File size: 16,467 Bytes
3ec9224
5be8df6
3ec9224
5db4902
5be8df6
5db4902
5be8df6
5db4902
 
5be8df6
 
ecf1633
3ec9224
1ef8d7c
 
aa98840
1ef8d7c
5be8df6
 
 
 
 
 
 
6803339
146ca67
f74a5f5
ce888d8
89262a4
 
3bf3580
 
37ae113
b1ec9ac
5be8df6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ef8d7c
5be8df6
1ef8d7c
5be8df6
 
 
1ef8d7c
 
6f79ea2
5be8df6
 
 
 
 
 
 
 
88fa380
5be8df6
 
 
 
 
 
 
 
067316d
5be8df6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
067316d
 
 
88fa380
ecf1633
88fa380
ecf1633
 
 
 
 
88fa380
4ce7fc5
 
 
 
 
 
 
 
88fa380
fa4eb33
ecf1633
067316d
ecf1633
 
 
 
 
 
e27125d
 
ecf1633
e27125d
ecf1633
 
 
 
067316d
fa4eb33
 
ecf1633
fa4eb33
ecf1633
 
 
 
fa4eb33
067316d
ecf1633
067316d
 
ecf1633
 
 
 
067316d
5be8df6
eb94a8f
5be8df6
 
9733941
5be8df6
 
 
 
 
 
 
 
 
 
 
9733941
138ca2e
 
5be8df6
 
00bd139
5be8df6
 
00bd139
5be8df6
 
 
32a58be
 
1ef8d7c
32a58be
6803339
32a58be
aa98840
 
6803339
989bff5
6803339
 
 
 
 
1ef8d7c
32a58be
5be8df6
 
 
32a58be
5be8df6
 
1ef8d7c
5be8df6
1ef8d7c
5be8df6
 
 
09444af
5be8df6
e0b45d0
00bd139
 
5be8df6
 
 
 
 
 
 
 
 
 
00bd139
5be8df6
 
 
 
 
9733941
04361a6
 
9733941
 
 
8bef1bd
9733941
 
 
8bef1bd
9733941
 
5be8df6
 
9733941
 
8bef1bd
5be8df6
 
 
 
 
 
 
 
 
 
 
 
 
 
3ca2785
00bd139
1ef8d7c
5be8df6
 
1da1e92
eb94a8f
 
9733941
ceae871
5be8df6
 
 
da842e1
5be8df6
 
a25f0eb
 
5be8df6
9733941
5be8df6
4f52c44
5be8df6
a25f0eb
5be8df6
1e93bea
5be8df6
8aeeae1
5be8df6
b1ec9ac
a25f0eb
 
e4c8a25
 
 
 
 
 
5be8df6
8aeeae1
5be8df6
1e93bea
5be8df6
8aeeae1
14155e5
9733941
 
 
 
 
 
 
8bef1bd
 
 
5be8df6
 
 
 
 
 
 
 
9733941
 
1ef8d7c
9733941
 
3b27ad5
9733941
8bef1bd
9733941
5be8df6
 
9733941
00bd139
8bef1bd
9733941
 
00bd139
8bef1bd
9733941
f8033f1
9733941
8bef1bd
9733941
323ccbe
5be8df6
 
 
 
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
import gradio as gr
import os

from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceEmbeddings 
from langchain_community.llms import HuggingFacePipeline
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import HuggingFaceEndpoint

from pathlib import Path
import chromadb
from unidecode import unidecode

from transformers import AutoTokenizer
import transformers
import torch
import tqdm 
import accelerate



# default_persist_directory = './chroma_HF/'
list_llm = ["mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.1", \
    "google/gemma-7b-it","google/gemma-2b-it", \
    "HuggingFaceH4/zephyr-7b-beta", "HuggingFaceH4/zephyr-7b-gemma-v0.1", \
    "meta-llama/Llama-2-7b-chat-hf", "microsoft/phi-2", \
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "mosaicml/mpt-7b-instruct", "tiiuae/falcon-7b-instruct", \
    "google/flan-t5-xxl"
]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]

# Load PDF document and create doc splits
def load_doc(list_file_path, chunk_size, chunk_overlap):
    # Processing for one document only
    # loader = PyPDFLoader(file_path)
    # pages = loader.load()
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    # text_splitter = RecursiveCharacterTextSplitter(chunk_size = 600, chunk_overlap = 50)
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size = chunk_size, 
        chunk_overlap = chunk_overlap)
    doc_splits = text_splitter.split_documents(pages)
    return doc_splits


# Create vector database
def create_db(splits, collection_name):
    embedding = HuggingFaceEmbeddings()
    new_client = chromadb.EphemeralClient()
    vectordb = Chroma.from_documents(
        documents=splits,
        embedding=embedding,
        client=new_client,
        collection_name=collection_name,
        # persist_directory=default_persist_directory
    )
    return vectordb


# Load vector database
def load_db():
    embedding = HuggingFaceEmbeddings()
    vectordb = Chroma(
        # persist_directory=default_persist_directory, 
        embedding_function=embedding)
    return vectordb


# Initialize langchain LLM chain
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    progress(0.1, desc="Initializing HF tokenizer...")
    # HuggingFacePipeline uses local model
    # Note: it will download model locally...
    # tokenizer=AutoTokenizer.from_pretrained(llm_model)
    # progress(0.5, desc="Initializing HF pipeline...")
    # pipeline=transformers.pipeline(
    #     "text-generation",
    #     model=llm_model,
    #     tokenizer=tokenizer,
    #     torch_dtype=torch.bfloat16,
    #     trust_remote_code=True,
    #     device_map="auto",
    #     # max_length=1024,
    #     max_new_tokens=max_tokens,
    #     do_sample=True,
    #     top_k=top_k,
    #     num_return_sequences=1,
    #     eos_token_id=tokenizer.eos_token_id
    #     )
    # llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
    
    # HuggingFaceHub uses HF inference endpoints
    progress(0.5, desc="Initializing HF Hub...")
    # Use of trust_remote_code as model_kwargs
    # Warning: langchain issue
    # URL: https://github.com/langchain-ai/langchain/issues/6080
    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
            temperature = temperature,
            max_new_tokens = max_tokens,
            top_k = top_k,
            load_in_8bit = True,
        )
    elif llm_model == "HuggingFaceH4/zephyr-7b-gemma-v0.1":
        raise gr.Error("zephyr-7b-gemma-v0.1 is too large to be loaded automatically on free inference endpoint")
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            temperature = temperature,
            max_new_tokens = max_tokens,
            top_k = top_k,
        )
    elif llm_model == "microsoft/phi-2":
        raise gr.Error("phi-2 model requires 'trust_remote_code=True', currently not supported by langchain HuggingFaceHub...")
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
            temperature = temperature,
            max_new_tokens = max_tokens,
            top_k = top_k,
            trust_remote_code = True,
            torch_dtype = "auto",
        )
    elif llm_model == "TinyLlama/TinyLlama-1.1B-Chat-v1.0":
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            # model_kwargs={"temperature": temperature, "max_new_tokens": 250, "top_k": top_k}
            temperature = temperature,
            max_new_tokens = 250,
            top_k = top_k,
        )
    elif llm_model == "meta-llama/Llama-2-7b-chat-hf":
        raise gr.Error("Llama-2-7b-chat-hf model requires a Pro subscription...")
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
            temperature = temperature,
            max_new_tokens = max_tokens,
            top_k = top_k,
        )
    else:
        llm = HuggingFaceEndpoint(
            repo_id=llm_model, 
            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
            temperature = temperature,
            max_new_tokens = max_tokens,
            top_k = top_k,
        )
    
    progress(0.75, desc="Defining buffer memory...")
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )
    # retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
    retriever=vector_db.as_retriever()
    progress(0.8, desc="Defining retrieval chain...")
    qa_chain = ConversationalRetrievalChain.from_llm(
        llm,
        retriever=retriever,
        chain_type="stuff", 
        memory=memory,
        # combine_docs_chain_kwargs={"prompt": your_prompt})
        return_source_documents=True,
        #return_generated_question=False,
        verbose=False,
    )
    progress(0.9, desc="Done!")
    return qa_chain


# Initialize database
def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
    # Create list of documents (when valid)
    list_file_path = [x.name for x in list_file_obj if x is not None]
    # Create collection_name for vector database
    progress(0.1, desc="Creating collection name...")
    collection_name = Path(list_file_path[0]).stem
    # Fix potential issues from naming convention
    ## Remove space
    collection_name = collection_name.replace(" ","-") 
    ## ASCII transliterations of Unicode text
    collection_name = unidecode(collection_name)
    ## Limit lenght to 50 characters
    collection_name = collection_name[:50]
    ## Enforce start and end as alphanumeric character
    if not collection_name[0].isalnum():
        collection_name[0] = 'A'
    if not collection_name[-1].isalnum():
        collection_name[-1] = 'Z'
    # print('list_file_path: ', list_file_path)
    print('Collection name: ', collection_name)
    progress(0.25, desc="Loading document...")
    # Load document and create splits
    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
    # Create or load vector database
    progress(0.5, desc="Generating vector database...")
    # global vector_db
    vector_db = create_db(doc_splits, collection_name)
    progress(0.9, desc="Done!")
    return vector_db, collection_name, "Complete!"


def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    # print("llm_option",llm_option)
    llm_name = list_llm[llm_option]
    print("llm_name: ",llm_name)
    qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
    return qa_chain, "Complete!"


def format_chat_history(message, chat_history):
    formatted_chat_history = []
    for user_message, bot_message in chat_history:
        formatted_chat_history.append(f"User: {user_message}")
        formatted_chat_history.append(f"Assistant: {bot_message}")
    return formatted_chat_history
    

def conversation(qa_chain, message, history):
    formatted_chat_history = format_chat_history(message, history)
    #print("formatted_chat_history",formatted_chat_history)
   
    # Generate response using QA chain
    response = qa_chain({"question": message, "chat_history": formatted_chat_history})
    response_answer = response["answer"]
    if response_answer.find("Helpful Answer:") != -1:
        response_answer = response_answer.split("Helpful Answer:")[-1]
    response_sources = response["source_documents"]
    response_source1 = response_sources[0].page_content.strip()
    response_source2 = response_sources[1].page_content.strip()
    response_source3 = response_sources[2].page_content.strip()
    # Langchain sources are zero-based
    response_source1_page = response_sources[0].metadata["page"] + 1
    response_source2_page = response_sources[1].metadata["page"] + 1
    response_source3_page = response_sources[2].metadata["page"] + 1
    # print ('chat response: ', response_answer)
    # print('DB source', response_sources)
    
    # Append user message and response to chat history
    new_history = history + [(message, response_answer)]
    # return gr.update(value=""), new_history, response_sources[0], response_sources[1] 
    return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
    

def upload_file(file_obj):
    list_file_path = []
    for idx, file in enumerate(file_obj):
        file_path = file_obj.name
        list_file_path.append(file_path)
    # print(file_path)
    # initialize_database(file_path, progress)
    return list_file_path


def demo():
    with gr.Blocks(theme="base") as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        collection_name = gr.State()
        
        gr.Markdown(
        """<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>
        <h3>Ask any questions about your PDF documents, along with follow-ups</h3>
        <b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
        When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.</i>
        <br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate an output.<br>
        """)
        with gr.Tab("Step 1 - Document pre-processing"):
            with gr.Row():
                document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
                # upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
            with gr.Row():
                db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
            with gr.Accordion("Advanced options - Document text splitter", open=False):
                with gr.Row():
                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
                with gr.Row():
                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
            with gr.Row():
                db_progress = gr.Textbox(label="Vector database initialization", value="None")
            with gr.Row():
                db_btn = gr.Button("Generate vector database...")
            
        with gr.Tab("Step 2 - QA chain initialization"):
            with gr.Row():
                llm_btn = gr.Radio(list_llm_simple, \
                    label="LLM models", value = list_llm_simple[0], type="index", info="Choose your LLM model")
            with gr.Accordion("Advanced options - LLM model", open=False):
                with gr.Row():
                    slider_temperature = gr.Slider(minimum = 0.0, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
                with gr.Row():
                    slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
                with gr.Row():
                    slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
            with gr.Row():
                llm_progress = gr.Textbox(value="None",label="QA chain initialization")
            with gr.Row():
                qachain_btn = gr.Button("Initialize question-answering chain...")

        with gr.Tab("Step 3 - Conversation with chatbot"):
            chatbot = gr.Chatbot(height=300)
            with gr.Accordion("Advanced - Document references", open=False):
                with gr.Row():
                    doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
                    source1_page = gr.Number(label="Page", scale=1)
                with gr.Row():
                    doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
                    source2_page = gr.Number(label="Page", scale=1)
                with gr.Row():
                    doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
                    source3_page = gr.Number(label="Page", scale=1)
            with gr.Row():
                msg = gr.Textbox(placeholder="Type message", container=True)
            with gr.Row():
                submit_btn = gr.Button("Submit")
                clear_btn = gr.ClearButton([msg, chatbot])
            
        # Preprocessing events
        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
        db_btn.click(initialize_database, \
            inputs=[document, slider_chunk_size, slider_chunk_overlap], \
            outputs=[vector_db, collection_name, db_progress])
        qachain_btn.click(initialize_LLM, \
            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
            outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
            inputs=None, \
            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
            queue=False)

        # Chatbot events
        msg.submit(conversation, \
            inputs=[qa_chain, msg, chatbot], \
            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
            queue=False)
        submit_btn.click(conversation, \
            inputs=[qa_chain, msg, chatbot], \
            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
            queue=False)
        clear_btn.click(lambda:[None,"",0,"",0,"",0], \
            inputs=None, \
            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
            queue=False)
    demo.queue().launch(debug=True)


if __name__ == "__main__":
    demo()