Upload app.py

app.py

@@ -0,0 +1,207 @@
+import os
+import fitz  # PyMuPDF
+import fasttext
+import requests
+import json
+import torch
+from PIL import Image
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+from IndicTransToolkit.processor import IndicProcessor
+import google.generativeai as genai
+import gradio as gr
+
+# === 1. CONFIGURATION & SECRETS ===
+# --- Load the Gemini API Key from Hugging Face Secrets ---
+GEMINI_API_KEY = os.environ.get("GEMINI_API_KEY")
+
+# --- Model IDs (Using the CPU-friendly TrOCR model) ---
+TRANSLATION_MODEL_REPO_ID = "ai4bharat/indictrans2-indic-en-1B"
+OCR_MODEL_ID = "microsoft/trocr-base-printed"
+
+# --- Language Settings ---
+LANGUAGE_TO_TRANSLATE = "mal"
+
+# --- Hardware Settings ---
+DEVICE = "cpu"  # Forcing CPU for compatibility with free tier
+
+# === 2. LOAD MODELS & CONFIGURE API ===
+# --- Configure Gemini API ---
+if not GEMINI_API_KEY:
+    print("🔴 ERROR: Gemini API key is not set in the Space Secrets.")
+else:
+    genai.configure(api_key=GEMINI_API_KEY)
+
+# --- Load Translation Model ---
+print(f"Loading tokenizer & model: {TRANSLATION_MODEL_REPO_ID} ...")
+translation_tokenizer = AutoTokenizer.from_pretrained(TRANSLATION_MODEL_REPO_ID, trust_remote_code=True)
+translation_model = AutoModelForSeq2SeqLM.from_pretrained(
+    TRANSLATION_MODEL_REPO_ID,
+    trust_remote_code=True,
+    torch_dtype=torch.float32  # Use float32 for CPU
+).to(DEVICE)
+print("✅ Translation model loaded.")
+ip = IndicProcessor(inference=True)
+
+# --- Load Language Detection Model ---
+print("Loading fastText language detector...")
+ft_model_path = hf_hub_download(repo_id="facebook/fasttext-language-identification", filename="model.bin")
+lang_detect_model = fasttext.load_model(ft_model_path)
+print("✅ fastText loaded.")
+
+# --- Load Standard OCR Model ---
+print(f"Loading Standard OCR model: {OCR_MODEL_ID}...")
+ocr_pipeline = pipeline("image-to-text", model=OCR_MODEL_ID, device=-1)  # device=-1 ensures CPU
+print("✅ Standard OCR model loaded.")
+
+
+# === 3. HELPER FUNCTIONS ===
+
+# --- Phase 1: Text Extraction ---
+def classify_image_with_gemini(image: Image.Image):
+    """Uses Gemini to classify an image as a 'document' or 'diagram'."""
+    model = genai.GenerativeModel('gemini-1.5-flash-latest')
+    prompt = "Is this image primarily a text document or an engineering/technical diagram? Answer with only 'document' or 'diagram'."
+    response = model.generate_content([prompt, image])
+    classification = response.text.strip().lower()
+    print(f"✅ Image classified as: {classification}")
+    return "diagram" if "diagram" in classification else "document"
+
+def summarize_diagram_with_gemini(image: Image.Image):
+    """Uses Gemini to generate a summary of an engineering diagram."""
+    model = genai.GenerativeModel('gemini-1.5-flash-latest')
+    prompt = "You are an engineering assistant for Kochi Metro Rail Limited (KMRL). Describe the contents of this technical diagram or engineering drawing in a concise summary. Identify key components and their apparent purpose."
+    response = model.generate_content([prompt, image])
+    print("✅ Diagram summary successful.")
+    return response.text.strip()
+
+def extract_text_from_image(path):
+    """
+    Classifies an image and routes it for either OCR (if a text doc) or summarization (if a diagram).
+    """
+    print("\n--- Starting Image Processing ---")
+    try:
+        image = Image.open(path).convert("RGB")
+        
+        # Step 1: Classify the image using Gemini
+        image_type = classify_image_with_gemini(image)
+        
+        # Step 2: Route to the correct function
+        if image_type == "diagram":
+            print("-> Image is a diagram. Summarizing with Gemini...")
+            return summarize_diagram_with_gemini(image)
+        else:
+            print("-> Image is a document. Extracting text with TrOCR...")
+            out = ocr_pipeline(image)
+            return out[0]["generated_text"] if out else ""
+            
+    except Exception as e:
+        print(f"❌ An error occurred during image processing: {e}")
+        return "Error during image processing."
+
+def extract_text_from_pdf(path):
+    doc = fitz.open(path)
+    return "".join(page.get_text("text") + "\n" for page in doc)
+
+def read_text_from_txt(path):
+    with open(path, "r", encoding="utf-8") as f:
+        return f.read()
+
+# --- Phase 2: Translation ---
+def detect_language(text_snippet):
+    s = text_snippet.replace("\n", " ").strip()
+    if not s: return None
+    preds = lang_detect_model.predict(s, k=1)
+    return preds[0][0].split("__")[-1] if preds and preds[0] else None
+
+def translate_chunk(chunk):
+    batch = ip.preprocess_batch([chunk], src_lang="mal_Mlym", tgt_lang="eng_Latn")
+    inputs = translation_tokenizer(batch, return_tensors="pt", padding=True, truncation=True, max_length=512).to(DEVICE)
+    with torch.no_grad():
+        generated_tokens = translation_model.generate(**inputs, num_beams=5, max_length=512, early_stopping=True)
+    decoded = translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+    return ip.postprocess_batch(decoded, lang=tgt_lang)[0]
+
+# --- Phase 3: Gemini Analysis ---
+def generate_structured_json(text_to_analyze):
+    """Generates the detailed JSON analysis."""
+    model = genai.GenerativeModel('gemini-1.5-flash-latest')
+    prompt = f"You are an AI assistant for KMRL. Analyze this document and extract key info as JSON: {text_to_analyze}"
+    json_schema = {"type": "OBJECT", "properties": {"summary": {"type": "STRING"}, "actions_required": {"type": "ARRAY", "items": {"type": "OBJECT", "properties": {"action": {"type": "STRING"}, "priority": {"type": "STRING", "enum": ["High", "Medium", "Low"]}, "deadline": {"type": "STRING"}, "notes": {"type": "STRING"}}, "required": ["action", "priority", "deadline", "notes"]}}, "departments_to_notify": {"type": "ARRAY", "items": {"type": "STRING"}}, "cross_document_flags": {"type": "ARRAY", "items": {"type": "OBJECT", "properties": {"related_document_type": {"type": "STRING"}, "related_issue": {"type": "STRING"}}, "required": ["related_document_type", "related_issue"]}}}, "required": ["summary", "actions_required", "departments_to_notify", "cross_document_flags"]}
+    generation_config = genai.types.GenerationConfig(response_mime_type="application/json", response_schema=json_schema)
+    response = model.generate_content(prompt, generation_config=generation_config)
+    return json.loads(response.text)
+
+def check_relevance_with_gemini(summary_text):
+    """Checks if the summary is relevant to KMRL."""
+    model = genai.GenerativeModel('gemini-1.5-flash-latest')
+    prompt = f'Is this summary related to transportation, infrastructure, railways, or metro systems? Answer only "Yes" or "No".\n\nSummary: {summary_text}'
+    response = model.generate_content(prompt)
+    return "yes" in response.text.strip().lower()
+
+# === 4. MAIN PROCESSING FUNCTION FOR GRADIO ===
+def process_and_analyze_document(input_file):
+    if not GEMINI_API_KEY:
+        raise gr.Error("Gemini API key is not configured. The administrator must set it in the Space Secrets.")
+    if input_file is None:
+        raise gr.Error("No file uploaded. Please upload a document.")
+    
+    try:
+        input_file_path = input_file.name
+        ext = os.path.splitext(input_file_path)[1].lower()
+        
+        # --- Phase 1: Get Original Text ---
+        if ext == ".pdf":
+            original_text = extract_text_from_pdf(input_file_path)
+        elif ext == ".txt":
+            original_text = read_text_from_txt(input_file_path)
+        elif ext in [".png", ".jpg", ".jpeg"]:
+            original_text = extract_text_from_image(input_file_path)
+        else:
+            raise gr.Error("Unsupported file type.")
+        
+        if not original_text or not original_text.strip():
+            raise gr.Error("No text could be extracted from the document.")
+
+        # --- Phase 2: Translate if Necessary ---
+        lines = original_text.split("\n")
+        translated_lines = []
+        for ln in lines:
+            if not ln.strip(): continue
+            lang = detect_language(ln)
+            if lang == LANGUAGE_TO_TRANSLATE:
+                translated_lines.append(translate_chunk(ln))
+            else:
+                translated_lines.append(ln)
+        final_text = "\n".join(translated_lines)
+
+        # --- Phase 3: Analyze with Gemini ---
+        summary_data = generate_structured_json(final_text)
+        if not summary_data or "summary" not in summary_data:
+            raise gr.Error("Failed to generate a valid analysis from the document.")
+
+        is_relevant = check_relevance_with_gemini(summary_data["summary"])
+
+        if is_relevant:
+            return summary_data
+        else:
+            return {"status": "Not Applicable", "reason": "The document was determined to be not relevant to KMRL."}
+
+    except Exception as e:
+        raise gr.Error(f"An unexpected error occurred: {str(e)}")
+
+
+iface = gr.Interface(
+    fn=process_and_analyze_document,
+    inputs=gr.File(label="Upload Document (.pdf, .txt, .png, .jpeg)"),
+    outputs=gr.JSON(label="Analysis Result"),
+    title="KMRL Document Analysis Pipeline",
+    description="Upload a document (Malayalam or English). The system will detect and translate Malayalam text to English, then send the full text to Gemini for structured analysis.",
+    allow_flagging="never",
+    examples=[
+        ["Malayalam-en.txt"] # If you upload this file to your Space
+    ]
+)
+
+if __name__ == "__main__":
+    iface.launch()