Added AWS Bedrock embeddings

[jackretterer]

Summary: Added support for AWS Bedrock embeddings. Leverages "amazon.titan-tg1-large" for the embedding model.

Test

• find your aws secret access key and key id; make sure the account has access to bedrock's tian embed model
• follow the instructions in https://github.com/Unstructured-IO/unstructured/blob/d5e797cd44bfe3d423750b4442148d03f942d27d/docs/source/bricks/embedding.rst#bedrockembeddingencoder

[badGarnet]

Suggested change

	embeddings = [np.array(self.bedrock_client.embed_query(str(e))) for e in elements]
	embeddings = [np.array(self.embed_query(str(e))) for e in elements]

[badGarnet]

actually langchain has a bulk embed api: https://api.python.langchain.com/en/latest/embeddings/langchain.embeddings.bedrock.BedrockEmbeddings.html#langchain.embeddings.bedrock.BedrockEmbeddings.embed_documents
so better to do:

embeddings = self.bedrock_client.embed_documents([str(e) for e in elements])

[badGarnet]

I see this is using

unstructured/unstructured/embed/openai.py

Lines 37 to 52 in 97f8326

	def _add_embeddings_to_elements(self, elements, embeddings) -> List[Element]:
	assert len(elements) == len(embeddings)
	elements_w_embedding = []
	for i, element in enumerate(elements):
	original_method = element.to_dict
	def new_to_dict(self):
	d = original_method()
	d["embeddings"] = self.embeddings
	return d
	element.embeddings = embeddings[i]
	elements_w_embedding.append(element)
	element.to_dict = types.MethodType(new_to_dict, element)
	return elements

this part of the code essentially depend on element data definition so can be shared by different embeddings. So we can consider refactor this either:

• as a method for elements
• as a utils func in unstructured/embed/utils.py so both here and openai version can import the same function

[ryannikolaidis]

also just FYI: that logic is buggy. fix on route over here

[badGarnet]

yeah; given this operation here I mean leaning even more toward moving this function to be part of element (here self refers to an element... NOT this class itself

[badGarnet]

I see this is using

unstructured/unstructured/embed/openai.py

Lines 37 to 52 in 97f8326

	def _add_embeddings_to_elements(self, elements, embeddings) -> List[Element]:
	assert len(elements) == len(embeddings)
	elements_w_embedding = []
	for i, element in enumerate(elements):
	original_method = element.to_dict
	def new_to_dict(self):
	d = original_method()
	d["embeddings"] = self.embeddings
	return d
	element.embeddings = embeddings[i]
	elements_w_embedding.append(element)
	element.to_dict = types.MethodType(new_to_dict, element)
	return elements

this part of the code essentially depend on element data definition so can be shared by different embeddings. So we can consider refactor this either:

• as a method for elements
• as a utils func in unstructured/embed/utils.py so both here and openai version can import the same function

[ahmetmeleq]

Reminder to update the docs: https://github.com/Unstructured-IO/unstructured/blob/main/docs/source/bricks/embedding.rst

[ahmetmeleq]

1. Let's form an extra named embed-aws-bedrock and add it as an extra key to setup.py as in:
   

   unstructured/setup.py

   Line 160 in 282b8f7

   "openai": load_requirements("requirements/ingest-openai.in"),

2. We need an .in requirements file for this to work (we can name it embed-aws-bedrock.in). Example:
   https://github.com/Unstructured-IO/unstructured/blob/282b8f700d9471b9430e8a37af73ca96b980e0f0/requirements/ingest-openai.in

3. Finally, let's add a @requires_dependencies decorator to the methods where we use the extra libraries explicitly. Example:
   

   unstructured/unstructured/embed/openai.py

   Lines 44 to 49 in 282b8f7

   	@EmbeddingEncoderConnectionError.wrap
   	@requires_dependencies(
   	["langchain", "openai", "tiktoken"],
   	extras="openai",
   	)
   	def get_openai_client(self):

[ahmetmeleq]

Let's wrap potential connection errors as in:

unstructured/unstructured/embed/openai.py

Line 44 in 282b8f7

@EmbeddingEncoderConnectionError.wrap

[ahmetmeleq]

Can we add a mock test as in: https://github.com/Unstructured-IO/unstructured/blob/main/test_unstructured/embed/test_openai.py

[ahmetmeleq]

LGTM!