Add Azure OpenAI client

[oxaroky02]

Updated to reflect various changes based on feedback; added some tests.

TLDR;

• This is big PR since adding Azure Open AI support also meant making all the rest of Archyve support any LLM client, which in turn meant removing assumptions about Ollama client-specifics.
• The Azure OpenAI client support chat, embedding and completion requests.

The detailed description is organized as follows.

• Model changes
  • ModelServer
  • ModelConfig
  • ApiCall
• LLM Clients
  • Common
  • OpenAI Client
  • Ollama Client
  • Base Client
• Services
  • Helpers
  • Common Client Support
  • API

Model changes

ModelServer

• Tweaked the provider enumeration, renaming openai to openai_azure and setting it up with a naming prefix so that we can query a model server with server.provider_ollama? and server.provider_openai_azure?.
• Added (via migration) optional api_key field for use with external servers that require API tokens.
• Added validation support methods api_key_required? and api_version_required? implemented based on a server's provider, both of which return true if the provider is openai_azure.
• Added validation to enforce presence of api_key.

ModelConfig

• Added two fields (via migration), an optional api_version and an optional model_server reference.
• Added validation support method api_version_required? which calls the same method for a model server if present
• Added validation to enforce presence of api_version

ApiCall

• Added from_faraday support method to create an API call record based on Faraday-middleware based request/responses.
• Used by the Open AI client via a Faraday request interceptor.

Caveat: When using Open AI "streaming", we can't retrieve the response body via the interceptor.

LLM Clients

Common

• Tweaked to #chat (in all clients) to receive Message and handle the chat history conversion within the client

OpenAI Client

• Added the ruby-openai gem to the project's bundler
• Re-implemented the openai/ client API as follows under the Openai:: module as follows:
  • Client base class for non-Azure-specific use of the ruby-openai gem, including integration with Faraday's interceptor middleware flow to track request/response details
  • Update: removed the mixins and put the chat_request and complete_request and embedding_request methods into the Client where they belong for clarity
  • ChatMessageHelper to help process Message for chat history formatting for server-specific requests
  • AzureClient implementation sub-class for all the Azure-specific aspects of the client.
  • Update: Renamed some internal method names to use "connection" when referring to the underling ruby-openai "client" to avoid ambiguity with the use of "client" in Archyve.

Ollama Client

• Modified #chat to use the OpenAI ChatMessageHelper for the chat history, since Ollama API is compatible with OpenAI.

Base Client

• Added optional api_version property
• Tweaked client_class_for factory method to return appropriate client class based on the provider

Services

Helpers

• Started a Helpers module
• Added a ModelClientHelper helper that is shared by various services, consolidating common functionality when instantiating an LLM client based on a ModelConfig.
  • Update: Converted from mixin to proper helper instantiated by services.
• Removed any Ollama client-specific references

Common Client Support

• Refactored the following service classes to leverage the shared ModelClientHelper (update: replaced mixin with helper instances) for the common functionality to be able to retrieve an LLM client based on the model configuration:
  • ResponseStreamer,
  • Embedder and
  • SummarizeMessage
• Removed any Ollama client-specific references, including letting the clients deal with the Message processing for chat history.
• Tweaked the TheIngestor, ReplyToMessage and Search services to reflect the changes in the refactored service classes

API

• Tweaked the ModelLoader to support the model server's API key

Testing

• Refactored the ollama/chat_spec to instead test the Open AI ChatMessageHelper shared by both clients
• Added an Ollama and Azure Open client test specs to test the three method #embed, #chat, and #complete
• Put in a spec_helper.rb config check to exclude tests tagged with :az_openai unless both AZURE_OPENAI_API_KEY and AZURE_OPENAI_URI env variables have been set.

[oxaronick]

I'm going to stop there for now. In general, I think it's great! Just some concerns about how the code is structured.

There are no tests, but I can contribute those afterwards. Archyve isn't near 100% test coverage yet, but there are 100-odd tests in there now, and I'm trying to test new code.

[oxaroky02]

There are no tests, but I can contribute those afterwards. Archyve isn't near 100% test coverage yet, but there are 100-odd tests in there now, and I'm trying to test new code.

Yep. I did want to include tests but I ran into some trouble running the tests. I promise to start adding testing in separate PRs.

[oxaroky02]

There are no tests, but I can contribute those afterwards. Archyve isn't near 100% test coverage yet, but there are 100-odd tests in there now, and I'm trying to test new code.

Yep. I did want to include tests but I ran into some trouble running the tests. I promise to start adding testing in separate PRs.

OK, figured out my problem with running tests; can run tests now, and found that the old ollama chat test failed due to my refactoring. So I've fixed that so at least all the current tests pass again. Sorry about that.