azure-genai-utils 0.0.2.17

Utility functions for Azure GenAI

Azure GenAI Utils

This repository contains a set of utilities for working with Azure GenAI. The utilities are written in Python and are designed to be used for Hackathons, Workshops, and other events where you need to quickly get started with Azure GenAI.

Requirements

• Azure Subscription
• Azure AI Foundry
• Bing Search API Key
• Python 3.8 or later
• .env file: Please do not forget to modify the .env file to match your account. Rename .env.sample to .env or copy and use it

  AZURE_OPENAI_ENDPOINT=xxxxx
  AZURE_OPENAI_API_KEY=xxxxx
  OPENAI_API_VERSION=2024-12-01-preview
  AZURE_OPENAI_DEPLOYMENT_NAME=gpt-4o-mini
  
  # Optinoal, but required for LangChain
  LANGCHAIN_TRACING_V2=false
  LANGCHAIN_ENDPOINT=https://api.smith.langchain.com
  LANGCHAIN_API_KEY=xxxxx
  LANGCHAIN_PROJECT="YOUR-PROJECT"
  

Installation

PyPI

• pip install azure-genai-utils

From Source

• python setup.py install

Usage

Azure OpenAI Test

Expand

from azure_genai_utils.aoai import AOAI
aoai = AOAI()
aoai.test_api_call()

PDF RAG Chain

Expand

from azure_genai_utils.rag.pdf import PDFRetrievalChain

pdf_path = "[YOUR-PDF-PATH]"

pdf = PDFRetrievalChain(
    source_uri=[pdf_path],
    loader_type="PDFPlumber",
    model_name="gpt-4o-mini",
    embedding_name="text-embedding-3-large",
    chunk_size=500,
    chunk_overlap=50,
).create_chain()

question = "[YOUR-QUESTION]"
docs = pdf.retriever.invoke(question)
results = pdf.chain.invoke({"chat_history": "", "question": question, "context": docs})

Bing Search

Please make sure to set the following environment variables in your .env file:

BING_SUBSCRIPTION_KEY=xxxxx

Expand

from azure_genai_utils.tools import BingSearch
from dotenv import load_dotenv

# You need to add BING_SUBSCRIPTION_KEY=xxxx in .env file
load_dotenv()

# Basic usage
bing = BingSearch(max_results=2, locale="ko-KR")
results = bing.invoke("Microsoft AutoGen")
print(results)

## Include news search results and format output
bing = BingSearch(
    max_results=2,
    locale="ko-KR",
    include_news=True,
    include_entity=False,
    format_output=True,
)
results = bing.invoke("Microsoft AutoGen")
print(results)

LangGraph Example (Bing Search + Azure GenAI)

Expand

import json
from typing import Annotated
from typing_extensions import TypedDict
from langchain_openai import AzureChatOpenAI
from langchain_core.messages import ToolMessage
from langgraph.graph.message import add_messages
from langgraph.graph import StateGraph
from langgraph.prebuilt import ToolNode
from langgraph.graph import START, END
from azure_genai_utils.tools import BingSearch
from dotenv import load_dotenv

load_dotenv()

class State(TypedDict):
    messages: Annotated[list, add_messages]

llm = AzureChatOpenAI(model="gpt-4o-mini")
tool = BingSearch(max_results=3, format_output=False)
tools = [tool]
llm_with_tools = llm.bind_tools(tools)

def chatbot(state: State):
    answer = llm_with_tools.invoke(state["messages"])
    return {"messages": [answer]}

def route_tools(
    state: State,
):
    if messages := state.get("messages", []):
        ai_message = messages[-1]
    else:
        raise ValueError(f"No messages found in input state to tool_edge: {state}")

    if hasattr(ai_message, "tool_calls") and len(ai_message.tool_calls) > 0:
        return "tools"

    return END

graph_builder = StateGraph(State)
graph_builder.add_node("chatbot", chatbot)
tool_node = ToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)

graph_builder.add_conditional_edges(
    source="chatbot",
    path=route_tools,
    path_map={"tools": "tools", END: END},
)

graph_builder.add_edge("tools", "chatbot")
graph_builder.add_edge(START, "chatbot")
graph = graph_builder.compile()

# Test
inputs = {"messages": "Microsoft AutoGen"}

for event in graph.stream(inputs, stream_mode="values"):
    for key, value in event.items():
        print(f"\n==============\nSTEP: {key}\n==============\n")
        print(value[-1])

Synthetic Data Generation

Expand

from azure_genai_utils.synthetic import (
    QADataGenerator,
    CustomQADataGenerator,
    QAType,
    generate_qas,
)

input_batch = [
    "The quick brown fox jumps over the lazy dog.",
    "What is the capital of France?",
]

model_config = {
    "deployment": "gpt-4o-mini",
    "model": "gpt-4o-mini",
    "max_tokens": 256,
}

try:
    qa_generator = QADataGenerator(model_config=model_config)
    # qa_generator = CustomQADataGenerator(
    #     model_config=model_config, templates_dir=f"./azure_genai_utils/synthetic/prompt_templates/ko"
    # )
    task = generate_qas(
        input_texts=input_batch,
        qa_generator=qa_generator,
        qa_type=QAType.LONG_ANSWER,
        num_questions=2,
        concurrency=3,
    )
except Exception as e:
    print(f"Error generating QAs: {e}")

Azure Custom Speech

Please make sure to set the following environment variables in your .env file:

AZURE_AI_SPEECH_REGION=xxxxx
AZURE_AI_SPEECH_API_KEY=xxxxx

Expand

from azure_genai_utils.stt.stt_generator import CustomSpeechToTextGenerator

# Initialize the CustomSpeechToTextGenerator
stt = CustomSpeechToTextGenerator(
    custom_speech_lang="Korean",
    synthetic_text_file="cc_support_expressions.jsonl",
    train_output_dir="synthetic_data_train",
    train_output_dir_aug="synthetic_data_train_aug",
    eval_output_dir="synthetic_data_eval",
)

### Training set
# Generate synthetic text
topic = "Call center QnA related expected spoken utterances"
content = stt.generate_synthetic_text(
    topic=topic, num_samples=2, model_name="gpt-4o-mini"
)
stt.save_synthetic_text(output_dir="plain_text")

# Generate synthetic wav files for training
train_tts_voice_list = [
    "ko-KR-InJoonNeural",
    "zh-CN-XiaoxiaoMultilingualNeural",
    "en-GB-AdaMultilingualNeural",
]
stt.generate_synthetic_wav(
    mode="train", tts_voice_list=train_tts_voice_list, delete_old_data=True
)

# Augment the train data (Optional)
stt.augment_wav_files(num_augments=4)
# Package the train data to be used in the training pipeline
stt.package_trainset(use_augmented_data=True)

### Evaluation set
# Generate synthetic wav files for evaluation
eval_tts_voice_list = ["ko-KR-YuJinNeural"]
stt.generate_synthetic_wav(
    mode="eval", tts_voice_list=eval_tts_voice_list, delete_old_data=True
)
# Package the eval data to be used in the evaluation pipeline
stt.package_evalset(eval_dataset_dir="eval_dataset")

License Summary

This sample code is provided under the Apache 2.0 license. See the LICENSE file.