interface.py

import pandas as pd
import json
import re
from json import loads, JSONDecodeError
import sys
import os
import ast
from util.vector_base import EmbeddingFunction, get_or_create_vector_base
from doubao_service import DouBaoService
from PROMPT_TEMPLATE import prompt_template
from util.Embeddings import TextEmb3LargeEmbedding
from langchain_core.documents import Document
from FlagEmbedding import FlagReranker
from retriever import retriever
import time
# from bm25s import BM25, tokenize
import contextlib
import io

import gradio as gr
import time

client = DouBaoService("DouBao128Pro")
embeddingmodel = TextEmb3LargeEmbedding(max_qpm=58)
embedding = EmbeddingFunction(embeddingmodel)
safeguard_vector_store = get_or_create_vector_base('safeguard_database', embedding)

# reranker_model = FlagReranker(
#     'C://Users//Admin//Desktop//PDPO//NLL_LLM//model//bge-reranker-v2-m3',
#     use_fp16=True,
#     devices=["cpu"],   
# )

OPTIONS = ['AI Governance',
 'Data Accuracy',
 'Data Minimization & Purpose Limitation',
 'Data Retention',
 'Data Security',
 'Data Sharing',
 'Individual Rights',
 'Privacy by Design',
 'Transparency']


def format_model_output(raw_output):
    """
    处理模型输出：
    - 将 \n 转换为实际换行
    - 提取 ```json ``` 中的内容并格式化为可折叠的 JSON
    """
    formatted = raw_output.replace('\\n', '\n')  
    def replace_json(match):
        json_str = match.group(1).strip()
        try:
            json_obj = loads(json_str)
            return f"```json\n{json.dumps(json_obj, indent=2, ensure_ascii=False)}\n```"
        except JSONDecodeError:
            return match.group(0) 
  
    formatted = re.sub(r'```json\n?(.*?)\n?```', replace_json, formatted, flags=re.DOTALL)
    return ast.literal_eval(formatted) 

def model_predict(input_text, if_split_po, topk, selected_items):
    """
    selected_items: 用户选择的项目（可能是["All"]或具体PO）
    """
    requirement = input_text
    requirement = requirement.replace("\t", "").replace("\n", "").replace("\r", "")
    if "All" in selected_items:
        PO = OPTIONS
    else:
        PO = selected_items
    if topk:
        topk = int(topk)
    else:
        topk = 10
    final_result = retriever(
                requirement,
                PO,
                safeguard_vector_store,
                reranker_model=None,
                using_reranker=False,
                using_BM25=False,
                using_chroma=True,
                k=topk,
                if_split_po=if_split_po
            )
    mapping_safeguards = {}
    for safeguard in final_result:
        if safeguard[3] not in mapping_safeguards:
            mapping_safeguards[safeguard[3]] = []
        mapping_safeguards[safeguard[3]].append(
            {
                "Score": safeguard[0],
                "Safeguard Number": safeguard[1],
                "Safeguard Description": safeguard[2]
            }
        )
    prompt = prompt_template(requirement, mapping_safeguards)
    response = client.chat_complete(messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": prompt},
    ])
    # return {"requirement": requirement, "safeguards": mapping_safeguards}
    print("requirement:", requirement)
    print("mapping safeguards:", mapping_safeguards)
    print("response:", response)
    return {"requirement": requirement, "safeguards": format_model_output(response)}

with gr.Blocks(title="New Law Landing") as demo:
    gr.Markdown("## 🏙️ New Law Landing")
    
    requirement = gr.Textbox(label="Input Requirements", placeholder="Example: Data Minimization Consent for incompatible purposes")
    details = gr.Textbox(label="Input Details", placeholder="Example: Require consent for...")
    
    # 修改为 Number 输入组件
    topk = gr.Number(
        label="Top K safeguards", 
        value=10,  
        precision=0,  
        minimum=1,  
        interactive=True
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            if_split_po = gr.Checkbox(
                label="If Split Privacy Objective", 
                value=True, 
                info="Recall K Safeguards for each Privacy Objective"
            )
        with gr.Column(scale=1):
            all_checkbox = gr.Checkbox(
                label="ALL Privacy Objective", 
                value=True, 
                info="No specific Privacy Objective is specified"
            )
        with gr.Column(scale=4):
            PO_checklist = gr.CheckboxGroup(
                label="Choose Privacy Objective",
                choices=OPTIONS,
                value=[],  
                interactive=True
            )

    submit_btn = gr.Button("Submit", variant="primary")
    result_output = gr.JSON(label="Related safeguards", open=True)
    

    def sync_checkboxes(selected_items, all_selected):
        if len(selected_items) > 0:
            return False
        return all_selected
    
    PO_checklist.change(
        fn=sync_checkboxes,
        inputs=[PO_checklist, all_checkbox],
        outputs=all_checkbox
    )
    
    def sync_all(selected_all, current_selection):
        if selected_all:
            return []
        return current_selection
    
    all_checkbox.change(
        fn=sync_all,
        inputs=[all_checkbox, PO_checklist],
        outputs=PO_checklist
    )
 
    def process_inputs(requirement, details, topk, if_split_po, all_selected, PO_selected):
        input_text = requirement + ": " + details
        if all_selected:
            return model_predict(input_text, if_split_po, int(topk), ["All"])
        else:
            return model_predict(input_text, if_split_po, int(topk), PO_selected)
    
    submit_btn.click(
        fn=process_inputs,
        inputs=[requirement, details, topk, if_split_po, all_checkbox, PO_checklist],
        outputs=[result_output]
    )

if __name__ == "__main__":
    demo.launch(share=True)