app.py

import os, base64
import requests, json
import gradio as gr


GREEN = '\033[1;32m'
BLUE = '\033[1;34m'
RESET = '\033[0m'
URL = "https://ai1071.4dstaging.com/v1/"

VALID_ANSWER, QUERY_FAIL, INVALID_ANSWER=0 , 1, 2

VICTORIA_HARBOUR, MIC =0, 1 # VICTORIA_HARBOUR=海港飲食集團
CUSTOMER = MIC

MODES = [
    {
        "name": " ",
        "query_mode_indx": 5,
        "retrieval_temperature": 0.2, #EC19Jun2024
        "path": r"E:\workspace\RAG_data\20240412_superQuery\db\EC_test_all\20240603_haigang_qa",
        "sample_questions": [
            "這裡可以book位嗎？", "可以book位嗎？", "Hi", "蟹", "魚", "會員", "訂枱"
            "锡我?", "可唔可以幫我寫一張菜單？",
            "可以加大長腳蟹嗎？","想查詢最新堂食優惠",
            "有什麼優惠", "宴會菜單", "有長腳蟹?", "積分如何運作?", "點加入會員？",
            "套餐可轉其他菜式嗎？", "網購限定優惠可以堂食嗎？", "當日海鮮供應情況?"
        ],
     
    },{
        "name": "MiC Modular Integrated Construction - HK (Beta)",
        "query_mode_indx": 4,
        "retrieval_temperature": 0.2, #EC19Jun2024
        # "path": r"E:\workspace\RAG_data\20??????????????S",
        "path": r"E:\workspace\RAG_data\20240412_superQuery\db\EC_test_all\20240619_mic_demo",
        "sample_questions": [
            "What is MIC?", "優惠措施", "Please introduce CIC", "Key Technologies of MIC",
            "組裝合成建築法", "物料或產品規格", "MIC safety."
        ],
    }
]


questions=MODES[CUSTOMER]['sample_questions']

def the_answer(response:dict): #extract answer from the response.
    a=response['msg'].split('Answer(GPT4):')[1].split('References:')[0] #get answers 
    a.strip() # remove all linefeeds
    return a

def the_references(response:dict, user_query: str):
    ref_contents=[]
    if response["code"]==VALID_ANSWER: #       
        for ref in response["data"]["source_docs"]:
            content=ref["page_content"]
            # ref_question=content['問題']
            # ref_question_answer=content['回答']
            ref_contents.append(content)      
    ref_contents_filtered = filter_repeated(user_query, ref_contents) #EC04Jun2024
    return ref_contents_filtered #ref_contents

def filter_repeated(user_query, ref_contents: list):
# This function help to filter out the reference question that is 100% same as the user's ASKED question.
    #EC04Jun2024
    ref_contents_filtered = []
    for ref in ref_contents:
        # ref_dict = json.loads(ref)
        try:
            question = next(iter(ref.get('問題').values()))
        except StopIteration as e:
            print(e)
            pass
        except Exception as e:
            print(e)
            ref_contents_filtered.append(ref)
            continue
        print(question)
        print("question == user_query: "+str(question == user_query))
        if not question == user_query:
            ref_contents_filtered.append(ref)
    return ref_contents_filtered

def get_images_from_source(source_docs):
    image_exts = [".jpg", ".jpeg", ".png"]
    source_list = [doc['source'] for doc in source_docs]
    source_img_list = [source for source in source_list if os.path.splitext(source)[1] in image_exts]
    
    buffer_img_str = ""
    for source in source_img_list:
        response = requests.get(URL+f"images?image_id={source}")
        if response.status_code == 200:
            image_data = response.content
            base64_image = base64.b64encode(image_data).decode("utf-8") #image_data.encode("base64").decode("utf-8")
        
            # Create an HTML <img> tag
            # img_name = os.path.basename(source)
            img_html = f'<img src="data:image/png;base64,{base64_image}" alt="img_name">'
            buffer_img_str += "\n"+img_html
            
            # Print or use img_html as needed
            # print(img_html)
        else:
            print("Error fetching image")
    return buffer_img_str

def all_info(response):
    info="\n".join([f"{GREEN}{key}{RESET}: {value}" for key, value in response.items()])
    return info

def request_stream_chat(question:str, history):
    global temp_source_docs
    
    if not question:
        yield "Hello! What would you like to know?"
        return
    
    payload = {
        "prompt": question,
        "retrieval_temperature": 0.2, #MODES[CUSTOMER]['retrieval_temperature'], #EC19Jun2024: from 0.2 -> MODES[CUSTOMER]['retrieval_temperature']
        
        # "query_mode_indx": 5,
        # "path": r"E:\workspace\RAG_data\20240412_superQuery\db\EC_test_all\20240603_haigang_qa",
        
        "query_mode_indx": MODES[CUSTOMER]['query_mode_indx'],
        "path": MODES[CUSTOMER]['path'],
        
        "stream": True,
        "LLM_type": "gpt"
    }
    reply_buffer = ""
    with requests.post(url=URL+"query", json=payload, stream=True) as r_stream:
        for line in r_stream.iter_lines():
            if line:               
                line = json.loads(line)
                if line['finished']: #all the steamed content
                    response = line
                    # print(f"{RESET}-end")
                    # response=filter_repeated(response)
                    
                    msg = response['msg']
                    
                    if payload['query_mode_indx'] == 5:
                        source_docs_content = the_references(response, question)
                        source_docs_content_str = "\n".join([str(content) for content in source_docs_content])
                        response_str = msg+"\n\nSource documents:\n"+source_docs_content_str
                    else:
                        response_str = msg+"\n\n"+response.get('reference') #EC19Jun2024: from [] -> .get()
                    
                    source_docs = response['data']['source_docs']
                    image_str = get_images_from_source(source_docs)
                    response_str += "\n"+image_str
                    yield response_str
                    
                    break
                else:
                    # yield line
                    # print(f"{BLUE}"+line['reply']+f"{RESET}", end="") #steaming chuncks.
                    reply_buffer += line['reply']
                    yield reply_buffer #line['reply']
                    
    # response=the_answer(response)+'\n' + str(the_references(response))
    # return response

def my_generator(x):
    for i in range(x):
        yield i

if __name__ == "__main__":
    # responses=[]; answers=[]; references=[]; all_infos=[]
    # for q in questions:
    #     response=request_stream_chat(q, "dummy history")
        
    #     responses.append(response)  
    #     all_infos.append(all_info(response))
    #     answers.append(the_answer(response))
    #     references.append(the_references(response))
    
    gr.ChatInterface(
        request_stream_chat, #a4o_response
        examples=questions,

        chatbot=gr.Chatbot(height=450), #300),
        textbox=gr.Textbox(placeholder="喺呢度問我問題.", container=False, scale=7),
        title=MODES[CUSTOMER]['name'],
        description="智能査詢",
        theme="soft",
        cache_examples=False, #True,
        retry_btn=None,
        undo_btn="Delete Previous",
        clear_btn="Clear",
        fill_height=True,
    ).launch(share=True) #False) #True)