app.py

# ---------------------------------------------------------------------------------------
# Imports and Options
# ---------------------------------------------------------------------------------------
import streamlit as st
import pandas as pd
import requests
import re
import fitz  # PyMuPDF
import io
import matplotlib.pyplot as plt
from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq
from docling_core.types.doc import DoclingDocument
from docling_core.types.doc.document import DocTagsDocument
import torch

# ---------------------------------------------------------------------------------------
# Streamlit Page Configuration
# ---------------------------------------------------------------------------------------
st.set_page_config(
    page_title="Choose Your Own Adventure (Topic Extraction) PDF Analysis App",
    page_icon=":bar_chart:",
    layout="centered",
    initial_sidebar_state="auto",
    menu_items={
        'Get Help': 'mailto:[email protected]',
        'About': "This app is built to support PDF analysis"
    }
)

# ---------------------------------------------------------------------------------------
# Session State Initialization
# ---------------------------------------------------------------------------------------
for key in ['pdf_processed', 'markdown_texts', 'df']:
    if key not in st.session_state:
        st.session_state[key] = False if key == 'pdf_processed' else []

# ---------------------------------------------------------------------------------------
# API Configuration
# ---------------------------------------------------------------------------------------
API_URL = "https://api.stack-ai.com/inference/v0/run/2df89a6c-a4af-4576-880e-27058e498f02/67acad8b0603ba4631db38e7"
headers = {
    'Authorization': 'Bearer a9e4979e-cdbe-49ea-a193-53562a784805',
    'Content-Type': 'application/json'
}

# ---------------------------------------------------------------------------------------
# Survey Analysis Class
# ---------------------------------------------------------------------------------------
class SurveyAnalysis:
    def prepare_llm_input(self, survey_response, topics):
        topic_descriptions = "\n".join([f"- **{t}**: {d}" for t, d in topics.items()])
        return f"""Extract and summarize PDF notes based on topics:
{topic_descriptions}

Instructions:
- Extract exact quotes per topic.
- Ignore irrelevant topics.

Format:
[Topic]
- "Exact quote"

Meeting Notes:
{survey_response}
"""

    def query_api(self, payload):
        try:
            res = requests.post(API_URL, headers=headers, json=payload, timeout=60)
            res.raise_for_status()
            return res.json()
        except requests.exceptions.RequestException as e:
            st.error(f"API request failed: {e}")
            return {'outputs': {'out-0': ''}}

    def extract_meeting_notes(self, response):
        return response.get('outputs', {}).get('out-0', '')

    def process_dataframe(self, df, topics):
        results = []
        for _, row in df.iterrows():
            llm_input = self.prepare_llm_input(row['Document_Text'], topics)
            payload = {"user_id": "user", "in-0": llm_input}
            response = self.query_api(payload)
            notes = self.extract_meeting_notes(response)
            results.append({'Document_Text': row['Document_Text'], 'Topic_Summary': notes})
        return pd.concat([df.reset_index(drop=True), pd.DataFrame(results)['Topic_Summary']], axis=1)

# ---------------------------------------------------------------------------------------
# Helper Functions
# ---------------------------------------------------------------------------------------
@st.cache_resource
def load_smol_docling():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
    model = AutoModelForVision2Seq.from_pretrained(
        "ds4sd/SmolDocling-256M-preview", torch_dtype=torch.float32
    ).to(device)
    return model, processor

model, processor = load_smol_docling()

def convert_pdf_to_images(pdf_file, dpi=150, max_size=1600):
    images = []
    doc = fitz.open(stream=pdf_file.read(), filetype="pdf")
    for page in doc:
        pix = page.get_pixmap(dpi=dpi)
        img = Image.open(io.BytesIO(pix.tobytes("png"))).convert("RGB")
        img.thumbnail((max_size, max_size), Image.LANCZOS)
        images.append(img)
    return images

def extract_markdown_from_image(image):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    prompt = processor.apply_chat_template([{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Convert this page to docling."}]}], add_generation_prompt=True)
    inputs = processor(text=prompt, images=[image], return_tensors="pt").to(device)
    with torch.no_grad():
        generated_ids = model.generate(**inputs, max_new_tokens=1024)
    doctags = processor.batch_decode(generated_ids[:, inputs.input_ids.shape[1]:], skip_special_tokens=False)[0].replace("<end_of_utterance>", "").strip()
    doctags_doc = DocTagsDocument.from_doctags_and_image_pairs([doctags], [image])
    doc = DoclingDocument(name="ExtractedDocument")
    doc.load_from_doctags(doctags_doc)
    return doc.export_to_markdown()

def extract_excerpts(processed_df):
    rows = []
    for _, r in processed_df.iterrows():
        for sec in re.split(r'\n(?=\[)', r['Topic_Summary']):
            topic_match = re.match(r'\[([^\]]+)\]', sec)
            if topic_match:
                topic = topic_match.group(1)
                excerpts = re.findall(r'- "([^"]+)"', sec)
                for excerpt in excerpts:
                    rows.append({'Document_Text': r['Document_Text'], 'Topic_Summary': r['Topic_Summary'], 'Excerpt': excerpt, 'Topic': topic})
    return pd.DataFrame(rows)

# ---------------------------------------------------------------------------------------
# Streamlit UI
# ---------------------------------------------------------------------------------------
st.title("Choose Your Own Adventure (Topic Extraction) PDF Analysis App")

uploaded_file = st.file_uploader("Upload PDF file", type=["pdf"])

if uploaded_file and not st.session_state['pdf_processed']:
    with st.spinner("Processing PDF..."):
        images = convert_pdf_to_images(uploaded_file)
        markdown_texts = [extract_markdown_from_image(img) for img in images]
        st.session_state['df'] = pd.DataFrame({'Document_Text': markdown_texts})
        st.session_state['pdf_processed'] = True
    st.success("PDF processed successfully!")

if st.session_state['pdf_processed']:
    st.markdown("### Extracted Text Preview")
    st.write(st.session_state['df'].head())

    st.markdown("### Enter Topics and Descriptions")
    num_topics = st.number_input("Number of topics", 1, 10, 1)
    topics = {}
    for i in range(num_topics):
        topic = st.text_input(f"Topic {i+1} Name", key=f"topic_{i}")
        desc = st.text_area(f"Topic {i+1} Description", key=f"description_{i}")
        if topic and desc:
            topics[topic] = desc

    if st.button("Run Analysis"):
        if not topics:
            st.warning("Please enter at least one topic and description.")
            st.stop()

        analyzer = SurveyAnalysis()
        processed_df = analyzer.process_dataframe(st.session_state['df'], topics)
        extracted_df = extract_excerpts(processed_df)

        st.markdown("### Extracted Excerpts")
        st.dataframe(extracted_df)

        csv = extracted_df.to_csv(index=False)
        st.download_button("Download CSV", csv, "extracted_notes.csv", "text/csv")

        topic_counts = extracted_df['Topic'].value_counts()
        fig, ax = plt.subplots()
        topic_counts.plot.bar(ax=ax, color='#3d9aa1')
        st.pyplot(fig)

if st.button("Reset / Upload New PDF"):
    for key in ['pdf_processed', 'markdown_texts', 'df']:
        st.session_state[key] = False if key == 'pdf_processed' else []
    st.experimental_rerun()

if not uploaded_file:
    st.info("Please upload a PDF file to begin.")