File size: 15,000 Bytes
49e32ea
 
 
 
1365c48
 
 
9795699
e78df68
49e32ea
 
 
2e536f9
 
1365c48
49e32ea
2e536f9
1365c48
49e32ea
 
 
 
 
 
 
 
 
2e536f9
 
49e32ea
275393f
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114048b
 
2e536f9
114048b
 
 
 
 
 
 
2e536f9
275393f
114048b
 
d5a8385
114048b
 
 
 
 
1365c48
2e536f9
d213c15
 
 
d5a8385
2e536f9
d213c15
 
 
d5a8385
 
2e536f9
 
 
275393f
2e536f9
 
 
 
 
 
 
 
 
1365c48
2e536f9
1365c48
2e536f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d5a8385
 
 
275393f
2e536f9
1365c48
275393f
1365c48
114048b
1365c48
275393f
114048b
2e536f9
49e32ea
 
 
 
41ed1b7
 
49e32ea
 
 
bc459f6
49e32ea
 
 
102df35
49e32ea
 
 
71c040a
49e32ea
 
102df35
 
 
 
2e536f9
102df35
 
 
2e536f9
 
 
102df35
 
 
49e32ea
 
275393f
9118536
275393f
 
 
49e32ea
 
 
 
2e536f9
d5a8385
275393f
 
49e32ea
 
 
 
 
 
30689f9
 
d5a8385
 
49e32ea
71c040a
 
 
49e32ea
 
71c040a
49e32ea
30689f9
 
 
49e32ea
 
 
 
 
 
 
ae4a7ec
49e32ea
 
 
ae4a7ec
49e32ea
 
 
1365c48
275393f
d5a8385
d213c15
d5a8385
 
2e536f9
49e32ea
2e536f9
49e32ea
 
71c040a
49e32ea
d5a8385
275393f
d5a8385
 
 
2e536f9
49e32ea
9118536
49e32ea
71c040a
 
49e32ea
 
9118536
49e32ea
71c040a
 
49e32ea
 
 
 
2e536f9
49e32ea
2e536f9
49e32ea
 
2e536f9
49e32ea
2e536f9
49e32ea
2e536f9
49e32ea
 
2e536f9
49e32ea
d5a8385
 
 
49e32ea
 
 
 
2e536f9
aa0ad5d
49e32ea
 
 
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
# # Load in packages

# +
import os

# Need to overwrite version of gradio present in Huggingface spaces as it doesn't have like buttons/avatars (Oct 2023)
#os.system("pip uninstall -y gradio")
os.system("pip install gradio==3.42.0")

from typing import TypeVar
from langchain.embeddings import HuggingFaceEmbeddings, HuggingFaceInstructEmbeddings
from langchain.vectorstores import FAISS
import gradio as gr

from transformers import AutoTokenizer

# Alternative model sources
from ctransformers import AutoModelForCausalLM

PandasDataFrame = TypeVar('pd.core.frame.DataFrame')

# Disable cuda devices if necessary
#os.environ['CUDA_VISIBLE_DEVICES'] = '-1' 

#from chatfuncs.chatfuncs import *
import chatfuncs.ingest as ing


##  Load preset embeddings, vectorstore, and model

embeddings_name = "BAAI/bge-base-en-v1.5"

def load_embeddings(embeddings_name = "thenlper/gte-base"):


    if embeddings_name == "hkunlp/instructor-large":
        embeddings_func = HuggingFaceInstructEmbeddings(model_name=embeddings_name,
        embed_instruction="Represent the paragraph for retrieval: ",
        query_instruction="Represent the question for retrieving supporting documents: "
        )

    else: 
        embeddings_func = HuggingFaceEmbeddings(model_name=embeddings_name)

    global embeddings

    embeddings = embeddings_func

    return embeddings

def get_faiss_store(faiss_vstore_folder,embeddings):
    import zipfile
    with zipfile.ZipFile(faiss_vstore_folder + '/' + faiss_vstore_folder + '.zip', 'r') as zip_ref:
        zip_ref.extractall(faiss_vstore_folder)

    faiss_vstore = FAISS.load_local(folder_path=faiss_vstore_folder, embeddings=embeddings)
    os.remove(faiss_vstore_folder + "/index.faiss")
    os.remove(faiss_vstore_folder + "/index.pkl")
    
    global vectorstore

    vectorstore = faiss_vstore

    return vectorstore

import chatfuncs.chatfuncs as chatf

chatf.embeddings = load_embeddings(embeddings_name)
chatf.vectorstore = get_faiss_store(faiss_vstore_folder="faiss_embedding",embeddings=globals()["embeddings"])

def load_model(model_type, gpu_layers, gpu_config=None, cpu_config=None, torch_device=None):
    print("Loading model")

    # Default values inside the function
    if gpu_config is None:
        gpu_config = chatf.gpu_config
    if cpu_config is None:
        cpu_config = chatf.cpu_config
    if torch_device is None:
        torch_device = chatf.torch_device

    if model_type == "Orca Mini (larger, slow)":

        gpu_config.update_gpu(gpu_layers)
        cpu_config.update_gpu(gpu_layers)

        print("Loading with", cpu_config.gpu_layers, "model layers sent to GPU.")

        print(vars(gpu_config))
        print(vars(cpu_config))

        try:
            #model = AutoModelForCausalLM.from_pretrained('juanjgit/orca_mini_3B-GGUF', model_type='llama', model_file='orca-mini-3b.q4_0.gguf', **vars(gpu_config)) # **asdict(CtransRunConfig_cpu())
            model = AutoModelForCausalLM.from_pretrained('Aryanne/Orca-Mini-3B-gguf', model_type='llama', model_file='q5_0-orca-mini-3b.gguf', **vars(gpu_config)) # **asdict(CtransRunConfig_cpu())
            #model = AutoModelForCausalLM.from_pretrained('Aryanne/Wizard-Orca-3B-gguf', model_type='llama', model_file='q4_1-wizard-orca-3b.gguf', **vars(gpu_config)) # **asdict(CtransRunConfig_cpu())
            #model = AutoModelForCausalLM.from_pretrained('TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF', model_type='llama', model_file='tinyllama-1.1b-1t-openorca.Q8_0.gguf', **vars(gpu_config)) # **asdict(CtransRunConfig_cpu())
        except:
            #model = AutoModelForCausalLM.from_pretrained('juanjgit/orca_mini_3B-GGUF', model_type='llama', model_file='orca-mini-3b.q4_0.gguf', **vars(cpu_config)) #**asdict(CtransRunConfig_gpu())
            model = AutoModelForCausalLM.from_pretrained('Aryanne/Orca-Mini-3B-gguf', model_type='llama', model_file='q5_0-orca-mini-3b.gguf', **vars(cpu_config)) #**asdict(CtransRunConfig_gpu())
            #model = AutoModelForCausalLM.from_pretrained('Aryanne/Wizard-Orca-3B-gguf', model_type='llama', model_file='q4_1-wizard-orca-3b.gguf', **vars(cpu_config)) # **asdict(CtransRunConfig_cpu())
            #model = AutoModelForCausalLM.from_pretrained('TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF', model_type='llama', model_file='tinyllama-1.1b-1t-openorca.Q8_0.gguf', **vars(cpu_config)) # **asdict(CtransRunConfig_cpu())


        tokenizer = []

    if model_type == "Flan Alpaca (small, fast)":
        # Huggingface chat model
        hf_checkpoint = 'declare-lab/flan-alpaca-large'
        
        def create_hf_model(model_name):

            from transformers import AutoModelForSeq2SeqLM,  AutoModelForCausalLM
            
            if torch_device == "cuda":
                if "flan" in model_name:
                    model = AutoModelForSeq2SeqLM.from_pretrained(model_name, device_map="auto")
                else:
                    model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
            else:
                if "flan" in model_name:
                    model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
                else: 
                    model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

            tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length = chatf.context_length)

            return model, tokenizer, model_type

        model, tokenizer, model_type = create_hf_model(model_name = hf_checkpoint)

    chatf.model = model
    chatf.tokenizer = tokenizer
    chatf.model_type = model_type

    load_confirmation = "Finished loading model: " + model_type

    print(load_confirmation)
    return model_type, load_confirmation, model_type

# Both models are loaded on app initialisation so that users don't have to wait for the models to be downloaded
model_type = "Orca Mini (larger, slow)"

load_model(model_type, chatf.gpu_layers, chatf.gpu_config, chatf.cpu_config, chatf.torch_device)

model_type = "Flan Alpaca (small, fast)"
load_model(model_type, 0, chatf.gpu_config, chatf.cpu_config, chatf.torch_device)

def docs_to_faiss_save(docs_out:PandasDataFrame, embeddings=embeddings):

    print(f"> Total split documents: {len(docs_out)}")

    print(docs_out)

    vectorstore_func = FAISS.from_documents(documents=docs_out, embedding=embeddings)


    chatf.vectorstore = vectorstore_func

    out_message = "Document processing complete"

    return out_message, vectorstore_func

 # Gradio chat

block = gr.Blocks(theme = gr.themes.Base())#css=".gradio-container {background-color: black}")

with block:
    ingest_text = gr.State()
    ingest_metadata = gr.State()
    ingest_docs = gr.State()

    model_type_state = gr.State(model_type)
    embeddings_state = gr.State(globals()["embeddings"])
    vectorstore_state = gr.State(globals()["vectorstore"])  

    model_state = gr.State() # chatf.model (gives error)
    tokenizer_state = gr.State() # chatf.tokenizer (gives error)

    chat_history_state = gr.State()
    instruction_prompt_out = gr.State()

    gr.Markdown("<h1><center>Lightweight PDF / web page QA bot</center></h1>")        
    
    gr.Markdown("Chat with PDF or web page documents. The default is a small model (Flan Alpaca), that can only answer specific questions that are answered in the text. It cannot give overall impressions of, or summarise the document. The alternative (Orca Mini (larger, slow)), can reason a little better, but is much slower (See Advanced tab).\n\nBy default the Lambeth Borough Plan '[Lambeth 2030 : Our Future, Our Lambeth](https://www.lambeth.gov.uk/better-fairer-lambeth/projects/lambeth-2030-our-future-our-lambeth)' is loaded. If you want to talk about another document or web page, please select from the second tab. If switching topic, please click the 'Clear chat' button.\n\nCaution: This is a public app. Please ensure that the document you upload is not sensitive is any way as other users may see it! Also, please note that LLM chatbots may give incomplete or incorrect information, so please use with care.")

    with gr.Row():
        current_source = gr.Textbox(label="Current data source(s)", value="Lambeth_2030-Our_Future_Our_Lambeth.pdf", scale = 10)
        current_model = gr.Textbox(label="Current model", value=model_type, scale = 3)

    with gr.Tab("Chatbot"):

        with gr.Row():
            chat_height = 500
            chatbot = gr.Chatbot(height=chat_height, avatar_images=('user.jfif', 'bot.jpg'),bubble_full_width = False, scale = 1)
            #sources = gr.HTML(value = "Source paragraphs with the most relevant text will appear here", height=chat_height, scale = 2)
            sources = gr.Markdown(value = "Source paragraphs with the most relevant text will appear here", height=chat_height, scale = 2)

        with gr.Row():
            message = gr.Textbox(
                label="What's your question?",
                lines=1,
            )     
        with gr.Row():
            submit = gr.Button(value="Send message", variant="secondary", scale = 1)
            clear = gr.Button(value="Clear chat", variant="secondary", scale=0) 
            stop = gr.Button(value="Stop generating", variant="secondary", scale=0)

        examples_set = gr.Radio(label="Examples for the Lambeth Borough Plan",
            #value = "What were the five pillars of the previous borough plan?",
            choices=["What were the five pillars of the previous borough plan?",
                "What is the vision statement for Lambeth?",
                "What are the commitments for Lambeth?",
                "What are the 2030 outcomes for Lambeth?"])

        
        current_topic = gr.Textbox(label="Feature currently disabled - Keywords related to current conversation topic.", placeholder="Keywords related to the conversation topic will appear here")
            


    with gr.Tab("Load in a different PDF file or web page to chat"):
        with gr.Accordion("PDF file", open = False):
            in_pdf = gr.File(label="Upload pdf", file_count="multiple", file_types=['.pdf'])
            load_pdf = gr.Button(value="Load in file", variant="secondary", scale=0)
        
        with gr.Accordion("Web page", open = False):
            with gr.Row():
                in_web = gr.Textbox(label="Enter webpage url")
                in_div = gr.Textbox(label="(Advanced) Webpage div for text extraction", value="p", placeholder="p")
            load_web = gr.Button(value="Load in webpage", variant="secondary", scale=0) 
        
        ingest_embed_out = gr.Textbox(label="File/webpage preparation progress")

    with gr.Tab("Advanced features"):
        model_choice = gr.Radio(label="Choose a chat model", value="Flan Alpaca (small, fast)", choices = ["Flan Alpaca (small, fast)", "Orca Mini (larger, slow)"])
        with gr.Row():
            gpu_layer_choice = gr.Slider(label="Choose number of model layers to send to GPU (WARNING: please don't modify unless you have a GPU).", value=0, minimum=0, maximum=6, step = 1, visible=False)
            change_model_button = gr.Button(value="Load model", scale=0)
        load_text = gr.Text(label="Load status")

    gr.HTML(
        "<center>This app is based on the models Flan Alpaca and Orca Mini. It powered by Gradio, Transformers, Ctransformers, and Langchain.</a></center>"
    )

    examples_set.change(fn=chatf.update_message, inputs=[examples_set], outputs=[message])

    change_model_button.click(fn=chatf.turn_off_interactivity, inputs=[message, chatbot], outputs=[message, chatbot], queue=False).\
    then(fn=load_model, inputs=[model_choice, gpu_layer_choice], outputs = [model_type_state, load_text, current_model]).\
    then(lambda: chatf.restore_interactivity(), None, [message], queue=False).\
    then(chatf.clear_chat, inputs=[chat_history_state, sources, message, current_topic], outputs=[chat_history_state, sources, message, current_topic]).\
    then(lambda: None, None, chatbot, queue=False)

    # Load in a pdf
    load_pdf_click = load_pdf.click(ing.parse_file, inputs=[in_pdf], outputs=[ingest_text, current_source]).\
             then(ing.text_to_docs, inputs=[ingest_text], outputs=[ingest_docs]).\
             then(docs_to_faiss_save, inputs=[ingest_docs], outputs=[ingest_embed_out, vectorstore_state]).\
             then(chatf.hide_block, outputs = [examples_set])

    # Load in a webpage
    load_web_click = load_web.click(ing.parse_html, inputs=[in_web, in_div], outputs=[ingest_text, ingest_metadata, current_source]).\
             then(ing.html_text_to_docs, inputs=[ingest_text, ingest_metadata], outputs=[ingest_docs]).\
             then(docs_to_faiss_save, inputs=[ingest_docs], outputs=[ingest_embed_out, vectorstore_state]).\
             then(chatf.hide_block, outputs = [examples_set])

    # Load in a webpage

    # Click/enter to send message action
    response_click = submit.click(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False, api_name="retrieval").\
                then(chatf.turn_off_interactivity, inputs=[message, chatbot], outputs=[message, chatbot], queue=False).\
                then(chatf.produce_streaming_answer_chatbot, inputs=[chatbot, instruction_prompt_out, model_type_state], outputs=chatbot)
    response_click.then(chatf.highlight_found_text, [chatbot, sources], [sources]).\
                then(chatf.add_inputs_answer_to_history,[message, chatbot, current_topic], [chat_history_state, current_topic]).\
                then(lambda: chatf.restore_interactivity(), None, [message], queue=False)

    response_enter = message.submit(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False).\
                then(chatf.turn_off_interactivity, inputs=[message, chatbot], outputs=[message, chatbot], queue=False).\
                then(chatf.produce_streaming_answer_chatbot, [chatbot, instruction_prompt_out, model_type_state], chatbot)    
    response_enter.then(chatf.highlight_found_text, [chatbot, sources], [sources]).\
                then(chatf.add_inputs_answer_to_history,[message, chatbot, current_topic], [chat_history_state, current_topic]).\
                then(lambda: chatf.restore_interactivity(), None, [message], queue=False)
    
    # Stop box
    stop.click(fn=None, inputs=None, outputs=None, cancels=[response_click, response_enter])

    # Clear box
    clear.click(chatf.clear_chat, inputs=[chat_history_state, sources, message, current_topic], outputs=[chat_history_state, sources, message, current_topic])
    clear.click(lambda: None, None, chatbot, queue=False)

    chatbot.like(chatf.vote, [chat_history_state, instruction_prompt_out, model_type_state], None)

block.queue(concurrency_count=1).launch(debug=True)
# -