AzureML Retrieval Augmented Generation Utilities

This package is in alpha stage at the moment, use at risk of breaking changes and unstable behavior.

It contains utilities for:

• Processing text documents into chunks appropriate for use in LLM prompts, with metadata such is source url.
• Embedding chunks with OpenAI or HuggingFace embeddings models, including the ability to update a set of embeddings over time.
• Create MLIndex artifacts from embeddings, a yaml file capturing metadata needed to deserialize different kinds of Vector Indexes for use in langchain. Supported Index types:
  • FAISS index (via langchain)
  • Azure Cognitive Search index
  • Pinecone index
  • Milvus index
  • Azure Cosmos Mongo vCore index

Getting started

You can install AzureMLs RAG package using pip.

pip install azureml-rag

There are various extra installs you probably want to include based on intended use:

• faiss: When using FAISS based Vector Indexes
• cognitive_search: When using Azure Cognitive Search Indexes
• pinecone: When using Pinecone Indexes
• azure_cosmos_mongo_vcore: When using Azure Cosmos Mongo vCore Indexes
• hugging_face: When using Sentence Transformer embedding models from HuggingFace (local inference)
• document_parsing: When cracking and chunking documents locally to put in an Index

MLIndex

MLIndex files describe an index of data + embeddings and the embeddings model used in yaml.

Azure Cognitive Search Index:

embeddings:
  dimension: 768
  kind: hugging_face
  model: sentence-transformers/all-mpnet-base-v2
  schema_version: '2'
index:
  api_version: 2021-04-30-Preview
  connection:
    id: /subscriptions/<subscription_id>/resourceGroups/<resource_group>/providers/Microsoft.MachineLearningServices/workspaces/<workspace>/connections/<acs_connection_name>
  connection_type: workspace_connection
  endpoint: https://<acs_name>.search.windows.net
  engine: azure-sdk
  field_mapping:
    content: content
    filename: filepath
    metadata: meta_json_string
    title: title
    url: url
    embedding: contentVector
  index: azureml-rag-test-206e03b6-3880-407b-9bc4-c0a1162d6c70
  kind: acs

Pinecone Index:

embeddings:
  dimension: 768
  kind: hugging_face
  model: sentence-transformers/all-mpnet-base-v2
  schema_version: '2'
index:
  connection:
    id: /subscriptions/<subscription_id>/resourceGroups/<resource_group>/providers/Microsoft.MachineLearningServices/workspaces/<workspace>/connections/<pinecone_connection_name>
  connection_type: workspace_connection
  engine: pinecone-sdk
  field_mapping:
    content: content
    filename: filepath
    metadata: metadata_json_string
    title: title
    url: url
  index: azureml-rag-test-206e03b6-3880-407b-9bc4-c0a1162d6c70
  kind: pinecone

Azure Cosmos Mongo vCore Index:

embeddings:
  dimension: 768
  kind: hugging_face
  model: sentence-transformers/all-mpnet-base-v2
  schema_version: '2'
index:
  connection:
    id: /subscriptions/<subscription_id>/resourceGroups/<resource_group>/providers/Microsoft.MachineLearningServices/workspaces/<workspace>/connections/<cosmos_connection_name>
  connection_type: workspace_connection
  engine: pymongo-sdk
  field_mapping:
    content: content
    filename: filepath
    metadata: metadata_json_string
    title: title
    url: url
    embedding: contentVector
  database: azureml-rag-test-db
  collection: azureml-rag-test-collection
  index: azureml-rag-test-206e03b6-3880-407b-9bc4-c0a1162d6c70
  kind: azure_cosmos_mongo_vcore

Create MLIndex

Examples using MLIndex remotely with AzureML and locally with langchain live here: https://github.com/Azure/azureml-examples/tree/main/sdk/python/generative-ai/rag

Consume MLIndex

from azureml.rag.mlindex import MLIndex

retriever = MLIndex(uri_to_folder_with_mlindex).as_langchain_retriever()
retriever.get_relevant_documents('What is an AzureML Compute Instance?')

Changelog

Please insert change log into "Next Release" ONLY.

Next release

0.2.31.1

• Update search index with azure-search-documents 11.4.0
• add azureml-core in the dependency list

0.2.31

• Categorize user error and system error, and update RH accordingly to show in logs
• Bugfix using obo credential for AAD connections.
• Prevention fix to support AadCredentialConfig in Connection object
• Update Pinecone legacy API
• Creating image embedding index with azure-search-documents 11.4.0

0.2.30.2

• Bugfix remove azure_ad_token_provider from EmbeddingContainer metadata
• set embeddings_model as optional argument

0.2.30.1

• Introduce elasticsearch extra to declare transitive dependency on the elasticsearch package when using Elasticsearch indices.

0.2.30

• Bugfix in models.py to handle empty deployment name.
• Supporting existing elasticsearch indices
• Bug fix in crack_and_chunk_and_embed_and_index
• Fixing bug in using AAD auth type ACS connections.

0.2.29.2

• Fixing ACS index creation failure with azure-search-documents 11.4.0

0.2.29.1

• Fixing FAISS, dependable_faiss_import import failure with Langchain 0.1.x

0.2.29

• Support AAD and MSI auth type in AOAI, ACS connection

0.2.28

• Ensure compatibility with newer versions of azure-ai-ml.
• Upgrade langchain to support up to 0.1

0.2.27

• Support Cohere serverless endpoint
• Support multiple ACS lookups in the same process, eliminating field mapping conflicts
• Support pass-in credential in get_connection_by_name_v2 to unblock managed vNet setup
• Update validate_deployments in crack_chunk_embed_index_and_register.py

0.2.26

• Support for .csv and .json file extensions in pipeline
• Ignore mlflow.exceptions.RestException in safe_mlflow_log_metric
• validate_deployments supports openai v1.0+
• Removing unexpected keyword argument 'engine'
• Checking ACS account has enough index quota
• infer_deployment supports openai v1.0+
• Create missing fields for existing index

0.2.25

• Using local cached encodings.
• Adding convert_to_dict() for openai v1.0+
• Check index_config before passing in validate_deployments.py
• Limit size of documents upload to ACS in one batch to solve RequestEntityTooLargeError

0.2.24.2

• Supporting *.cognitiveservices.* endpoint
• Adding azureml-rag specific user_agent when using DocumentIntelligence
• Refactored update index tasks
• Supporting uppercase file extensions name in crack_and_chunk
• Fixing Deployment importing bug in utils
• Adding the playgroundType tag in MLIndex Asset used for Azure AI studio playground
• Remove mandatory module-level imports of optional extra packages

0.2.24.1

• Fixing is_florence key detection
• Using 'embedding_connection_id' instead of 'florence_connection_id' as parameter name

0.2.24

• Introducing image ingestion with florence embedding API
• Adding dummy output to validate_deployments for holding the right order
• Fixing DeploymentNotFound bug

0.2.23.5

• Deprecate pkg_resources in logging.py (https://setuptools.pypa.io/en/latest/pkg_resources.html)

0.2.23.4

• Make the api_type parameter non-case sensitive in OpenAIEmbedder
• Bug fix in embeddings container path

0.2.23.3

• Set upper bound for langchain to 0.0.348

0.2.23.2

• Make tiktoken pull from a cache instead of making the outgoing network call to get encodings files
• Add support for Azure Cosmos Mongo vCore

0.2.23.1

• Fixing exception handling in validate_deployments to support OpenAI v1.0+

0.2.23

• Support OpenAI v1.0 +
• Handle FAISS.load_local() change since Langchain 0.0.318
• Handle mailto links in url crawling component.
• Add support for Milvus vector store

0.2.22

• update pypdf's version to 3.17.1 in document-parsing.

0.2.21

• Use workspace connection tags instead of metadata since it's deprecated.
• Fix bug handling single files in files_to_document_sources

0.2.20

• Initial introduction of validate_deployments.
• Asset registration in *_and_register attempts to infer target workspace from asset_uri and handle multiple auth options
• activity_logger moved out as first arg, this is an intermediate step as logger also shouldn't be first arg and instead handled by get_logger, activity_logger should be truly optional.
• validate_deployments itself was modified to make its interface closer to what existing tasks expect as input, and callable from other tasks as a function.

0.2.19

• Introduce a new path parameter in the index section of MLIndex documents over FAISS indices, to allow the path to FAISS index files to be different from the MLIndex document path.
• Ensure MLIndex.base_uri is never undefined for a valid MLIndex object.

0.2.18.1

• Only save out metadata before embedding in crack_and_chunk_and_embed_and_index
• Update create_embeddings to return num_embedded value.
  • This enables crack_and_chunk_and_embed to skip loading EmbeddedDocument partitions when no documents were embedded (all reused).

0.2.18

• Add new task to crack, chunk, embed, index to ACS, and register MLIndex in one step.
• Handle openai.api_type being None

0.2.17

• Fix loading MLIndex failure. Don't need to get the endpoint from connection when it is already provided.
• Try use langchain VectorStore and fallback to vendor
• Support `azure-search-documents==11.4.0b11``
• Add support for Pinecone in DataIndex

0.2.16

• Use Retry-After when aoai embedding endpoint throws RateLimitError

0.2.15.1

• Fix vendored FAISS langchain VectorStore to only error when a doc is None (rather than when a Document isn't exactly the right class)

0.2.15

• Support PDF cracking with Azure Document Intelligence service
• crack_and_chunk_and_embed now pulls documents through to embedding (streaming) and embeds documents in parallel batches
• Update default field names.
• Fix long file name bug when writing to output during crack and chunk

0.2.14

• Fix git_clone to handle WorkspaceConnections, again.

0.2.13

• Fix git_clone to handle WorkspaceConnection objects and urls with usernames already in them.

0.2.12

• Only process .jsonl and .csv files when reading chunks for embedding.

0.2.11

• Check casing for model kind and api_type
• Ensure api_version not being set is supported and default make sense.
• Add support for Pinecone indexes

0.2.10

• Fix QA generator and connections check for ApiType metadata

0.2.9

• QA data generation accepts connection as input

0.2.8

• Remove allowed_special="all" from tiktoken usage as it encodes special tokens like <|endoftext|> as their special token rather then as plain text (which is the case when only disallowed_special=() is set on its own)
• Stop truncating texts to embed (to model ctx length) as new azureml.rag.embeddings.OpenAIEmbedder handles batching and splitting long texts pre-embed then averaging the results into a single final embedding.
• Loosen tiktoken version range from ~=0.3.0 to <1

0.2.7

• Don't try and use MLClient for connections if azure-ai-ml<1.10.0
• Handle Custom Conenctions which azure-ai-ml can't deserialize today.
• Allow passing faiss index engine to MLIndex local
• Pass chunks directly into write_chunks_to_jsonl

0.2.6

• Fix jsonl output mode of crack_and_chunk writing csv internally.

0.2.5

• Ensure EmbeddingsContainer.mount_and_load sets create_destination=True when mounting to create embeddings_cache location if it's not already created.
• Fix safe_mlflow_start_run to yield None when mlflow not available
• Handle custom field_mappings passed to update_acs task.

0.2.4

• Introduce crack_and_chunk_and_embed task which tracks deletions and reused source + documents to enable full sync with indexes, levering EmbeddingsContainer for storage of this information across Snapshots.
• Restore workspace_connection_to_credential function.

0.2.3

• Fix git clone url format bug

0.2.2

• Fix all langchain splitter to use tiktoken in an airgap friendly way.

0.2.1

• Introduce DataIndex interface for scheduling Vector Index Pipeline in AzureML and creating MLIndex Assets
• Vendor various langchain components to avoid breaking changes to MLIndex internal logic

0.1.24.2

• Fix all langchain splitter to use tiktoken in an airgap friendly way.

0.1.24.1

• Fix subsplitter init bug in MarkdownHeaderSplitter
• Support getting langchain retriever for ACS based MLIndex with embeddings.kind: none.

0.1.24

• Don't mlflow log unless there's an active mlflow run.
• Support langchain.vectorstores.azuresearch after langchain>=0.0.273 upgraded to azure-search-documents==11.4.0b8
• Use tiktoken encodings from package for other splitter types

0.1.23.2

• Handle Path objects passed into MLIndex init.

0.1.23.1

• Handle .api.cognitive style aoai endpoints correctly

0.1.23

• Ensure tiktoken encodings are packaged in wheel

0.1.22

• Set environment variables to pull encodings files from directory with cache key to avoid tiktoken external network call
• Fix mlflow log error when there's no files input

0.1.21

• Fix top level imports in update_acs task failing without helpful reason when old azure-search-documents is installed.

0.1.20

• Fix Crack'n'Chunk race-condition where same named files would overwrite each other.

0.1.19

• Various bug fixes:
  • Handle some malformed git urls in git_clone task
  • Try fall back when parsing csv with pandas fails
  • Allow chunking special tokens
  • Ensure logging with mlflow can't fail a task
• Update to support latest azure-search-documents==11.4.0b8

0.1.18

• Add FaissAndDocStore and FileBasedDocStore which closely mirror langchains' FAISS and InMemoryDocStore without the langchain or pickle dependency. These are default not used until PromptFlow support has been added.
• Pin azure-documents-search==11.4.0b6 as there's breaking changes in 11.4.0b7 and 11.4.0b8

0.1.17

• Update interactions with Azure Cognitive Search to use latest azure-documents-search SDK

0.1.16

• Convert api_type from Workspace Connections to lower case to appease langchains case sensitive checking.

0.1.15

• Add support for custom loaders
• Added logging for MLIndex.init to understand usage of MLindex

0.1.14

• Add Support for CustomKeys connections
• Add OpenAI support for QA Gen and Embeddings

0.1.13 (2023-07-12)

• Implement single node non-PRS embed task to enable clearer logs for users.

0.1.12 (2023-06-29)

• Fix casing check of ApiVersion, ApiType in infer_deployment util

0.1.11 (2023-06-28)

• Update casing check for workspace connection ApiVersion, ApiType
• int casting for temperature, max_tokens

0.1.10 (2023-06-26)

• Update data asset registering to have adjustable output_type
• Remove asset registering from generate_qa.py

0.1.9 (2023-06-22)

• Add azureml.rag.data_generation module.
• Fixed bug that would cause crack_and_chunk to fail for documents that contain non-utf-8 characters. Currently these characters will be ignored.
• Improved heading extraction from Markdown files. When use_rcts=False Markdown files will be split on headings and each chunk with have the heading context up to the root as a prefix (e.g. # Heading 1\n## Heading 2\n# Heading 3\n{content})

0.1.8 (2023-06-21)

• Add deployment inferring util for use in azureml-insider notebooks.

0.1.7 (2023-06-08)

• Improved telemetry for tasks (used in RAG Pipeline Components)

0.1.6 (2023-05-31)

• Fail crack_and_chunk task when no files were processed (usually because of a malformed input_glob)
• Change update_acs.py to default push_embeddings=True instead of False.

0.1.5 (2023-05-19)

• Add api_base back to MLIndex embeddings config for back-compat (until all clients start getting it from Workspace Connection).
• Add telemetry for tasks used in pipeline components, not enabled by default for SDK usage.

0.1.4 (2023-05-17)

• Fix bug where enabling rcts option on split_documents used nltk splitter instead.

0.1.3 (2023-05-12)

• Support Workspace Connection based auth for Git, Azure OpenAI and Azure Cognitive Search usage.

0.1.2 (2023-05-05)

• Refactored document chunking to allow insertion of custom processing logic

0.0.1 (2023-04-25)

Features Added

• Introduced package
• langchain Retriever for Azure Cognitive Search