app.py

from flask import Flask, jsonify, request, render_template
import pandas as pd
from flask_cors import CORS
import os
import json
import torch
from sentence_transformers import SentenceTransformer
import logging
from transformers import GPT2LMHeadModel, GPT2Tokenizer
from datetime import datetime
from transformers import BertTokenizer, BertForSequenceClassification
import random
import re

app = Flask(__name__)
app.json.sort_keys = False
CORS(app)

# Configure logging
logging.basicConfig(level=logging.DEBUG)


# Load the SentenceTransformer model
model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
print("---"*30)
print(device)

# Load tokenizer and model
loaded_model = BertForSequenceClassification.from_pretrained('saved_model')
loaded_tokenizer = BertTokenizer.from_pretrained('saved_model')

tokenizer = GPT2Tokenizer.from_pretrained("checkpoint-15000")
model_gpt = GPT2LMHeadModel.from_pretrained("checkpoint-15000")
model_gpt.to(device)
print("===="*20)

# df_case = pd.read_csv('case_clustering24.csv', on_bad_lines="skip")
grouped = pd.read_csv('grouped_22_23_24.csv', on_bad_lines="skip")


from openai import OpenAI
api_key = "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"
#api_key =  "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"
client = OpenAI(api_key = api_key)

def ask_gpt(question):
    # Using the text completion model
    response = client.chat.completions.create(
        model = "gpt-4o-mini",
        messages = [
            {"role":"system","content":"you are a helpful assistant."},
             {"role":"user","content": question}
            
        ]
    )
    return response.choices[0].message.content


class DataFrameManager:
    def __init__(self, file_path):
        self.file_path = file_path
        self.df = pd.DataFrame()  # Initialize an empty DataFrame
        self.load_dataframe()

    def load_dataframe(self):
        if os.path.exists(self.file_path):
            constant_date = "2024-01-01"
            constant_policy_number = "POL123456"
            constant_status = "Active"
            self.df = pd.read_csv(self.file_path)
            self.df['date'] = constant_date
            self.df['policy_number'] = constant_policy_number
            self.df['status'] = constant_status
            self.df = self.df.rename(columns={
                'note_id': 'id',
                'cleaned_comments': 'summary',
                'summarized_text': 'suggested_summary'
            })
        else:
            print(f"File not found: {self.file_path}")

    def get_dataframe(self):
        return self.df.copy().head(200)

df_manager = DataFrameManager('client_notes_Sneha.csv')

# Define a function for text generation
def generate_text(prompt_text, max_length=100,num_return_sequences=10):
    # Tokenize the prompt text and convert to tensor
    input_ids = tokenizer(prompt_text, return_tensors="pt").input_ids.to(device)
    attention_mask = tokenizer(
    prompt_text, return_tensors="pt").attention_mask.to(device)
    print("........")
    try:
        # Move input_ids and attention_mask tensor to GPU
        input_ids = input_ids.to(device)
        attention_mask = attention_mask.to(device)

        outputs = model_gpt.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        pad_token_id=tokenizer.pad_token_id,
        #max_length=10,
        max_new_tokens=3,
        num_beams=50,
        temperature=0.7,
        top_k=50,
        top_p=0.9,
        do_sample=True,
        num_return_sequences=num_return_sequences
        )
        print(outputs)
        print(",,,")
        # Decode the generated text
        generated_texts = [tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
        print(generated_texts)
        unique_texts = list(set(generated_texts))
        return unique_texts[:5]
    except Exception as e:
        print(str(e))


@app.route('/get-csv')
def get_csv():
    df = df_manager.get_dataframe()
    data = df.to_dict(orient='records')
    return jsonify(data)

@app.route('/search_notes', methods=['POST'])
def search_notes():
    request_data = request.get_json()
    claim_id_to_search = request_data.get('name', '')

    full_df = df_manager.get_dataframe()
    # print("DataFrame columns:", full_df.columns)  # Debug output
    # print("DataFrame first few rows:", full_df.head())  # Debug output

    if claim_id_to_search:
        try:
            claim_id_to_search = int(claim_id_to_search)  # Convert the claim ID from string to integer
            # print("Searching for ID:", claim_id_to_search)  # Debug output

            filtered_df = full_df[full_df['id'] == claim_id_to_search]
            # print("Filtered DataFrame:", filtered_df)  # Debug output

            if filtered_df.empty:
                print("No matching records found, returning full DataFrame.")
                return jsonify({})
            else:
                return jsonify(filtered_df.to_dict(orient='records'))
        except ValueError as  e:
            print(e)
            return jsonify({"error": "Invalid claim ID format"}), 400
    else:
        print("No claim ID provided, returning full DataFrame.")
        return jsonify({})


@app.route('/get-similarity', methods=['POST'])
def get_similarity():
    data = request.json
    logging.debug(f"Received payload: {data}")

    if not data or 'id' not in data:
        return jsonify({'error': 'No valid data provided'}), 400

    note_id = data['id']
    logging.debug(f"Note ID: {note_id}")
    df = df_manager.get_dataframe()
    print(df.columns)
    filtered_df = df[df['id'] == note_id]
    if filtered_df.empty:
        return jsonify({'error': 'No matching record found'}), 404

    row = filtered_df.iloc[0]
    summarized_text = row['suggested_summary']
    logging.debug(f"Summarized Text: {summarized_text}")

    # Encode the target summarized_text
    target_embedding = model.encode(summarized_text, convert_to_tensor=True, device=device).unsqueeze(0)

    # Calculate similarities with all entries in the suggested_summary column
    similarities = []
    for index, row in df.iterrows():
        text = row['suggested_summary']
        embedding = model.encode(text, convert_to_tensor=True, device=device).unsqueeze(0)
        similarity = torch.nn.functional.cosine_similarity(target_embedding, embedding).item()
        similarities.append({
            'id': row['id'],
            'status': row['status'],
            'policy_number': row['policy_number'],
            'date': row['date_created'].split(' ')[0],
            'summary': row['summary'],
            'suggested_summary': text,
            'similarity': similarity
        })

    # Convert the results to a dataframe
    similarity_df = pd.DataFrame(similarities)

    # Sort the dataframe by similarity in descending order
    similarity_df = similarity_df.sort_values(by='similarity', ascending=False)
    print(similarity_df.head())
    print(similarity_df.columns)
    # Convert the dataframe to a dictionary
    result = similarity_df.to_dict(orient='records')

    return jsonify(result), 200


@app.route('/autocomplete', methods=['GET'])
def auto_complete():
    data = request.args
    print("-----------------------")
    print(data)
    prompt = data.get('prompt', '')
    print(prompt)
    if not prompt:
        return jsonify({"error":"No prompt Provided"}),400
    try:
        print("====")
        generated_texts = generate_text(prompt)
        print(generated_texts)
        return jsonify({'generated_text': generated_texts})
    
    except Exception as e:
        return jsonify({"error":str(e)}),500
        

@app.route('/assign-case-id', methods=['POST'])
def classify_claim():
    data = request.json
    case_id = int(data['case_id'])
    # claim_line_id = data['claim_line_id']
    diagnosis = data['diagnosis']
    claim_line_note = data['claim_line_notes']
    print(data['service_date'])
    service_date = datetime.strptime(data['service_date'], "%Y-%m-%d")
    print("++++++"*30)
    print("dddddddd")

    # Convert all `case_id` values in `grouped` to integers
    grouped['case_id'] = grouped['case_id'].astype(int)

    record = grouped[grouped['case_id'] == case_id]


    # Check if case_id is present
    if case_id not in grouped['case_id'].values:
        new_case_id = random.randint(100000, 999999)
        return jsonify({"message": f"New Case: Customer id {case_id} not found. \n"
                                   f"A new case has been created with Case ID: {new_case_id}." })

    # Check if the record is empty and return an appropriate response
    if record.empty:
        return jsonify({"error": "No record found for the given case_id"}), 404
    

    print("--"*2)
    print(record['service_date'])
    # Compare service_date
    existing_service_date = datetime.strptime(eval(record['service_date'].values[0])[-1], "%Y-%m-%d")
    #existing_service_date = datetime.strptime(eval(record['service_date'].values[0])[-1],'%d-%m-%Y')
    print("-")
    print(existing_service_date)
    print("-")
    print(service_date)
    #
    is_recent = (service_date - existing_service_date).days < 90
    print(is_recent)

    if case_id in grouped['case_id'].values and not is_recent:
        return jsonify({
            "message": (
                f"New case (Customer id {case_id} found, however service date is more than 90 days), "
                f"Last service date: {existing_service_date.strftime('%Y-%m-%d')}"
            )
        })
    

    # history for bert
    past_claims_data = {}
    for _, row in record.iterrows():
        case_id = row['case_id']  # Extract the case_id for reference
        num_claims = len(row['service_date'])
        
        # Create sequences of claims within the same case
        for i in range(1, num_claims):
            row['claim_line_note'] = str([i for i in eval(row['claim_line_note'],{'nan':'nan'}) if i != 'nan'])
            input_sequence = (
                f"Diagnosis History: {', '.join(map(str, eval(row['diagnosis'])))}, "
                f"Claim Line Notes History: {', '.join(map(str, eval(row['claim_line_note'])))}, "
                f"Service Dates History: {', '.join(map(str, eval(row['service_date'])))}")
        
            past_claims_data["input_sequence"]= input_sequence


    # history for llm
    past_claims_data_llm = {}
    for _, row in record.iterrows():
        case_id = row['case_id']  # Extract the case_id for reference
        num_claims = len(row['service_date'])
        
        # Create sequences of claims within the same case
        for i in range(1, num_claims):
            row['claim_line_note'] = str([i for i in eval(row['claim_line_note'],{'nan':'nan'}) if i != 'nan'])
            past_claims_data_llm["Diagnosis History"]= ', '.join(map(str, eval(row['diagnosis'])))
            past_claims_data_llm["Claim Line Notes"]= ', '.join(map(str, eval(row['claim_line_note'])))


    print("***********************Past claim History***********************")
    print(past_claims_data['input_sequence'])
    print()

    
    # new claim info
    new_claim = (
        f"New Diagnosis: {', '.join(map(str, [diagnosis]))}, "
        f"New Claim Line Note: {', '.join(map(str, [claim_line_note]))}, "
        f"New Service Date: {', '.join(map(str, [service_date]))}"
    )
    
    print("***********************New claim Data***********************")
    print(new_claim)
    print("***********************")

    # Tokenize the test data
    inputs = loaded_tokenizer(past_claims_data['input_sequence'], new_claim, padding=True, truncation=True, return_tensors="pt")

    # Get model predictions
    with torch.no_grad():
        outputs = loaded_model(**inputs)
        predictions = torch.argmax(outputs.logits, dim=-1)
    bert_probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)

    pred_label = predictions.tolist()[0]
    new_case_id = random.randint(100000, 999999)
    new_claim_id = random.randint(100000, 999999)
    # Generate final output based on prediction
    if pred_label == 1:
        # New Case: Generate a 6-digit random case ID
        
        final_output = (
            f"New Claim ID: {new_claim_id}."
            f"New Case: A new case has been created with Case ID: {new_case_id}. "
            f"The diagnosis and claim notes indicate it's a New Case. "
            f"Diagnosis: {diagnosis}, Claim Line Note: {claim_line_note}, Service Date: {service_date.strftime('%Y-%m-%d')}."
        )
    else:
        # Follow-up Case: Add reasoning
        final_output = (
            f"New Claim ID: {new_claim_id}."
            f"Follow-up Case: The claim has been classified as a follow-up case for Case ID: {case_id}. "
            f"The diagnosis and claim notes indicate a follow-up claim, and the service date is within 30 days of the last service date."
        )

    ## LLM
    system_prompt = """Respond to the human as helpfully and accurately as possible. You are an expert in analyzing medical claims. Your task is to compare the new claim with past claims and determine if the New Claim is a "Follow-up Claim" (related to an existing issue) or a "Different Claim" (a separate, unrelated issue). 

    To make this determination, carefully analyze both the below diagnosis and the claim line notes for patterns, similarities, or differences.



    **Existing Claims:**

    - Diagnosis: "{past_claims_data_diagnosis}"

    - Claim Line Note: "{past_claims_data_claim_line_note}"



    **New Claim:**

    - Diagnosis: "{diagnosis}"

    - Claim Line Note: "{claim_line_note}"



    Use a json blob to output a confidence sore along with a reasoning.

    Valid "category" values: Follow-up Case, New Case

    Valid "confidence_score" values: 1-100

    Provide only ONE action per $JSON_BLOB, as shown:

    ```

    {{

    "action": $CATEGORY_NAME

    "confidence_score": "$CONFIDENCE_SCORE",

    "reasoning": $REASONING

    }}

    ```

    Begin! Reminder to ALWAYS respond with a valid json blob of a single action.

    Respond directly if appropriate. Format is Action:```$JSON_BLOB```"""

    system_prompt = system_prompt.format(diagnosis=diagnosis, claim_line_note=claim_line_note,
                                     past_claims_data_diagnosis=past_claims_data_llm["Diagnosis History"],
                                     past_claims_data_claim_line_note=past_claims_data_llm["Claim Line Notes"])

    # Get the LLM's response
    llm_response = ask_gpt(system_prompt)
    print(llm_response) 
    # Function to extract the JSON blob from the LLM response
    def extract_json_blobs(response):
        try:
            # Use regex to find all JSON blobs within the backticks
            matches = re.findall(r'\{.*?\}', response, re.DOTALL)
            # matches = re.findall(r'\{(?:[^{}]|(?R))*\}', response)
            json_blobs = [json.loads(match) for match in matches]
            for blob in json_blobs:
                if 'confidence_score' in blob:
                    blob['confidence_score'] = float(blob['confidence_score']) / 100
            return json_blobs
        except json.JSONDecodeError as e:
            print(f"Error in parsing JSON: {e}")
            return []

    final_output_llm = extract_json_blobs(llm_response)[0]
    print(final_output_llm)

    # Assign weights to BERT and LLM responses (adjust as per requirement)
    json_confidence_follow_up = final_output_llm['confidence_score'] if final_output_llm['action'] == 'Follow-up Case' else 1-final_output_llm['confidence_score']
    json_confidence_new_case = final_output_llm['confidence_score'] if final_output_llm['action'] == 'New Case' else 1-final_output_llm['confidence_score']  

    # Extract BERT probabilities
    bert_confidence_follow_up = bert_probabilities[0][0].item()
    bert_confidence_new_case = bert_probabilities[0][1].item()

    # (simple average)
    combined_confidence_follow_up = 0.35 * bert_confidence_follow_up + 0.65 * json_confidence_follow_up
    combined_confidence_new_case = 0.35 * bert_confidence_new_case + 0.65 * json_confidence_new_case

    final_prediction = "New Case" if combined_confidence_new_case > combined_confidence_follow_up else "Follow-Up Case"

    if final_prediction == "Follow-Up Case":
        # Construct the response with Markdown-style formatting
        return jsonify({
                #"BERT Prediction": "New Case" if pred_label == 1 else "Follow-Up Case",
                #"LLM Prediction": final_output_llm["action"],
                "Ensembled model Prediction": final_prediction,
                "New Claim ID": new_claim_id,
                "Weighted Confidence Score": round(max(combined_confidence_follow_up, combined_confidence_new_case), 2),
                "Reasoning": final_output_llm['reasoning']
        })
    else:
        return jsonify({
               # "BERT Prediction": "New Case" if pred_label == 1 else "Follow-Up Case",
               # "LLM Prediction": final_output_llm["action"],
                "Ensembled models Prediction": final_prediction,
                "New Claim ID": new_claim_id,
                "Weighted Confidence Score": round(max(combined_confidence_follow_up, combined_confidence_new_case), 2),
                "New Case ID": new_case_id,
                "Reasoning": final_output_llm['reasoning']
        })



if __name__ == '__main__':
    app.run(host='0.0.0.0', port=5000, debug=True)