deepset-ai / haystack-core-integrations Goto Github PK

Summary and motivation

As a user, I'd like to use models on Amazon Bedrock in my Haystack 2.0 pipelines. This was brought up in deepset-ai/haystack#6545

Detailed design

A new generator class for Amazon Bedrock with the same features as the 1.x version, which was implemented as an invocation layer. All currently available models on Bedrock should be supported and streaming should be supported for all models that support streaming. For example, we should be able to load Llama 2 Chat 13B model:

from amazon_bedrock_haystack.generators import AmazonBedrockGenerator
generator = AmazonBedrockGenerator(model_name_or_path="meta.llama2-13b-chat-v1")

Related PRs for the support in Haystack 1.x were deepset-ai/haystack#6226 and deepset-ai/haystack#6406

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• deepset-ai/haystack-integrations#104
• The integration has been listed in the Inventory section of this repo README
• #173
• #226

Summary and motivation

Add the Jina AI embedding functionality to Haystack

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations
• The integration has been listed in the Inventory section of this repo README
• https://github.com/deepset-ai/devrel-board/issues/232
• The feature was announced through social media

Summary and motivation

Similar to 1.x we'd like to have a document store for MongoDB's Atlas.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch:
  • Docstore itself: #413
  • First retriever: #427
  • Filtering:
• Docs are published at https://docs.haystack.deepset.ai/: Draft at https://www.notion.so/deepsetai/MongoDBAtlasDocumentStore-WIP-e2b17f3079804ac6a69384789f83b2ce
• There is a Github workflow running the tests for the integration nightly and at every PR: #413
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated: #413
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations: deepset-ai/haystack-integrations#180
• The integration has been listed in the Inventory section of this repo README: #413
• There is an example available to demonstrate the feature: #413
• The feature was announced through social media

• The current logic to understand if the user is using the hosted API is not great (here)

• The Unstructured SaaS API was recently announced. This involves several changes. (IDK if we support it as of now)

In conclusion, we can make our implementation more robust and flexible, so that users can use:

• the free API via Docker
• the free hosted API
• the SAAS API

Summary and motivation

DataStax Astra DB is a serverless vector database that’s perfect for managing mission-critical AI workloads. It’s built on Apache Cassandra®, making Astra DB a highly scalable, reliable database technology. This makes it a powerful all-in-one data storage solution, ideal for Generative AI projects.

Detailed design

Implement the DocumentStore protocol and provide a specific documents retriever.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch #144
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• #164
• The integration has been listed in the Inventory section of this repo README
• #163
• The feature was announced through social media

• replace Document.text with Document.content
• replace Document.metadata with Document.meta

Summary and motivation

We want to support Weaviate as a Document Store in Haystack 2.x, much like we did for Haystack 1.x

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• The feature was announced through social media

When they're published, we should link them from the "Google Vertex AI integration" page

Is your feature request related to a problem? Please describe.
We should support Azure Open AI services. It's a widely used service that is particularly relevant for professional production use.

See also: deepset-ai/haystack#6620

Describe the solution you'd like
Similar to 1.x we need a component that supports OpenAI models hosted on Azure.

Describe alternatives you've considered
None

Problem:
I want to retrieve all relevant (similar) documents from the ElasticsearchDocumentStore based on the _score using the EmbeddingRetriever (I am not using the Reader). Prior to the search, I don't know how many relevant Documents exist. To make sure, that I retrieve all relevant entries from the ElasticsearchDocumentStore I need to set top_k=10000 or higher and filter the results afterwards - only taking documents with a _score higher than x. Retrieving this many documents takes several seconds.

Solution
Filtering your query results by a minimum score value is already implemented in the Python Elasticsearch client. You could add another parameter (min_score) similar to tok_k and add it to the body that you use in client.search(). See my example:

body = { "size": top_k, "min_score": min_score, "query": self._get_vector_similarity_query(query_emb, top_k) }

I changed the body form the function def query_by_embedding(...) from the file haystack/document_stores/elasticsearch.py. Now the results contain only documents that have a _score higher than min_score.

Additional context
In case the user wants to filter the results by the cosine similarity metric the min_score parameter needs to be scaled appropriately before using it in the body.

Hi , wondering if you will support supabse in near future

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

Describe the bug

We recently moved (#103) package versioning from hardcoded strings in the __about__.py file to git tags using setuptools_scm through hatch.

Problem is setuptools_scm splits the tag using - and having hyphens in the string prepending the version in the tag name, for example integrations/google_vertex-v0.0.1 confuses the plugin.

There isn't an easy fix, so I propose the following workaround:

• Rename all the tags integrations/google-vertex-vXXX to integrations/google_vertex-vXXX
• Rename all the tags integrations/instructor-embedders-vXXX to integrations/instructor_embedders-vXXX
• Push the new tags, CI will attempt to rebuild the packages but will fail because the path google_vertex doesn't exist
• Rename the folder integrations/google-vertex to integrations/google_vertex #114
• Rename the folder integrations/instructor-embedders to integrations/instructor_embedders #114
• Enforce the new naming convention with a job in the CI #119

The workaround won't affect the name of the package on PyPI, nor the import paths.

To Reproduce

Checkout the latest tag for google vertex and call hatch version

Describe your environment (please complete the following information):

• OS: [e.g. iOS]
• Haystack version:
• Integration version:

Describe the bug
According to Vertex AI documentation, the model for image generation is imagegeneration. Default model right now is imagetext which is a model for image captioning

To Reproduce

from elasticsearch_haystack.bm25_retriever import ElasticsearchBM25Retriever
from elasticsearch_haystack.document_store import ElasticsearchDocumentStore

document_store = ElasticsearchDocumentStore(hosts= "http://localhost:9200/")

documents = [Document(content="There are over 7,000 languages spoken around the world today."),
  Document(content="Elephants have been observed to behave in a way that indicates a high level of self-awareness, such as recognizing themselves in mirrors."),
  Document(content="In certain parts of the world, like the Maldives, Puerto Rico, and San Diego, you can witness the phenomenon of bioluminescent waves.")]

retriever = ElasticsearchBM25Retriever(document_store=document_store)
print(retriever.run(query="How much self awareness do elephants have?"))
# {'documents': []}

This should return the 2nd Document but this does not happen because of this AND operator:

See for comparison the same query in Haystack 1.x:
https://github.com/deepset-ai/haystack/blob/c812250453ab7da35f526a5f2a53e18c058fe2ff/haystack/document_stores/search_engine.py#L1100

Is your feature request related to a problem? Please describe.
Weaviate supports hybrid search (vector + sparse). However, Haystack doesn't support it currently.

Describe the solution you'd like
To be able to control alpha parameter and other necessary configurations for hybrid search through WeaviteDocumentStore

Additional context
Weaviate docs on hybrid
Related issue on Weaviate repo

LlamaHub has added a function to convert outputs to haystack format, which luckily, seems compatible with 2.0 although it was built for 1.x.

The 'verified' data loaders have a standardized way to be loaded and used. Let's create an integration that consists of a custom component where the run function takes only a few things, such as: the name of the loader for download_loader('name of loader'), and the input expected for load_data()

Summary and motivation

See deepset-ai/haystack#6514

Detailed design

See deepset-ai/haystack#6514

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• The feature was announced through social media

Ollama Generator should support streaming, as done in other Generators with streaming_callback: see for example OpenAIGenerator.

google-vertex-haystack package has been released.
We need to add it in the integrations list.

Is your feature request related to a problem? Please describe.
As a user, I'd like to have predictable names for components and other retriever names contain the retrieval method

Describe the solution you'd like
Renaming the QdrantRetriever as QdrantEmbeddingRetriever to align with the current convention. Also, since we foresee QdrantSparseRetriever and QdrantHybridRetriever, it makes sense to create a new retrievers folder and move QdrantEmbeddingRetriever into it.

Describe alternatives you've considered
N/A

Additional context
N/A

• Retrievers of InMemoryDocumentStore (InMemoryBM25Retriever...) have the attribute document_store

• Retrievers of Elasticsearch and Opensearch Document Store have the attribute _document_store

We should probably agree on a consistent approach,
so that for example pipeline.get_component("retriever").document_store always works.

@silvanocerza

Describe the bug
Example colab of Chroma is not working. Link to the notebook: https://colab.research.google.com/drive/1YpDetI8BRbObPDEVdfqUcwhEX9UUXP-m?usp=sharing. The code in example.py should be updated accordingly

To Reproduce
Most up to date code block 👇 It seems like DocumentWriter is not writing any documents to the document store

import os
from pathlib import Path

from haystack import Pipeline
from haystack.components.converters import TextFileToDocument
from haystack.components.writers import DocumentWriter

from chroma_haystack import ChromaDocumentStore
from chroma_haystack.retriever import ChromaQueryRetriever

file_paths = ["data" / Path(name) for name in os.listdir("data")]

document_store = ChromaDocumentStore()

indexing = Pipeline()
indexing.add_component("converter", TextFileToDocument())
indexing.add_component("writer", DocumentWriter(document_store))
indexing.connect("converter", "writer")
indexing.run({"converter": {"sources": file_paths}})

querying = Pipeline()
querying.add_component("retriever", ChromaQueryRetriever(document_store))
results = querying.run({"retriever": {"queries": ["Variable declarations"], "top_k": 3}})

for d in results["retriever"][0]:
    print(d.metadata, d.score)

Describe your environment (please complete the following information):

• OS: Colab
• Haystack version: haystack-ai==2.0.0b3
• Integration version: 0.8.1

Currently, Embedding Retrieval in the Elasticsearch Document Store does not allow scaling scores in the range [0, 1].

I have not implemented this feature for two reasons:

• I have the impression that in the latest versions of Elasticsearch, it comes out of the box
• It's not trivial to do it right

We should investigate these points further.

Is your feature request related to a problem? Please describe.
Huggin Face's Optimum library provides faster inference through ONNX and TensorRT. This can be used to create blazing fast embedding components. The concepts used in Optimum also play well with some of the concepts that we have in Haystack. For example:

Loading non-ONNX checkpoints requires a conversion step, this takes some time. We can do that step in our warmup function (https://huggingface.co/docs/optimum/onnxruntime/usage_guides/models#loading-a-vanilla-transformers-model).

Describe the solution you'd like

• fast embedders based on Optimum
• apply other tricks such as sorting by sequence length and dynamic padding to bring down inference time (see https://github.com/UKPLab/sentence-transformers/blob/c006921e9e9977bc107b05676266b581091688a2/sentence_transformers/SentenceTransformer.py#L179)

Describe alternatives you've considered

• Not having a fast embedder component
• https://github.com/qdrant/fastembed > supported models are limited

Additional context
https://colab.research.google.com/drive/10UAtpz26Gv2LtamT8j33LmI5UFQFwF4T?usp=sharing
https://github.com/huggingface/optimum-benchmark/tree/main/examples/fast-mteb

Currently, Haystack supports storing data into ElasticSearch, InMemory, and RDBMS.
It would be nice to add support of Object storage like S3, which is very cheap and have less hassle to maintain.

In the first step, AWS s3 can be supported as they recently added s3 select option which can help retrieve only a subset of data from an object (currently support CSV file object in compressed or uncompressed format).

Ideally, we can add a Metadata service as well which may help to use Haystack along with Data Lakes.

Describe the bug
When intializing the ChromaDocumentStore, if you rerun the cell without changing the collection name, you will get an error saying telling you collection already exists.

Expected behaviour: Instead of the init calling 'create' by default, if collection already exists, it should simply connect to the existing collection. The current state means we cannot reuse a collection lateron.

To Reproduce
Any colab that has a cell with document_store = ChromaDocumentSrore()
Here is one for example: https://colab.research.google.com/drive/19NzliNb5ZUo1fUbplwnUrZpNBxs9HvoR?usp=sharing

Describe your environment (please complete the following information):

• Integration version: chroma_haystack-0.9.0

Summary and motivation

Google Gemini isn't only provided via Vertex AI. Let's create an integration for the REST API offering, too. There you only need a simple API Key - no need for a GCP account.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• The feature was announced through social media

Is your feature request related to a problem? Please describe.
Sagemaker is widely used for LLM inference in production use cases. We should support Sagemaker in 2.0.

Describe the solution you'd like
Generator similar to the Sagemaker support in 1.x

It should be possible to pass ChatMessages to LLMs served via Ollama, as done in OpenAIChatGenerator.

We should investigate whether this is feasible and if so create a ChatGenerator.

Summary and motivation

Briefly explain the request: why do we need this integration? What use cases does it support?

Detailed design

Explain the design in enough detail for somebody familiar with Haystack to understand, and for somebody familiar with
the implementation to implement. Get into specifics and corner-cases, and include examples of how the feature is used.
Also, if there's any new terminology involved, define it here.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• #161
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• deepset-ai/haystack-integrations#107
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• deepset-ai/devrel-board#241

We need to create an API reference for haystack-core-integrations that would be loaded to our Haystack docs to Readme editor: https://docs.haystack.deepset.ai/v2.0/reference

The instructor embedder (and probably others) are a causing errors right now because the imports still have the '.preview' in them. There might also be some other out of date imports.
(Reported on discord)

why? why would you do such a thing? 😭

Summary and motivation

Add support for CohereGenerator and CohereChatGenerator.

Detailed design

Cohere is one of the leading LLM providers and we should have Haystack Cohere LLM integration. See https://docs.cohere.com/reference/generate and https://docs.cohere.com/reference/chat for more details.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/ (Chat generator missing)
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• deepset-ai/haystack-integrations#150
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• #265

Is your feature request related to a problem? Please describe.
Anthropic models are widely used. We should support them in 2.0.

Describe the solution you'd like
Generator similar to the Anthropic support in 1.x

Is your feature request related to a problem? Please describe.
OpenAI's embedding models hosted on Azure are widely used. We should add the relevant embedders to support them.

See also: deepset-ai/haystack#6620

Describe the solution you'd like
Similar to 1.x support the OpenAI embedding models hosted on Azure.

Describe alternatives you've considered
None

Drafts are ready

Assign labels to PR with the name of the integration being changed.

Summary and motivation

PGVector is a popular request from our community. We should have it in 2.0.

Detailed design

TBD

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations (deepset-ai/haystack-integrations#162)
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• The feature was announced through social media

Is your feature request related to a problem? Please describe.
Recently, Pinecone announced support for Sparse-dense embeddings, allowing for hybrid vector search (both semantic and keyword search). However, haystack doesn't currently support these.

Describe the solution you'd like
Native support for sparse-dense vectors!

Describe alternatives you've considered
I've created a small package called haystack-hybrid-embedding which can be installed to support hybrid vectors for now, but it's a hack!

Currently, we assume that all document writing errors in the Document Store are related to duplicate documents.

We should handle this part better so as not to show misleading exceptions.

Summary and motivation

The integration has already been made and published.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• Docs are published at https://docs.haystack.deepset.ai/
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• An integration tile has been added to https://github.com/deepset-ai/haystack-integrations
• The integration has been listed in the Inventory section of this repo README
• There is an example available to demonstrate the feature
• The feature was announced through social media

Is your feature request related to a problem? Please describe.
Documentation for UnstructuredFileConverter is missing. Here's the integration page for it: https://haystack.deepset.ai/integrations/unstructured-file-converter

Describe the solution you'd like
Adding a new documentation page for this converter. We' can then use this page in its integration page as well.

Is your feature request related to a problem? Please describe.
As a user, I'm familiar with the convention at haystack repository and having the same convention here would make it easy for me to navigate.

Describe the solution you'd like
Create a /generators folder under cohere_haystack integration folder.
The same can apply to jina_haystack and unstructured. This way, we can be more open to new components coming from the same providers

Describe alternatives you've considered
We can leave it as it is.

Additional context
N/A

Is your feature request related to a problem? Please describe.
Cohere has a new type of embedding models: Embed v3. Let's add support for it for 2.0 pipelines
https://txt.cohere.com/introducing-embed-v3/
https://docs.cohere.com/reference/embed

Describe the solution you'd like
It seems like the only change we need to make in CohereEmbedder is to add the input_type parameter. See the documentation: https://docs.cohere.com/reference/embed

Describe alternatives you've considered
There's no other way of using these v3 models. When we pass the v3 model names, the CohereEmbedder (or cohere API, not quite sure which one) may not raise any error but the model fallbacks to a v2 model

Additional context
N/A

Summary and motivation

This already mostly implemented but I'm creating this issue to track the remaining work.

Checklist

If the request is accepted, ensure the following checklist is complete before closing this issue.

• The code was merged in the main branch
• #155
• There is a Github workflow running the tests for the integration nightly and at every PR
• A label named like integration:<your integration name> has been added to this repo
• The labeler.yml file has been updated
• The package has been released on PyPI
• #120
• The integration has been listed in the Inventory section of this repo README
• #117
• #116