app.py

from unsloth import FastLanguageModel
from peft import PeftModel
import pandas as pd
from unsloth.chat_templates import get_chat_template
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from sentence_transformers import SentenceTransformer, util
import nltk
import json
from google.oauth2.service_account import Credentials
import gspread
import gradio as gr

# Download stopwords
nltk.download("stopwords")
from nltk.corpus import stopwords

# Load the base model with FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Llama-3.2-3B-Instruct",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True
)
adapter_path = "FosterSystemsDatabase/model"  # Path to LoRA adapter on Hugging Face
model = PeftModel.from_pretrained(model, adapter_path)

# Load CSV data
file_path = 'Clean Missouri Data.csv'
df = pd.read_csv(file_path, encoding='MacRoman')

def search_relevant_policies(query, df, top_n=10, max_chars=40000):
    # Convert policies into a TF-IDF matrix
    tfidf = TfidfVectorizer(stop_words='english')
    tfidf_matrix = tfidf.fit_transform(df['Content'])
    query_vector = tfidf.transform([query])
    cosine_sim = cosine_similarity(query_vector, tfidf_matrix).flatten()
    
    # Get the top N relevant policies
    top_indices = cosine_sim.argsort()[-top_n:][::-1]
    relevant_policies = df.iloc[top_indices].copy()

    # Ensure total text is capped at max_chars
    char_count = 0
    valid_indices = []
    for idx, row in relevant_policies.iterrows():
        content_length = len(row["Content"])
        if char_count + content_length > max_chars:
            break
        char_count += content_length
        valid_indices.append(idx)
    
    truncated_policies = relevant_policies.loc[valid_indices]
    return truncated_policies

def get_content_after_query(response_text, query):
    query_position = response_text.lower().find(query.lower())
    if query_position != -1:
        res = response_text[query_position + len(query):].strip()
        return res[11:]
    else:
        return response_text.strip()

def process_query(query, tokenizer):
    relevant_policies = search_relevant_policies(query, df)
    formatted_policies = [row['Content'] for _, row in relevant_policies.iterrows()]
    relevant_policy_text = "\n\n".join(formatted_policies)
    
    messages_with_relevant_policies = [
        {"role": "system", "content": relevant_policy_text},
        {"role": "user", "content": query},
    ]
    
    tokenizer = get_chat_template(tokenizer, chat_template="llama-3.1")
    inputs = tokenizer.apply_chat_template(
        messages_with_relevant_policies,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to("cuda")
    
    FastLanguageModel.for_inference(model)
    outputs = model.generate(input_ids=inputs, max_new_tokens=512, use_cache=True, temperature=0.7, min_p=0.1)
    generated_response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    response = get_content_after_query(generated_response, query)
    
    # Rank the top policy using SBERT
    model_sbert = SentenceTransformer('all-MiniLM-L6-v2')
    response_embedding = model_sbert.encode(generated_response, convert_to_tensor=True)
    policy_embeddings = model_sbert.encode(relevant_policies['Content'].tolist(), convert_to_tensor=True)
    cosine_similarities = util.cos_sim(response_embedding, policy_embeddings).flatten()
    most_relevant_index = cosine_similarities.argmax().item()
    most_relevant_link = relevant_policies.iloc[most_relevant_index]['Link to Content']
    
    return {
        "response": response,
        "most_relevant_link": most_relevant_link
    }

# Set up Google Sheets
json_file_path = "fostercare-449201-85282f81c3b7.json"
with open(json_file_path, 'r') as file:
    service_account_data = json.load(file)
scopes = ["https://www.googleapis.com/auth/spreadsheets", "https://www.googleapis.com/auth/drive"]
creds = Credentials.from_service_account_info(service_account_data, scopes=scopes)
client = gspread.authorize(creds)
spreadsheet = client.open("Fostercare Responses").sheet1

# Gradio functions
def greet(query):
    result_1 = process_query(query, tokenizer)
    result_2 = process_query(query, tokenizer)
    return [result_1["response"], result_2["response"]]

def choose_preference(name, output1, output2, preference, query, broken):
    if not name:
        return "Please enter your name before submitting."
    
    broken_flag = "Yes" if broken else "No"
    
    if preference == "Output 1":
        new_row = [query, output1, output2, name, broken_flag]
        spreadsheet.append_row(new_row)
        return f"You preferred: Output 1 - {output1}"
    elif preference == "Output 2":
        new_row = [query, output2, output1, name, broken_flag]
        spreadsheet.append_row(new_row)
        return f"You preferred: Output 2 - {output2}"
    else:
        return "No preference selected."

# Gradio UI
with gr.Blocks() as demo:
    name_input = gr.Textbox(label="Enter your name")
    query_input = gr.Textbox(label="Enter your query")
    # "Generate Outputs" button is now placed right after the query input
    generate_button = gr.Button("Generate Outputs")
    output_1 = gr.Textbox(label="Output 1", interactive=False)
    output_2 = gr.Textbox(label="Output 2", interactive=False)
    preference = gr.Radio(["Output 1", "Output 2"], label="Choose your preferred output")
    broken_flag = gr.Checkbox(label="Mark as Broken Response")
    submit_button = gr.Button("Submit Preference")
    
    generate_button.click(greet, inputs=query_input, outputs=[output_1, output_2])
    submit_button.click(
        choose_preference,
        inputs=[name_input, output_1, output_2, preference, query_input, broken_flag],
    ).then(
        fn=lambda: ("", "", "", "", "", False, ""),
        inputs=[],
        outputs=[name_input, query_input, output_1, output_2, preference, broken_flag, preference_result]
    )

demo.launch()