File size: 11,400 Bytes
8e5a242
 
258d8d2
8e5a242
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a5238c2
 
a6787f7
8e5a242
1d7cc1f
 
8e5a242
 
b0ed36b
8e5a242
 
 
 
 
b0ed36b
8e5a242
258d8d2
 
 
8e5a242
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a6787f7
 
 
 
 
 
 
 
 
 
8e5a242
a6787f7
 
 
 
 
 
8e5a242
a6787f7
8e5a242
 
a6787f7
8e5a242
 
 
 
 
df6ace3
a6787f7
8e5a242
 
 
 
 
 
 
a6787f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e5a242
 
 
 
 
 
 
 
 
 
 
 
 
 
a6787f7
 
8e5a242
 
 
 
 
 
 
 
 
a6787f7
8e5a242
 
 
 
 
 
a6787f7
8e5a242
 
 
 
 
 
 
 
 
 
 
47ee788
8e5a242
 
 
a6787f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e5a242
 
a6787f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e5a242
 
 
 
 
 
 
 
 
 
 
 
 
a6787f7
8e5a242
 
 
 
a6787f7
 
8e5a242
a6787f7
 
8e5a242
 
 
a6787f7
8e5a242
 
 
 
 
a6787f7
 
 
 
 
 
8e5a242
a6787f7
 
 
 
 
 
 
 
 
8e5a242
 
 
 
a6787f7
8e5a242
 
 
 
 
a6787f7
 
 
 
 
 
8e5a242
a6787f7
 
 
 
 
 
 
 
 
8e5a242
 
 
a6787f7
 
8e5a242
 
 
 
 
 
 
a6787f7
 
8e5a242
 
 
60e0df0
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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
import gradio as gr

from uuid import uuid4
from huggingface_hub import snapshot_download
from langchain.document_loaders import (
    CSVLoader,
    EverNoteLoader,
    PDFMinerLoader,
    TextLoader,
    UnstructuredEmailLoader,
    UnstructuredEPubLoader,
    UnstructuredHTMLLoader,
    UnstructuredMarkdownLoader,
    UnstructuredODTLoader,
    UnstructuredPowerPointLoader,
    UnstructuredWordDocumentLoader,
)
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.docstore.document import Document
from chromadb.config import Settings
from llama_cpp import Llama


SYSTEM_PROMPT = "Ты — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им."
SYSTEM_TOKEN = 1788
USER_TOKEN = 1404
BOT_TOKEN = 9225
LINEBREAK_TOKEN = 13

ROLE_TOKENS = {
    "user": USER_TOKEN,
    "bot": BOT_TOKEN,
    "system": SYSTEM_TOKEN
}

LOADER_MAPPING = {
    ".csv": (CSVLoader, {}),
    ".doc": (UnstructuredWordDocumentLoader, {}),
    ".docx": (UnstructuredWordDocumentLoader, {}),
    ".enex": (EverNoteLoader, {}),
    ".epub": (UnstructuredEPubLoader, {}),
    ".html": (UnstructuredHTMLLoader, {}),
    ".md": (UnstructuredMarkdownLoader, {}),
    ".odt": (UnstructuredODTLoader, {}),
    ".pdf": (PDFMinerLoader, {}),
    ".ppt": (UnstructuredPowerPointLoader, {}),
    ".pptx": (UnstructuredPowerPointLoader, {}),
    ".txt": (TextLoader, {"encoding": "utf8"}),
}


repo_name = "IlyaGusev/saiga_13b_lora_llamacpp"
model_name = "ggml-model-q4_1.bin"
embedder_name = "sentence-transformers/paraphrase-multilingual-mpnet-base-v2"

snapshot_download(repo_id=repo_name, local_dir="/data/saiga_13b_lora_llamacpp", allow_patterns=model_name, cache_dir="/data/cache")
snapshot_download(repo_id=embedder_name, local_dir="/data/paraphrase-multilingual-mpnet-base-v2", cache_dir="/data/cache")

model = Llama(
    model_path=f"/data/saiga_13b_lora_llamacpp/{model_name}",
    n_ctx=2000,
    n_parts=1,
)

max_new_tokens = 1500
embeddings = HuggingFaceEmbeddings(model_name="/data/paraphrase-multilingual-mpnet-base-v2")

def get_uuid():
    return str(uuid4())


def load_single_document(file_path: str) -> Document:
    ext = "." + file_path.rsplit(".", 1)[-1]
    assert ext in LOADER_MAPPING
    loader_class, loader_args = LOADER_MAPPING[ext]
    loader = loader_class(file_path, **loader_args)
    return loader.load()[0]


def get_message_tokens(model, role, content):
    message_tokens = model.tokenize(content.encode("utf-8"))
    message_tokens.insert(1, ROLE_TOKENS[role])
    message_tokens.insert(2, LINEBREAK_TOKEN)
    message_tokens.append(model.token_eos())
    return message_tokens


def get_system_tokens(model):
    system_message = {"role": "system", "content": SYSTEM_PROMPT}
    return get_message_tokens(model, **system_message)


def upload_files(files, file_paths):
    file_paths = [f.name for f in files]
    return file_paths


def process_text(text):
    lines = text.split("\n")
    lines = [line for line in lines if len(line.strip()) > 2]
    text = "\n".join(lines).strip()
    if len(text) < 10:
        return None
    return text


def build_index(file_paths, db, chunk_size, chunk_overlap, file_warning):
    documents = [load_single_document(path) for path in file_paths]
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
    documents = text_splitter.split_documents(documents)
    fixed_documents = []
    for doc in documents:
        doc.page_content = process_text(doc.page_content)
        if not doc.page_content:
            continue
        fixed_documents.append(doc)

    db = Chroma.from_documents(
        fixed_documents,
        embeddings,
        client_settings=Settings(
            anonymized_telemetry=False
        )
    )
    file_warning = f"Загружено {len(fixed_documents)} фрагментов! Можно задавать вопросы."
    return db, file_warning


def user(message, history, system_prompt):
    new_history = history + [[message, None]]
    return "", new_history


def retrieve(history, db, retrieved_docs, k_documents):
    context = ""
    if db:
        last_user_message = history[-1][0]
        retriever = db.as_retriever(search_kwargs={"k": k_documents})
        docs = retriever.get_relevant_documents(last_user_message)
        retrieved_docs = "\n\n".join([doc.page_content for doc in docs])
    return retrieved_docs


def bot(
    history,
    system_prompt,
    conversation_id,
    retrieved_docs,
    top_p,
    top_k,
    temp
):
    if not history:
        return

    tokens = get_system_tokens(model)[:]
    tokens.append(LINEBREAK_TOKEN)

    for user_message, bot_message in history[:-1]:
        message_tokens = get_message_tokens(model=model, role="user", content=user_message)
        tokens.extend(message_tokens)
        if bot_message:
            message_tokens = get_message_tokens(model=model, role="bot", content=bot_message)
            tokens.extend(message_tokens)

    last_user_message = history[-1][0]
    if retrieved_docs:
        last_user_message = f"Контекст: {retrieved_docs}\n\nИспользуя контекст, ответь на вопрос: {last_user_message}"
    message_tokens = get_message_tokens(model=model, role="user", content=last_user_message)
    tokens.extend(message_tokens)

    role_tokens = [model.token_bos(), BOT_TOKEN, LINEBREAK_TOKEN]
    tokens.extend(role_tokens)
    generator = model.generate(
        tokens,
        top_k=top_k,
        top_p=top_p,
        temp=temp
    )

    partial_text = ""
    for i, token in enumerate(generator):
        if token == model.token_eos() or (max_new_tokens is not None and i >= max_new_tokens):
            break
        partial_text += model.detokenize([token]).decode("utf-8", "ignore")
        history[-1][1] = partial_text
        yield history


with gr.Blocks(
    theme=gr.themes.Soft()
) as demo:
    db = gr.State(None)
    conversation_id = gr.State(get_uuid)
    favicon = '<img src="https://cdn.midjourney.com/b88e5beb-6324-4820-8504-a1a37a9ba36d/0_1.png" width="48px" style="display: inline">'
    gr.Markdown(
        f"""<h1><center>{favicon}Saiga 13B llama.cpp: retrieval QA</center></h1>
        """
    )

    with gr.Row():
        with gr.Column(scale=5):
            file_output = gr.File(file_count="multiple", label="Загрузка файлов")
            file_paths = gr.State([])
            file_warning = gr.Markdown(f"Фрагменты ещё не загружены!")

        with gr.Column(min_width=200, scale=3):
            with gr.Tab(label="Параметры нарезки"):
                chunk_size = gr.Slider(
                    minimum=50,
                    maximum=2000,
                    value=250,
                    step=50,
                    interactive=True,
                    label="Размер фрагментов",
                )
                chunk_overlap = gr.Slider(
                    minimum=0,
                    maximum=500,
                    value=30,
                    step=10,
                    interactive=True,
                    label="Пересечение"
                )


    with gr.Row():
        k_documents = gr.Slider(
            minimum=1,
            maximum=10,
            value=2,
            step=1,
            interactive=True,
            label="Кол-во фрагментов для контекста"
        )
    with gr.Row():
        retrieved_docs = gr.Textbox(
            lines=6,
            label="Извлеченные фрагменты",
            placeholder="Появятся после задавания вопросов",
            interactive=False
        )
    with gr.Row():
        with gr.Column(scale=5):
            system_prompt = gr.Textbox(label="Системный промпт", placeholder="", value=SYSTEM_PROMPT, interactive=False)
            chatbot = gr.Chatbot(label="Диалог").style(height=400)
        with gr.Column(min_width=80, scale=1):
            with gr.Tab(label="Параметры генерации"):
                top_p = gr.Slider(
                    minimum=0.0,
                    maximum=1.0,
                    value=0.9,
                    step=0.05,
                    interactive=True,
                    label="Top-p",
                )
                top_k = gr.Slider(
                    minimum=10,
                    maximum=100,
                    value=30,
                    step=5,
                    interactive=True,
                    label="Top-k",
                )
                temp = gr.Slider(
                    minimum=0.0,
                    maximum=2.0,
                    value=0.1,
                    step=0.1,
                    interactive=True,
                    label="Temp"
                )

    with gr.Row():
        with gr.Column():
            msg = gr.Textbox(
                label="Отправить сообщение",
                placeholder="Отправить сообщение",
                show_label=False,
            ).style(container=False)
        with gr.Column():
            with gr.Row():
                submit = gr.Button("Отправить")
                stop = gr.Button("Остановить")
                clear = gr.Button("Очистить")

    # Upload files
    upload_event = file_output.change(
        fn=upload_files,
        inputs=[file_output, file_paths],
        outputs=[file_paths],
        queue=True,
    ).success(
        fn=build_index,
        inputs=[file_paths, db, chunk_size, chunk_overlap, file_warning],
        outputs=[db, file_warning],
        queue=True
    )

    # Pressing Enter
    submit_event = msg.submit(
        fn=user,
        inputs=[msg, chatbot, system_prompt],
        outputs=[msg, chatbot],
        queue=False,
    ).success(
        fn=retrieve,
        inputs=[chatbot, db, retrieved_docs, k_documents],
        outputs=[retrieved_docs],
        queue=True,
    ).success(
        fn=bot,
        inputs=[
            chatbot,
            system_prompt,
            conversation_id,
            retrieved_docs,
            top_p,
            top_k,
            temp
        ],
        outputs=chatbot,
        queue=True,
    )

    # Pressing the button
    submit_click_event = submit.click(
        fn=user,
        inputs=[msg, chatbot, system_prompt],
        outputs=[msg, chatbot],
        queue=False,
    ).success(
        fn=retrieve,
        inputs=[chatbot, db, retrieved_docs, k_documents],
        outputs=[retrieved_docs],
        queue=True,
    ).success(
        fn=bot,
        inputs=[
            chatbot,
            system_prompt,
            conversation_id,
            retrieved_docs,
            top_p,
            top_k,
            temp
        ],
        outputs=chatbot,
        queue=True,
    )

    # Stop generation
    stop.click(
        fn=None,
        inputs=None,
        outputs=None,
        cancels=[submit_event, submit_click_event],
        queue=False,
    )

    # Clear history
    clear.click(lambda: None, None, chatbot, queue=False)

demo.queue(max_size=128, concurrency_count=1)
demo.launch()