PR?
As of the latest Argilla release, we'll change and refactor some things
to make some workflows easier, one of those is how everything's pushed
to Argilla, so that now there's no need to call `push_to_argilla` over a
`FeedbackDataset` when either `push_to_argilla` is called for the first
time, or `from_argilla` is called; among others.
We also add some class variables to make sure those are easy to update
in case we update those internally in the future, also to make the
`warnings.warn` message lighter from the code view.
P.S. Regarding the Twitter/X mention feel free to do so at either
https://twitter.com/argilla_io or https://twitter.com/alvarobartt, or
both if applicable, otherwise, just the first Twitter/X handle.
## Description:
This PR adds the Titan Takeoff Server to the available LLMs in
LangChain.
Titan Takeoff is an inference server created by
[TitanML](https://www.titanml.co/) that allows you to deploy large
language models locally on your hardware in a single command. Most
generative model architectures are included, such as Falcon, Llama 2,
GPT2, T5 and many more.
Read more about Titan Takeoff here:
-
[Blog](https://medium.com/@TitanML/introducing-titan-takeoff-6c30e55a8e1e)
- [Docs](https://docs.titanml.co/docs/titan-takeoff/getting-started)
#### Testing
As Titan Takeoff runs locally on port 8000 by default, no network access
is needed. Responses are mocked for testing.
- [x] Make Lint
- [x] Make Format
- [x] Make Test
#### Dependencies
No new dependencies are introduced. However, users will need to install
the titan-iris package in their local environment and start the Titan
Takeoff inferencing server in order to use the Titan Takeoff
integration.
Thanks for your help and please let me know if you have any questions.
cc: @hwchase17 @baskaryan
- Description: Fixes an issue with Metaphor Search Tool throwing when
missing keys in API response.
- Issue: #9048
- Tag maintainer: @hinthornw @hwchase17
- Twitter handle: @pelaseyed
This PR adds the ability to temporarily cache or persistently store
embeddings.
A notebook has been included showing how to set up the cache and how to
use it with a vectorstore.
<!-- Thank you for contributing to LangChain!
Replace this 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 you're PR is passing linting and testing before
submitting. Run `make format`, `make lint` and `make test` to check this
locally.
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.
Maintainer responsibilities:
- General / Misc / if you don't know who to tag: @baskaryan
- DataLoaders / VectorStores / Retrievers: @rlancemartin, @eyurtsev
- Models / Prompts: @hwchase17, @baskaryan
- Memory: @hwchase17
- Agents / Tools / Toolkits: @hinthornw
- Tracing / Callbacks: @agola11
- Async: @agola11
If no one reviews your PR within a few days, feel free to @-mention the
same people again.
See contribution guidelines for more information on how to write/run
tests, lint, etc:
https://github.com/hwchase17/langchain/blob/master/.github/CONTRIBUTING.md
-->
FileCallbackHandler cannot handle some language, for example: Chinese.
Open file using UTF-8 encoding can fix it.
@agola11
**Issue**: #6919
**Dependencies**: NO dependencies,
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
DirectoryLoader can now return a random sample of files in a directory.
Parameters added are:
sample_size
randomize_sample
sample_seed
@rlancemartin, @eyurtsev
---------
Co-authored-by: Andrew Oseen <amovfx@protonmail.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Description: Allow GoogleDriveLoader to handle empty spreadsheets
- Issue: Currently GoogleDriveLoader will crash if it tries to load a
spreadsheet with an empty sheet
- Dependencies: n/a
- Tag maintainer: @rlancemartin, @eyurtsev
This pull request aims to ensure that the `OpenAICallbackHandler` can
properly calculate the total cost for Azure OpenAI chat models. The
following changes have resolved this issue:
- The `model_name` has been added to the ChatResult llm_output. Without
this, the default values of `gpt-35-turbo` were applied. This was
causing the total cost for Azure OpenAI's GPT-4 to be significantly
inaccurate.
- A new parameter `model_version` has been added to `AzureChatOpenAI`.
Azure does not include the model version in the response. With the
addition of `model_name`, this is not a significant issue for GPT-4
models, but it's an issue for GPT-3.5-Turbo. Version 0301 (default) of
GPT-3.5-Turbo on Azure has a flat rate of 0.002 per 1k tokens for both
prompt and completion. However, version 0613 introduced a split in
pricing for prompt and completion tokens.
- The `OpenAICallbackHandler` implementation has been updated with the
proper model names, versions, and cost per 1k tokens.
Unit tests have been added to ensure the functionality works as
expected; the Azure ChatOpenAI notebook has been updated with examples.
Maintainers: @hwchase17, @baskaryan
Twitter handle: @jjczopek
---------
Co-authored-by: Jerzy Czopek <jerzy.czopek@avanade.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
