Support AzureOpenAI LLMs & fix research agent defaulting to OpenAI (#94)

* Try get rid of OpenAI default embeddings in various LlamaIndex objects

* Support AzureOpenAI LLMs

* Restore research agent persisted storage

motleycrew/applications/research_agent/question_prioritizer.py

@@ -1,7 +1,10 @@
+from typing import Optional
+
 from langchain.prompts import PromptTemplate
 from langchain_core.prompts.base import BasePromptTemplate
 from langchain_core.runnables import RunnableLambda, RunnablePassthrough, chain
 from langchain_core.tools import StructuredTool
+from langchain_core.language_models import BaseLanguageModel
 from pydantic import BaseModel, Field
 
 from motleycrew.applications.research_agent.question import Question
@@ -16,8 +19,9 @@ class QuestionPrioritizerTool(MotleyTool):
     def __init__(
         self,
         prompt: str | BasePromptTemplate = None,
+        llm: Optional[BaseLanguageModel] = None,
     ):
-        langchain_tool = create_question_prioritizer_langchain_tool(prompt=prompt)
+        langchain_tool = create_question_prioritizer_langchain_tool(prompt=prompt, llm=llm)
 
         super().__init__(langchain_tool)
 
@@ -47,6 +51,7 @@ class QuestionPrioritizerInput(BaseModel, arbitrary_types_allowed=True):
 
 def create_question_prioritizer_langchain_tool(
     prompt: str | BasePromptTemplate = None,
+    llm: Optional[BaseLanguageModel] = None,
 ) -> StructuredTool:
     if prompt is None:
         prompt = _default_prompt
@@ -56,6 +61,7 @@ def create_question_prioritizer_langchain_tool(
         name="Question prioritizer",
         description="Takes the original question and a list of derived questions, "
         "and selects from the latter the one most pertinent to the former",
+        llm=llm,
     )
 
     @chain

motleycrew/applications/research_agent/question_task.py

@@ -41,7 +41,7 @@ def __init__(
         self.n_iter = 0
         self.question = Question(question=question)
         self.graph_store.insert_node(self.question)
-        self.question_prioritization_tool = QuestionPrioritizerTool()
+        self.question_prioritization_tool = QuestionPrioritizerTool(llm=llm)
         self.question_generation_tool = QuestionGeneratorTool(
             query_tool=query_tool, graph=self.graph_store, llm=llm
         )

motleycrew/common/enums.py

@@ -8,8 +8,9 @@ class LLMProvider:
     TOGETHER = "together"
     GROQ = "groq"
     OLLAMA = "ollama"
+    AZURE_OPENAI = "azure_openai"
 
-    ALL = {OPENAI, ANTHROPIC, REPLICATE, TOGETHER, GROQ, OLLAMA}
+    ALL = {OPENAI, ANTHROPIC, REPLICATE, TOGETHER, GROQ, OLLAMA, AZURE_OPENAI}
 
 
 class LLMFramework:

motleycrew/common/llms.py

@@ -1,7 +1,6 @@
 """Helper functions to initialize Language Models (LLMs) from different frameworks."""
 
-from motleycrew.common import Defaults
-from motleycrew.common import LLMProvider, LLMFramework
+from motleycrew.common import Defaults, LLMFramework, LLMProvider
 from motleycrew.common.exceptions import LLMProviderNotSupported
 from motleycrew.common.utils import ensure_module_is_installed
 
@@ -209,6 +208,47 @@ def llama_index_ollama_llm(
     return Ollama(model=llm_name, temperature=llm_temperature, **kwargs)
 
 
+def langchain_azure_openai_llm(
+    llm_name: str = Defaults.DEFAULT_LLM_NAME,
+    llm_temperature: float = Defaults.DEFAULT_LLM_TEMPERATURE,
+    **kwargs,
+):
+    """Initialize an Azure OpenAI LLM client for use with Langchain.
+
+    Args:
+        llm_name: Name of the LLM in Azure OpenAI API.
+        llm_temperature: Temperature for the LLM.
+    """
+    from langchain_openai import AzureChatOpenAI
+
+    return AzureChatOpenAI(model=llm_name, temperature=llm_temperature, **kwargs)
+
+
+def llama_index_azure_openai_llm(
+    llm_name: str = Defaults.DEFAULT_LLM_NAME,
+    llm_temperature: float = Defaults.DEFAULT_LLM_TEMPERATURE,
+    **kwargs,
+):
+    """Initialize an Azure OpenAI LLM client for use with LlamaIndex.
+
+    Args:
+        llm_name: Name of the LLM in Azure OpenAI API.
+        llm_temperature: Temperature for the LLM.
+    """
+    ensure_module_is_installed("llama_index")
+    from llama_index.llms.azure_openai import AzureOpenAI
+
+    if "azure_deployment" in kwargs:
+        kwargs["engine"] = kwargs.pop("azure_deployment")
+
+    if "engine" not in kwargs:
+        raise ValueError(
+            "For using Azure OpenAI with LlamaIndex, you must specify an engine/deployment name."
+        )
+
+    return AzureOpenAI(model=llm_name, temperature=llm_temperature, **kwargs)
+
+
 LLM_MAP = {
     (LLMFramework.LANGCHAIN, LLMProvider.OPENAI): langchain_openai_llm,
     (LLMFramework.LLAMA_INDEX, LLMProvider.OPENAI): llama_index_openai_llm,
@@ -222,6 +262,8 @@ def llama_index_ollama_llm(
     (LLMFramework.LLAMA_INDEX, LLMProvider.GROQ): llama_index_groq_llm,
     (LLMFramework.LANGCHAIN, LLMProvider.OLLAMA): langchain_ollama_llm,
     (LLMFramework.LLAMA_INDEX, LLMProvider.OLLAMA): llama_index_ollama_llm,
+    (LLMFramework.LANGCHAIN, LLMProvider.AZURE_OPENAI): langchain_azure_openai_llm,
+    (LLMFramework.LLAMA_INDEX, LLMProvider.AZURE_OPENAI): llama_index_azure_openai_llm,
 }
 
 

motleycrew/tools/simple_retriever_tool.py

@@ -65,14 +65,16 @@ def make_retriever_langchain_tool(
         # load the documents and create the index
         documents = SimpleDirectoryReader(data_dir).load_data()
         index = VectorStoreIndex.from_documents(
-            documents, transformations=[SentenceSplitter(chunk_size=512), embeddings]
+            documents,
+            transformations=[SentenceSplitter(chunk_size=512), embeddings],
+            embed_model=embeddings,
         )
         # store it for later
         index.storage_context.persist(persist_dir=persist_dir)
     else:
         # load the existing index
         storage_context = StorageContext.from_defaults(persist_dir=persist_dir)
-        index = load_index_from_storage(storage_context)
+        index = load_index_from_storage(storage_context, embed_model=embeddings)
 
     retriever = index.as_retriever(
         similarity_top_k=10,