File size: 8,938 Bytes
bdb1da0
 
 
 
 
 
 
 
 
 
 
 
 
 
7e9bcc1
bdb1da0
 
7e9bcc1
 
bdb1da0
 
7e9bcc1
 
 
 
 
 
 
bdb1da0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7e9bcc1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import gradio as gr
import os
import time
import pdfplumber
from dotenv import load_dotenv
import torch
from transformers import (
    BertJapaneseTokenizer,
    BertModel,
    AutoTokenizer,
    AutoModelForCausalLM,
    pipeline,
    BitsAndBytesConfig
)
from langchain_community.vectorstores import FAISS  # 修正
from langchain.chains import ConversationalRetrievalChain
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import HuggingFacePipeline  # 修正
from langchain_community.embeddings import HuggingFaceEmbeddings  # 修正
from langchain_huggingface import HuggingFaceEndpoint

# Pydanticの警告を無視
import warnings
warnings.filterwarnings(
    "ignore",
    message=r"Field \"model_name\" in HuggingFaceInferenceAPIEmbeddings has conflict with protected namespace"
)

load_dotenv()

list_llm = [
    "meta-llama/Meta-Llama-3-8B-Instruct",
    "rinna/llama-3-youko-8b",
]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]

# 日本語PDFのテキスト抽出
def extract_text_from_pdf(file_path):
    with pdfplumber.open(file_path) as pdf:
        pages = [page.extract_text() for page in pdf.pages]
    return " ".join(pages)

# モデルとトークナイザの初期化
tokenizer_bert = BertJapaneseTokenizer.from_pretrained(
    'cl-tohoku/bert-base-japanese',
    clean_up_tokenization_spaces=True
)
model_bert = BertModel.from_pretrained('cl-tohoku/bert-base-japanese')

def split_text_simple(text, chunk_size=1024):
    return [text[i:i + chunk_size] for i in range(0, len(text), chunk_size)]

def create_db(splits):
    embeddings = HuggingFaceEmbeddings(
        model_name='sonoisa/sentence-bert-base-ja-mean-tokens'
    )
    vectordb = FAISS.from_texts(splits, embeddings)
    return vectordb

def initialize_llmchain(
    llm_model,
    temperature,
    max_tokens,
    top_k,
    vector_db,
    retries=5,
    delay=5
):
    attempt = 0
    while attempt < retries:
        try:
            # ローカルモデルの場合
            if "rinna" in llm_model.lower():
                # デバイスの自動検出
                if torch.cuda.is_available():
                    device_map = "auto"
                    torch_dtype = torch.float16
                    # GPUがある場合は量子化を使用
                    quantization_config = BitsAndBytesConfig(
                        load_in_4bit=True,
                        bnb_4bit_compute_dtype=torch.float16,
                        bnb_4bit_use_double_quant=True,
                        bnb_4bit_quant_type="nf4"
                    )
                    model = AutoModelForCausalLM.from_pretrained(
                        llm_model,
                        device_map=device_map,
                        quantization_config=quantization_config
                    )
                else:
                    device_map = {"": "cpu"}
                    torch_dtype = torch.float32
                    # CPUの場合は量子化を使用せずにモデルをロード
                    model = AutoModelForCausalLM.from_pretrained(
                        llm_model,
                        device_map=device_map,
                        torch_dtype=torch_dtype
                    )
                tokenizer = AutoTokenizer.from_pretrained(llm_model, use_fast=False)
                pipe = pipeline(
                    "text-generation",
                    model=model,
                    tokenizer=tokenizer,
                    max_new_tokens=max_tokens,
                    temperature=temperature
                )
                llm = HuggingFacePipeline(pipeline=pipe)
            # エンドポイントモデルの場合
            elif "meta-llama" in llm_model.lower() or "mistralai" in llm_model.lower():
                # パラメータを直接指定
                llm = HuggingFaceEndpoint(
                    endpoint_url=f"https://api-inference.huggingface.co/models/{llm_model}",
                    huggingfacehub_api_token=os.getenv("HF_TOKEN"),
                    temperature=temperature,
                    max_new_tokens=max_tokens,
                    top_k=top_k
                )
            else:
                # その他のモデルの場合(必要に応じて追加)
                raise Exception(f"Unsupported model: {llm_model}")

            # 共通の処理
            memory = ConversationBufferMemory(
                memory_key="chat_history",
                output_key='answer',
                return_messages=True
            )
            retriever = vector_db.as_retriever()
            qa_chain = ConversationalRetrievalChain.from_llm(
                llm,
                retriever=retriever,
                memory=memory,
                return_source_documents=True,
                verbose=False
            )
            return qa_chain
        except Exception as e:
            if "Could not authenticate with huggingface_hub" in str(e):
                time.sleep(delay)
                attempt += 1
            else:
                raise Exception(f"Error initializing QA chain: {str(e)}")
    raise Exception(f"Failed to initialize after {retries} attempts")

def process_pdf(file):
    try:
        if file is None:
            return None, "Please upload a PDF file."
        text = extract_text_from_pdf(file.name)
        splits = split_text_simple(text)
        vdb = create_db(splits)
        return vdb, "PDF processed and vector database created."
    except Exception as e:
        return None, f"Error processing PDF: {str(e)}"

def initialize_qa_chain(
    llm_index,
    temperature,
    max_tokens,
    top_k,
    vector_db
):
    try:
        if vector_db is None:
            return None, "Please process a PDF first."
        llm_name = list_llm[llm_index]
        chain = initialize_llmchain(
            llm_name,
            temperature,
            max_tokens,
            top_k,
            vector_db
        )
        return chain, "QA Chatbot initialized with selected LLM."
    except Exception as e:
        return None, f"Error initializing QA chain: {str(e)}"

def update_chat(msg, history, chain):
    try:
        if chain is None:
            return history + [("User", msg), ("Assistant", "Please initialize the QA Chatbot first.")]
        response = chain({"question": msg, "chat_history": history})
        return history + [("User", msg), ("Assistant", response['answer'])]
    except Exception as e:
        return history + [("User", msg), ("Assistant", f"Error: {str(e)}")]

def demo():
    with gr.Blocks() as demo:
        vector_db = gr.State(value=None)
        qa_chain = gr.State(value=None)

        with gr.Tab("Step 1 - Upload and Process"):
            with gr.Row():
                document = gr.File(label="Upload your Japanese PDF document", file_types=["pdf"])
            with gr.Row():
                process_btn = gr.Button("Process PDF")
                process_output = gr.Textbox(label="Processing Output")

        with gr.Tab("Step 2 - Initialize QA Chatbot"):
            with gr.Row():
                llm_btn = gr.Radio(list_llm_simple, label="Select LLM Model", type="index")
                llm_temperature = gr.Slider(minimum=0.1, maximum=1.0, step=0.1, label="Temperature", value=0.7)
                max_tokens = gr.Slider(minimum=128, maximum=2048, step=128, label="Max Tokens", value=1024)
                top_k = gr.Slider(minimum=1, maximum=10, step=1, label="Top K", value=3)
            with gr.Row():
                init_qa_btn = gr.Button("Initialize QA Chatbot")
                init_output = gr.Textbox(label="Initialization Output")

        with gr.Tab("Step 3 - Chat with your Document"):
            chatbot = gr.Chatbot()
            message = gr.Textbox(label="Ask a question")
            with gr.Row():
                send_btn = gr.Button("Send")
                clear_chat_btn = gr.Button("Clear Chat")
                reset_all_btn = gr.Button("Reset All")

        process_btn.click(
            process_pdf,
            inputs=[document],
            outputs=[vector_db, process_output]
        )

        init_qa_btn.click(
            initialize_qa_chain,
            inputs=[llm_btn, llm_temperature, max_tokens, top_k, vector_db],
            outputs=[qa_chain, init_output]
        )

        send_btn.click(
            update_chat,
            inputs=[message, chatbot, qa_chain],
            outputs=[chatbot]
        )

        # Clear Chatボタン:チャット履歴のみをクリア
        clear_chat_btn.click(
            lambda: None,
            outputs=[chatbot]
        )

        # Reset Allボタン:チャット履歴、PDFデータ、チャットボットの状態をすべてクリア
        reset_all_btn.click(
            lambda: (None, None, None),
            outputs=[chatbot, vector_db, qa_chain]
        )
    return demo

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