app.py

"""
app.py

A unified Gradio chat application for Multimodal OCR Granite Vision.
Commands (enter these as a prefix in the text input):
  - @rag: For retrieval‐augmented generation (e.g. PDF or text-based queries).
  - @granite: For image understanding.
  - @video-infer: For video understanding (video is downsampled into frames).

The app uses gr.MultimodalTextbox to support text input together with file uploads.
"""

import os
import time
import uuid
import random
import logging
from threading import Thread
from pathlib import Path
from datetime import datetime, timezone

import torch
import spaces
import numpy as np
import cv2
from PIL import Image
import gradio as gr

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    AutoProcessor,
    AutoModelForVision2Seq,
)

# ---------------------------
# Utility functions and setup
# ---------------------------

def get_device():
    if torch.backends.mps.is_available():
        return "mps"  # mac GPU
    elif torch.cuda.is_available():
        return "cuda"
    else:
        return "cpu"

device = get_device()

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
)

def downsample_video(video_path):
    """
    Downsamples the video into 10 evenly spaced frames.
    Returns a list of (PIL Image, timestamp in seconds) tuples.
    """
    vidcap = cv2.VideoCapture(video_path)
    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = vidcap.get(cv2.CAP_PROP_FPS)
    frames = []
    frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
    for i in frame_indices:
        vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
        success, image = vidcap.read()
        if success:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
            pil_image = Image.fromarray(image)
            timestamp = round(i / fps, 2)
            frames.append((pil_image, timestamp))
    vidcap.release()
    return frames

# ---------------------------
# HF Embedding and LLM classes
# ---------------------------

class HFEmbedding:
    def __init__(self, model_id: str):
        self.model_name = model_id
        logging.info(f"Loading embeddings model from: {self.model_name}")
        # Using langchain_huggingface for embeddings
        from langchain_huggingface import HuggingFaceEmbeddings  # ensure installed
        # For simplicity, force CPU (adjust if needed)
        self.embeddings_service = HuggingFaceEmbeddings(
            model_name=self.model_name,
            model_kwargs={"device": "cpu"},
        )

    def embed_documents(self, texts: list[str]) -> list[list[float]]:
        return self.embeddings_service.embed_documents(texts)

    def embed_query(self, text: str) -> list[float]:
        return self.embed_documents([text])[0]

class HFLLM:
    def __init__(self, model_name: str):
        self.device = device
        self.model_name = model_name
        logging.info("Loading HF language model...")
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForCausalLM.from_pretrained(model_name).to(self.device)

    def generate(self, prompt: str) -> list:
        # Tokenize prompt and generate text
        model_inputs = self.tokenizer(prompt, return_tensors="pt").to(self.device)
        generated_ids = self.model.generate(**model_inputs, max_new_tokens=1024)
        generated_texts = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=False)
        # Extract answer assuming a marker in the generated text
        response = [{"answer": generated_texts[0].split("<|end_of_role|>")[-1].split("<|end_of_text|>")[0]}]
        return response

# ---------------------------
# LightRAG: Retrieval-Augmented Generation (Dummy)
# ---------------------------

class LightRAG:
    def __init__(self, config: dict):
        self.config = config
        # Load generation and embedding models immediately (or lazy load as needed)
        self.gen_model = HFLLM(config['generation_model_id'])
        self._embedding_model = HFEmbedding(config['embedding_model_id'])

    def search(self, query: str, top_n: int = 5) -> list:
        # Dummy retrieval: In practice, integrate with a vector store
        from langchain_core.documents import Document  # ensure langchain_core is installed
        dummy_doc = Document(
            page_content="Dummy context for query: " + query,
            metadata={"type": "text"}
        )
        return [dummy_doc]

    def generate(self, query, context=None):
        if context is None:
            context = []
        # Build prompt by concatenating retrieved context with the query.
        prompt = ""
        for doc in context:
            prompt += doc.page_content + "\n"
        prompt += "\nQuestion: " + query + "\nAnswer:"
        results = self.gen_model.generate(prompt)
        answer = results[0]["answer"]
        return answer, prompt

# Global configuration for LightRAG
rag_config = {
    "embedding_model_id": "ibm-granite/granite-embedding-125m-english",
    "generation_model_id": "ibm-granite/granite-3.1-8b-instruct",
}
light_rag = LightRAG(rag_config)

# ---------------------------
# Granite Vision functions (for image and video)
# ---------------------------

# Set the Granite Vision model ID (adjust version as needed)
GRANITE_MODEL_ID = "ibm-granite/granite-vision-3.2-2b"
granite_processor = None
granite_model = None

def load_granite_model():
    """Lazy load the Granite vision processor and model."""
    global granite_processor, granite_model
    if granite_processor is None or granite_model is None:
        granite_processor = AutoProcessor.from_pretrained(GRANITE_MODEL_ID)
        granite_model = AutoModelForVision2Seq.from_pretrained(GRANITE_MODEL_ID, device_map="auto").to(device)
    return granite_processor, granite_model

def create_single_turn(image, text):
    """
    Creates a single-turn conversation message.
    If an image is provided, it is added along with the text.
    """
    if image is None:
        return {"role": "user", "content": [{"type": "text", "text": text}]}
    else:
        return {"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": text}]}

def generate_granite(image, prompt_text, max_new_tokens=1024, temperature=0.7, top_p=0.85, top_k=50, repetition_penalty=1.05):
    """
    Generates a response from the Granite Vision model given an image and prompt.
    """
    processor, model = load_granite_model()
    conversation = [create_single_turn(image, prompt_text)]
    inputs = processor.apply_chat_template(
        conversation, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
    ).to(device)
    output = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
    )
    decoded = processor.decode(output[0], skip_special_tokens=True)
    parts = decoded.strip().split("<|assistant|>")
    return parts[-1].strip()

def generate_video_infer(video_path, prompt_text, max_new_tokens=1024, temperature=0.7, top_p=0.85, top_k=50, repetition_penalty=1.05):
    """
    Processes a video file by downsampling frames and sending them along with a prompt
    to the Granite Vision model.
    """
    frames = downsample_video(video_path)
    conversation_content = []
    for img, ts in frames:
        conversation_content.append({"type": "text", "text": f"Frame at {ts} sec:"})
        conversation_content.append({"type": "image", "image": img})
    conversation_content.append({"type": "text", "text": prompt_text})
    conversation = [{"role": "user", "content": conversation_content}]
    processor, model = load_granite_model()
    inputs = processor.apply_chat_template(
        conversation, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
    ).to(device)
    output = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
    )
    decoded = processor.decode(output[0], skip_special_tokens=True)
    parts = decoded.strip().split("<|assistant|>")
    return parts[-1].strip()

# ---------------------------
# Unified generation function for ChatInterface
# ---------------------------
@spaces.GPU
def generate(input_dict: dict, chat_history: list[dict],
             max_new_tokens: int, temperature: float,
             top_p: float, top_k: int, repetition_penalty: float):
    """
    Chat function that inspects the input text for special commands and routes:
      - @rag: Uses the RAG pipeline.
      - @granite: Uses Granite Vision for image understanding.
      - @video-infer: Uses Granite Vision for video processing.
    """
    text = input_dict["text"]
    files = input_dict.get("files", [])
    lower_text = text.strip().lower()

    # Optionally yield a progress message
    yield "Processing your request..."
    time.sleep(1)  # simulate processing delay

    if lower_text.startswith("@rag"):
        query = text[len("@rag"):].strip()
        logging.info(f"@rag command: {query}")
        context = light_rag.search(query)
        answer, _ = light_rag.generate(query, context)
        yield answer

    elif lower_text.startswith("@granite"):
        prompt_text = text[len("@granite"):].strip()
        logging.info(f"@granite command: {prompt_text}")
        if files:
            # Expecting an image file (as a PIL image)
            image = files[0]
            answer = generate_granite(image, prompt_text, max_new_tokens, temperature, top_p, top_k, repetition_penalty)
            yield answer
        else:
            yield "No image provided for @granite command."

    elif lower_text.startswith("@video-infer"):
        prompt_text = text[len("@video-infer"):].strip()
        logging.info(f"@video-infer command: {prompt_text}")
        if files:
            # Expecting a video file (the file path)
            video_path = files[0]
            answer = generate_video_infer(video_path, prompt_text, max_new_tokens, temperature, top_p, top_k, repetition_penalty)
            yield answer
        else:
            yield "No video provided for @video-infer command."

    else:
        # Default behavior: use RAG pipeline for text query.
        query = text.strip()
        logging.info(f"Default text query: {query}")
        context = light_rag.search(query)
        answer, _ = light_rag.generate(query, context)
        yield answer

# ---------------------------
# Gradio ChatInterface using MultimodalTextbox
# ---------------------------

demo = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Slider(label="Max new tokens", minimum=1, maximum=2048, step=1, value=1024),
        gr.Slider(label="Temperature", minimum=0.1, maximum=2.0, step=0.1, value=0.7),
        gr.Slider(label="Top-p", minimum=0.1, maximum=1.0, step=0.1, value=0.85),
        gr.Slider(label="Top-k", minimum=1, maximum=100, step=1, value=50),
        gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.05),
    ],
    examples=[
        # Examples show how to use the command prefixes.
        [{"text": "@rag What models are available in Watsonx?"}],
        [{"text": "@granite Describe the image", "files": [str(Path("examples") / "sample_image.png")]}],
        [{"text": "@video-infer Summarize the event in the video", "files": [str(Path("examples") / "sample_video.mp4")]}],
    ],
    cache_examples=False,
    type="messages",
    description=(
        "# **Multimodal OCR Granite Vision**\n\n"
        "Enter a command in the text input (with optional file uploads) using one of the following prefixes:\n\n"
        "- **@rag**: For retrieval-augmented generation (e.g. PDFs, documents).\n"
        "- **@granite**: For image understanding using Granite Vision.\n"
        "- **@video-infer**: For video understanding (video is downsampled into frames).\n\n"
        "For example:\n```\n@rag What is the revenue trend?\n```\n```\n@granite Describe this image\n```\n```\n@video-infer Summarize the event in this video\n```"
    ),
    fill_height=True,
    textbox=gr.MultimodalTextbox(
        label="Query Input",
        file_types=["image", "video", "pdf"],
        file_count="multiple",
        placeholder="@rag, @granite, or @video-infer followed by your prompt"
    ),
    stop_btn="Stop Generation",
    multimodal=True,
)

if __name__ == "__main__":
    demo.queue(max_size=20).launch()