app.py

import gradio as gr
import torch
import requests
import tempfile
import threading
import numpy as np
from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import Neo4jVector
from langchain_community.graphs import Neo4jGraph
from langchain_core.prompts import ChatPromptTemplate
import time
import os
from dataclasses import dataclass

# Define AppState to store audio state information
@dataclass
class AppState:
    stream: np.ndarray | None = None
    sampling_rate: int = 0
    pause_detected: bool = False
    started_talking: bool = False

# Neo4j setup
graph = Neo4jGraph(
    url="neo4j+s://c62d0d35.databases.neo4j.io",
    username="neo4j",
    password="_x8f-_aAQvs2NB0x6s0ZHSh3W_y-HrENDbgStvsUCM0"
)

# Initialize the vector index with Neo4j
vector_index = Neo4jVector.from_existing_graph(
    OpenAIEmbeddings(api_key=os.environ['OPENAI_API_KEY']),
    graph=graph,
    search_type="hybrid",
    node_label="Document",
    text_node_properties=["text"],
    embedding_node_property="embedding",
)

# Define the ASR model with Whisper
model_id = 'openai/whisper-large-v3'
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype).to(device)
processor = AutoProcessor.from_pretrained(model_id)

pipe_asr = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=15,
    batch_size=16,
    torch_dtype=torch_dtype,
    device=device,
    return_timestamps=True
)

# Adjusted function to determine if a pause occurred
def determine_pause(audio: np.ndarray, sampling_rate: int, state: AppState) -> bool:
    """Take in the stream, determine if a pause happened."""
    temp_audio = audio
    dur_vad = len(temp_audio) / sampling_rate  # Simulating VAD duration for this example
    duration = len(audio) / sampling_rate

    # Log the duration and VAD result for debugging
    print(f"Duration after VAD: {dur_vad:.3f} s, Total Duration: {duration:.3f} s")

    # Check if speech has started
    if dur_vad > 0.5 and not state.started_talking:
        print("Started talking")
        state.started_talking = True
        return False

    # If the difference between total duration and VAD duration is significant, consider it a pause
    # Adjust the threshold for pause detection (e.g., 0.5 seconds)
    pause_threshold = 0.5  # This value can be adjusted to be more sensitive
    if (duration - dur_vad) > pause_threshold and state.started_talking:
        print("Pause detected")
        return True

    return False


# Function to process audio input, detect pauses, and handle state
def process_audio(audio: tuple, state: AppState):
    # Ensure audio input is not None and has valid data
    if audio is None or audio[1] is None:
        print("Audio input is None or empty.")
        return None, state

    if state.stream is None:
        state.stream = audio[1]
        state.sampling_rate = audio[0]
    else:
        state.stream = np.concatenate((state.stream, audio[1]))

    # Check for a pause in speech
    pause_detected = determine_pause(state.stream, state.sampling_rate, state)
    state.pause_detected = pause_detected

    if state.pause_detected and state.started_talking:
        # Transcribe the audio when a pause is detected
        _, transcription, _ = transcribe_function(state.stream, (state.sampling_rate, state.stream))
        print(f"Transcription: {transcription}")

        # Check if transcription is empty
        if not transcription:
            print("No transcription available.")
            return None, state

        # Retrieve hybrid response using Neo4j and other methods
        response_text = retriever(transcription)
        print(f"Response: {response_text}")

        # Check if the response is empty before proceeding
        if not response_text:
            print("No response generated.")
            return None, state

        # Generate audio from the response text
        audio_path = generate_audio_elevenlabs(response_text)

        # Reset state for the next input
        state.stream = None
        state.started_talking = False
        state.pause_detected = False

        return audio_path, state

    return None, state


# Function to process audio input and transcribe it
def transcribe_function(stream, new_chunk):
    try:
        sr, y = new_chunk[0], new_chunk[1]
    except TypeError:
        print(f"Error chunk structure: {type(new_chunk)}, content: {new_chunk}")
        return stream, "", None

    if y is None or len(y) == 0:
        return stream, "", None

    y = y.astype(np.float32)
    max_abs_y = np.max(np.abs(y))
    if max_abs_y > 0:
        y = y / max_abs_y

    if stream is not None and len(stream) > 0:
        stream = np.concatenate([stream, y])
    else:
        stream = y

    result = pipe_asr({"array": stream, "sampling_rate": sr}, return_timestamps=False)
    full_text = result.get("text", "")

    return stream, full_text, full_text

# Function to generate a full-text search query for Neo4j
def generate_full_text_query(input: str) -> str:
    words = [el for el in input.split() if el]
    if not words:
        return ""  # Return an empty string or a default query if desired
    full_text_query = ""
    for word in words[:-1]:
        full_text_query += f" {word}~2 AND"
    full_text_query += f" {words[-1]}~2"
    return full_text_query.strip()

# Function to generate audio with Eleven Labs TTS
def generate_audio_elevenlabs(text):
    XI_API_KEY = os.environ['ELEVENLABS_API']
    VOICE_ID = 'ehbJzYLQFpwbJmGkqbnW'
    tts_url = f"https://api.elevenlabs.io/v1/text-to-speech/{VOICE_ID}/stream"
    headers = {
        "Accept": "application/json",
        "xi-api-key": XI_API_KEY
    }
    data = {
        "text": str(text),
        "model_id": "eleven_multilingual_v2",
        "voice_settings": {
            "stability": 1.0,
            "similarity_boost": 0.0,
            "style": 0.60,
            "use_speaker_boost": False
        }
    }
    response = requests.post(tts_url, headers=headers, json=data, stream=True)
    if response.ok:
        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as f:
            for chunk in response.iter_content(chunk_size=1024):
                if chunk:
                    f.write(chunk)
            audio_path = f.name
        return audio_path
    else:
        print(f"Error generating audio: {response.text}")
        return None

# Define the template for generating responses based on context
template = """I am a guide for Birmingham, Alabama. I can provide recommendations and insights about the city, including events and activities.
Ask your question directly, and I'll provide a precise and quick, short and crisp response in a conversational and straightforward way without any Greet.
Context:
{context}

Question: {question}
Answer concisely:"""

# Create a prompt object using the template
prompt = ChatPromptTemplate.from_template(template)

# Function to generate a response using the prompt and the context
def generate_response_with_prompt(context, question):
    formatted_prompt = prompt.format(
        context=context,
        question=question
    )
    llm = ChatOpenAI(temperature=0, api_key=os.environ['OPENAI_API_KEY'])
    response = llm(formatted_prompt)
    return response.content.strip()

# Define the function to generate a hybrid response using Neo4j and other retrieval methods
def retriever(question: str):
    structured_query = f"""
    CALL db.index.fulltext.queryNodes('entity', $query, {{limit: 2}})
    YIELD node, score
    RETURN node.id AS entity, node.text AS context, score
    ORDER BY score DESC
    LIMIT 2
    """
    structured_data = graph.query(structured_query, {"query": generate_full_text_query(question)})
    structured_response = "\n".join([f"{record['entity']}: {record['context']}" for record in structured_data])

    unstructured_data = [el.page_content for el in vector_index.similarity_search(question)]
    unstructured_response = "\n".join(unstructured_data)

    combined_context = f"Structured data:\n{structured_response}\n\nUnstructured data:\n{unstructured_response}"
    final_response = generate_response_with_prompt(combined_context, question)
    return final_response

# Create Gradio interface for audio input and output
interface = gr.Interface(
    fn=lambda audio, state: process_audio(audio, state),
    inputs=[
        gr.Audio(sources="microphone", type="numpy", streaming=True),
        gr.State(AppState())
    ],
    outputs=[
        gr.Audio(type="filepath", autoplay=True, interactive=False),
        gr.State()
    ],
    live=True,
    description="Ask questions via audio and receive audio responses.",
    allow_flagging="never"
)

# Launch the Gradio app
interface.launch()