搜尋結果增強技術神經網路重新排序與生成式LLM

def drdecr_reranker(question, max_reranked_documents=10):
    reranker = ColBERTReranker(model=model_name_or_path)
    reranker.load()
    search_results = solr_retriever(question)
    if len(search_results) > 0:
        reranked_results = reranker.predict(queries=[question], documents=[search_results], max_num_documents=max_reranked_documents)
        print(reranked_results)
        reranked_results_to_display = [result['document'] for result in reranked_results[0]]
        df = pd.DataFrame.from_records(reranked_results_to_display, columns=['rank', 'document_id', 'text'])
        print('=========================================================')
        print(f'QUERY: {question}')
        return df['text'][0]
    else:
        return "0 documents found", "None"

output = query({
    "inputs": "根據以下上下文回答問題。\
    上下文：X700 行動電話的一些功能。透過多媒體訊息和電子郵件訊息傳送彩色圖片和影片剪輯給朋友和家人。\
    使用藍牙、紅外線或USB連線將圖片、影片、電子名片等傳送到其他手機或裝置。\
    使用附帶的miniSD卡和介面卡在適當的裝置之間交換內容。\
    使用組態好的無線網際網路瀏覽器瀏覽網際網路。\
    對VGA尺寸的影像新增框架、文字和流行藝術。\
    捕捉影片剪輯。播放40音多聲道鈴聲。不再需要有線耳機！使用藍牙無線耳機。\
    將您的聯絡人和日曆與其他網路裝置同步。在X700外部顯示屏上檢視來電者的身份。\
    下載遊戲、影像、主題、聲音、影片和Java及Series 60應用程式。在一個地方檢視所有媒體檔案，即媒體函式庫。\
    搜尋您的檔案；包括遊戲、影像、安裝包以及影片和聲音剪輯。接收、編輯和檢視Microsoft Word、Excel和PowerPoint檔案。\
    將檔案傳送到印表機和投影機（需搭配適當的配件）。使用語音命令撥打號碼。\
    說“開啟訊息”或“開啟相機”，手機將直接進入應用程式。1 \
    問題：Panasonic X700的功能是什麼？",
    "parameters": {'temperature': 0.5,
                   'max_new_tokens': 200,
                   'return_full_text': False,},
})

import os
import time
from textwrap import dedent
import dash
from dash import html
from dash import dcc
import dash_bootstrap_components as dbc
from dash.dependencies import Input, Output, State

# .env檔案內容
export HF_API_KEY="Bearer hf_XXXXXXXXX"
export OPENAI_API_KEY="sk-XXXXXXXXX"

import os
openai.api_key = os.getenv("OPENAI_API_KEY")
hf_key = os.getenv("HF_API_KEY")

def Header(name, app):
    title = html.H1(name, style={"margin-top": 5})
    logo = html.Img(
        src=app.get_asset_url("apress-logo.png"), 
        style={"float": "right", "height": 60}
    )
    return dbc.Row([dbc.Col(title, md=8), dbc.Col(logo, md=4)])

def textbox(text, box="AI", name="Knowa"):
    # 省略部分程式碼...
    if box == "user":
        style["margin-left"] = "auto"
        style["margin-right"] = 0
        return dbc.Card(text, style=style, body=True, color="primary", inverse=True)
    elif box == "AI":
        style["margin-left"] = 0
        style["margin-right"] = "auto"
        thumbnail = html.Img(
            src=app.get_asset_url("apress-logo.png"),
            style={
                "border-radius": 50,
                "height": 36,
                "margin-right": 5,
                "float": "left",
            },
        )
        textbox = dbc.Card(text, style=style, body=True, color="light", inverse=False)
        return html.Div([thumbnail, textbox])

app.layout = dbc.Container(
    fluid=False,
    children=[
        Header("Generative Q&A", app),
        # 省略部分程式碼...
        dcc.Tabs(id="tabs-example-graph", value='tab-1-example-graph',
                 children=[
                     dcc.Tab(label='AI powered QA', children=[
                         # 省略部分程式碼...
                     ]),
                 ]),
        html.Br(),
        html.Footer(children="Please note that this content is made available with the Generative AI book. \
                              The content may include systems & methods pending patent with USPTO and protected under US Patent Laws. \
                              Copyright - Apress Publication")
    ],
)

@app.callback(
    Output("display-conversation", "children"),
    [Input("store-conversation", "data")]
)
def update_display(chat_history):
    conversation_parts = chat_history.split("<split>")[:-1]
    conversation_components = [
        textbox(x, box="user") if i % 2 == 0 else textbox(x, box="AI")
        for i, x in enumerate(conversation_parts)
    ]
    return conversation_components

@app.callback(
    Output("user-input", "value"),
    [Input("submit", "n_clicks"), Input("user-input", "n_submit")],
)
def clear_input(n_clicks, n_submit):
    return "" if n_clicks or n_submit else dash.no_update

@app.callback(
    [Output("store-conversation", "data"), Output("loading-component", "children")],
    [Input("submit", "n_clicks"), Input("user-input", "n_submit")],
    [State("user-input", "value"), State("store-conversation", "data")],
)
def run_chatbot(n_clicks, n_submit, user_input, chat_history):
    if n_clicks == 0 and n_submit is None:
        return "", None
    if user_input is None or user_input == "":
        return chat_history, None
    context = drdecr_reranker(user_input)
    chat_history += f"根據以下內容回答問題。內容：{context} 問題：{user_input}"
    model_input = chat_history.replace("<split>", "\n")
    # HuggingFace LLM 模型
    model_output = call_huggingface_api(model_input)
    model_output = format_model_output(model_output)
    # OpenAI LLM 模型 - 取消註解以使用 OpenAI 模型
    # model_output = call_openai_api(model_input)
    model_output = f"{model_output}<split>"
    return model_output, None