app.py

import streamlit as st
import os
import json
import faiss
import numpy as np
from vosk import Model, KaldiRecognizer
import wave
from pydub import AudioSegment
from sentence_transformers import SentenceTransformer
from transformers import pipeline

# ------------------------------
# 1. LOAD MODELS OFFLINE
# ------------------------------

@st.cache_resource
def load_stt_model():
    model_path = "vosk-model-small-en-us-0.15"
    return Model(model_path)

@st.cache_resource
def load_embedding_model():
    return SentenceTransformer("all-MiniLM-L6-v2")

@st.cache_resource
def load_qa_model():
    return pipeline("question-answering", model="distilbert-base-cased-distilled-squad")

stt_model = load_stt_model()
embedder = load_embedding_model()
qa_model = load_qa_model()

# ------------------------------
# 2. FUNCTIONS
# ------------------------------

def transcribe_audio(file_path):
    wf = wave.open(file_path, "rb")
    rec = KaldiRecognizer(stt_model, wf.getframerate())
    rec.SetWords(True)

    text_result = ""

    while True:
        data = wf.readframes(4000)
        if len(data) == 0:
            break
        if rec.AcceptWaveform(data):
            res = rec.Result()
            text_result += json.loads(res)["text"] + " "

    final_res = rec.FinalResult()
    text_result += json.loads(final_res)["text"]

    return text_result.strip()


def convert_to_wav(uploaded_file):
    audio = AudioSegment.from_file(uploaded_file)
    output_path = "temp.wav"
    audio.export(output_path, format="wav")
    return output_path


def save_text(text):
    with open(f"transcripts/data.txt", "a", encoding="utf-8") as f:
        f.write(text + "\n")


def build_vector_db():
    with open("transcripts/data.txt", "r", encoding="utf-8") as f:
        docs = f.readlines()

    embeddings = embedder.encode(docs)
    index = faiss.IndexFlatL2(embeddings.shape[1])
    index.add(np.array(embeddings).astype("float32"))
    return docs, index


def retrieve(query, docs, index, top_k=3):
    q_emb = embedder.encode([query])
    D, I = index.search(np.array(q_emb).astype("float32"), top_k)
    results = [docs[i] for i in I[0]]
    return " ".join(results)


# ------------------------------
# 3. STREAMLIT UI
# ------------------------------

st.title("🔴 Offline GenAI RAG from Audio (No API • No Internet)")
st.write("🎤 Upload or record audio → 📄 Convert to text → 🤖 Ask questions offline")

menu = st.sidebar.radio("Navigation", ["Upload Audio", "Ask Questions"])

# -----------------------------------
#  UPLOAD AUDIO PAGE
# -----------------------------------
if menu == "Upload Audio":
    st.header("🎤 Upload or Record Audio")

    audio_file = st.file_uploader("Upload audio file", type=["wav", "mp3", "m4a"])

    if audio_file:
        st.success("File uploaded successfully")

        wav_path = convert_to_wav(audio_file)

        st.info("Transcribing offline... please wait")

        text = transcribe_audio(wav_path)

        st.subheader("📝 Transcribed Text")
        st.write(text)

        save_text(text)

        st.success("Saved locally in transcripts/ folder")

# -----------------------------------
#  ASK QUESTION PAGE
# -----------------------------------
if menu == "Ask Questions":
    st.header("❓ Ask Questions From Your Audio Knowledge Base")

    docs, index = build_vector_db()

    user_q = st.text_input("Enter your question")

    if st.button("Get Answer"):
        context = retrieve(user_q, docs, index)
        result = qa_model(question=user_q, context=context)

        st.subheader("🧠 Answer")
        st.write(result["answer"])

        st.caption("Based only on your stored audio transcriptions")