File size: 37,095 Bytes
49e32ea
 
 
 
 
 
 
 
 
d2ddc62
49e32ea
 
 
 
d2ddc62
49e32ea
 
 
 
 
 
 
 
 
 
 
 
9a3229c
49e32ea
 
 
 
9a3229c
49e32ea
 
 
 
 
 
 
 
 
 
d2ddc62
 
e4df9f2
d2ddc62
 
49e32ea
f6036ad
49e32ea
 
 
 
 
 
 
ed48a70
 
49e32ea
 
 
 
 
 
 
 
aa0ad5d
49e32ea
 
 
 
 
 
 
e4df9f2
aa0ad5d
f6036ad
49e32ea
d2ddc62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49e32ea
aa0ad5d
49e32ea
 
 
 
 
 
 
 
1ae7b34
49e32ea
aa0ad5d
 
d2ddc62
49e32ea
d2ddc62
aa0ad5d
 
 
 
49e32ea
aa0ad5d
 
 
 
 
 
49e32ea
aa0ad5d
49e32ea
aa0ad5d
 
 
 
 
 
 
 
 
49e32ea
aa0ad5d
 
 
 
 
 
49e32ea
aa0ad5d
49e32ea
aa0ad5d
49e32ea
aa0ad5d
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d2ddc62
e4df9f2
 
 
 
 
 
 
d2ddc62
49e32ea
 
 
 
 
 
 
 
 
ed48a70
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102df35
49e32ea
 
102df35
 
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ae4a7ec
f6036ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102df35
49e32ea
 
 
 
 
 
 
 
 
 
 
102df35
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6036ad
 
49e32ea
 
 
102df35
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102df35
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ae4a7ec
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6036ad
49e32ea
 
 
f6036ad
49e32ea
 
 
 
 
aa0ad5d
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d2ddc62
 
aa0ad5d
49e32ea
 
aa0ad5d
 
 
 
 
49e32ea
d2ddc62
 
 
aa0ad5d
49e32ea
aa0ad5d
 
 
 
 
 
 
 
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71c040a
 
 
 
 
 
49e32ea
 
 
 
 
 
 
 
 
 
ed48a70
49e32ea
ed48a70
49e32ea
 
ed48a70
 
49e32ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed48a70
49e32ea
 
 
e4df9f2
 
 
49e32ea
 
ed48a70
49e32ea
 
ed48a70
 
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
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
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
import re
import datetime
from typing import TypeVar, Dict, List, Tuple
from itertools import compress
import pandas as pd
import numpy as np

# Model packages
import torch
torch.cuda.empty_cache()
from threading import Thread
from transformers import AutoTokenizer, pipeline, TextIteratorStreamer

# Alternative model sources
from ctransformers import AutoModelForCausalLM#, AutoTokenizer
from dataclasses import asdict, dataclass

# Langchain functions
from langchain import PromptTemplate
from langchain.prompts import PromptTemplate
from langchain.vectorstores import FAISS
from langchain.retrievers import SVMRetriever 
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.docstore.document import Document

# For keyword extraction
import nltk
nltk.download('wordnet')
from nltk.corpus import stopwords
from nltk.tokenize import RegexpTokenizer
from nltk.stem import WordNetLemmatizer
import keybert

# For Name Entity Recognition model
from span_marker import SpanMarkerModel

# For BM25 retrieval
from gensim.corpora import Dictionary
from gensim.models import TfidfModel, OkapiBM25Model
from gensim.similarities import SparseMatrixSimilarity

import gradio as gr

if torch.cuda.is_available():
    torch_device = "cuda"
    gpu_layers = 5
else: torch_device =  "cpu"

print("Running on device:", torch_device)
threads = 8#torch.get_num_threads()
print("CPU threads:", threads)

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

embeddings = None  # global variable setup
vectorstore = None # global variable setup

max_memory_length = 0 # How long should the memory of the conversation last?

full_text = "" # Define dummy source text (full text) just to enable highlight function to load

ctrans_llm = [] # Define empty list to hold CTrans LLMs for functions to run

temperature: float = 0.1
top_k: int = 3
top_p: float = 1
repetition_penalty: float = 1.05
flan_alpaca_repetition_penalty: float = 1.3
last_n_tokens: int = 64
max_new_tokens: int = 125
#seed: int = 42
reset: bool = False
stream: bool = True
threads: int = threads
batch_size:int = 512
context_length:int = 4096
gpu_layers:int = 0#5#gpu_layers For serving on Huggingface set to 0 as using free CPU instance
sample = True

@dataclass
class GenerationConfig:
    temperature: float = temperature
    top_k: int = top_k
    top_p: float = top_p
    repetition_penalty: float = repetition_penalty
    last_n_tokens: int = last_n_tokens
    max_new_tokens: int = max_new_tokens
    #seed: int = 42
    reset: bool = reset
    stream: bool = stream
    threads: int = threads
    batch_size:int = batch_size
    context_length:int = context_length
    gpu_layers:int = gpu_layers
    #stop: list[str] = field(default_factory=lambda: [stop_string])


## Highlight text constants
hlt_chunk_size = 15
hlt_strat = [" ", ".", "!", "?", ":", "\n\n", "\n", ","]
hlt_overlap = 0

## Initialise NER model ##
ner_model = SpanMarkerModel.from_pretrained("tomaarsen/span-marker-mbert-base-multinerd")

## Initialise keyword model ##
# Used to pull out keywords from chat history to add to user queries behind the scenes
kw_model = pipeline("feature-extraction", model="sentence-transformers/all-MiniLM-L6-v2")

## Set model type ##
model_type = "ctrans"

## Chat models ##

if model_type == "ctrans":
    ctrans_llm = AutoModelForCausalLM.from_pretrained('juanjgit/orca_mini_3B-GGUF', model_type='llama', model_file='orca-mini-3b.q4_0.gguf', **asdict(GenerationConfig()))
    #ctrans_llm = AutoModelForCausalLM.from_pretrained('TheBloke/Mistral-7B-OpenOrca-GGUF', model_type='mistral', model_file='mistral-7b-openorca.Q4_K_M.gguf', **asdict(GenerationConfig()))
    #ctrans_llm = AutoModelForCausalLM.from_pretrained('TheBloke/Mistral-7B-OpenOrca-GGUF', model_type='mistral', model_file='mistral-7b-openorca.Q2_K.gguf', **asdict(GenerationConfig()))

if model_type == "hf":
    # Huggingface chat model
    #hf_checkpoint = 'jphme/phi-1_5_Wizard_Vicuna_uncensored'
    hf_checkpoint = 'declare-lab/flan-alpaca-large'
    
    def create_hf_model(model_name):

        from transformers import AutoModelForSeq2SeqLM,  AutoModelForCausalLM

    #    model_id = model_name
        
        if torch_device == "cuda":
            if "flan" in model_name:
                model = AutoModelForSeq2SeqLM.from_pretrained(model_name, load_in_8bit=True, device_map="auto")
            elif "mpt" in model_name:
                model = AutoModelForCausalLM.from_pretrained(model_name, load_in_8bit=True, device_map="auto", trust_remote_code=True)
            else:
                model = AutoModelForCausalLM.from_pretrained(model_name, load_in_8bit=True, device_map="auto")
        else:
            if "flan" in model_name:
                model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
            elif "mpt" in model_name:    
                model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
            else: 
                model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

        tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length = 2048)

        return model, tokenizer, torch_device

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

# Vectorstore funcs

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

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

    vectorstore_func = FAISS.from_documents(documents=docs_out, embedding=embeddings)
        
    '''  
    #with open("vectorstore.pkl", "wb") as f:
        #pickle.dump(vectorstore, f) 
    ''' 

    #if Path(save_to).exists():
    #    vectorstore_func.save_local(folder_path=save_to)
    #else:
    #    os.mkdir(save_to)
    #    vectorstore_func.save_local(folder_path=save_to)

    global vectorstore

    vectorstore = vectorstore_func

    out_message = "Document processing complete"

    #print(out_message)
    #print(f"> Saved to: {save_to}")

    return out_message

# # Prompt functions

def create_prompt_templates():    
  
    #EXAMPLE_PROMPT = PromptTemplate(
    #    template="\nCONTENT:\n\n{page_content}\n\nSOURCE: {source}\n\n",
    #    input_variables=["page_content", "source"],
    #)

    CONTENT_PROMPT = PromptTemplate(
        template="{page_content}\n\n",#\n\nSOURCE: {source}\n\n",
        input_variables=["page_content"]
    )


# The main prompt:

    instruction_prompt_template_alpaca_quote = """### Instruction:
    Quote directly from the SOURCE below that best answers the QUESTION. Only quote full sentences in the correct order. If you cannot find an answer, start your response with "My best guess is: ".
    
    CONTENT: {summaries}
    
    QUESTION: {question}

    Response:"""

    instruction_prompt_template_orca = """
    ### System:
    You are an AI assistant that follows instruction extremely well. Help as much as you can.
    ### User:
    Answer the QUESTION using information from the following CONTENT.
    CONTENT: {summaries}
    QUESTION: {question}

    ### Response:"""


    instruction_prompt_mistral_orca = """<|im_start|>system\n
You are an AI assistant that follows instruction extremely well. Help as much as you can.
<|im_start|>user\n
Answer the QUESTION using information from the following CONTENT.
CONTENT: {summaries}
QUESTION: {question}\n
<|im_end|>"""

 

   
    INSTRUCTION_PROMPT=PromptTemplate(template=instruction_prompt_template_orca, input_variables=['question', 'summaries'])
    
    return INSTRUCTION_PROMPT, CONTENT_PROMPT

def adapt_q_from_chat_history(question, chat_history, extracted_memory, keyword_model=""):#keyword_model): # new_question_keywords, 
 
        chat_history_str, chat_history_first_q, chat_history_first_ans, max_memory_length = _get_chat_history(chat_history)

        if chat_history_str:
            # Keyword extraction is now done in the add_inputs_to_history function
            extracted_memory = extracted_memory#remove_q_stopwords(str(chat_history_first_q) + " " + str(chat_history_first_ans))
            
           
            new_question_kworded = str(extracted_memory) + ". " + question #+ " " + new_question_keywords
            #extracted_memory + " " + question
            
        else:
            new_question_kworded = question #new_question_keywords

        #print("Question output is: " + new_question_kworded)
            
        return new_question_kworded

def create_doc_df(docs_keep_out):
    # Extract content and metadata from 'winning' passages.
            content=[]
            meta=[]
            meta_url=[]
            page_section=[]
            score=[]

            for item in docs_keep_out:
                content.append(item[0].page_content)
                meta.append(item[0].metadata)
                meta_url.append(item[0].metadata['source'])
                page_section.append(item[0].metadata['page_section'])
                score.append(item[1])       

            # Create df from 'winning' passages

            doc_df = pd.DataFrame(list(zip(content, meta, page_section, meta_url, score)),
               columns =['page_content', 'metadata', 'page_section', 'meta_url', 'score'])

            docs_content = doc_df['page_content'].astype(str)
            doc_df['full_url'] = "https://" + doc_df['meta_url'] 

            return doc_df

def hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val, out_passages,
                           vec_score_cut_off, vec_weight, bm25_weight, svm_weight): # ,vectorstore, embeddings

            #vectorstore=globals()["vectorstore"]
            #embeddings=globals()["embeddings"]


            docs = vectorstore.similarity_search_with_score(new_question_kworded, k=k_val)

            print("Docs from similarity search:")
            print(docs)

            # Keep only documents with a certain score
            docs_len = [len(x[0].page_content) for x in docs]
            docs_scores = [x[1] for x in docs]

            # Only keep sources that are sufficiently relevant (i.e. similarity search score below threshold below)
            score_more_limit = pd.Series(docs_scores) < vec_score_cut_off
            docs_keep = list(compress(docs, score_more_limit))

            if docs_keep == []:
                docs_keep_as_doc = []
                docs_content = []
                docs_url = []
                return docs_keep_as_doc, docs_content, docs_url

            # Only keep sources that are at least 100 characters long
            length_more_limit = pd.Series(docs_len) >= 100
            docs_keep = list(compress(docs_keep, length_more_limit))

            if docs_keep == []:
                docs_keep_as_doc = []
                docs_content = []
                docs_url = []
                return docs_keep_as_doc, docs_content, docs_url

            docs_keep_as_doc = [x[0] for x in docs_keep]
            docs_keep_length = len(docs_keep_as_doc)


                
            if docs_keep_length == 1:

                content=[]
                meta_url=[]
                score=[]
                
                for item in docs_keep:
                    content.append(item[0].page_content)
                    meta_url.append(item[0].metadata['source'])
                    score.append(item[1])       

                # Create df from 'winning' passages

                doc_df = pd.DataFrame(list(zip(content, meta_url, score)),
                columns =['page_content', 'meta_url', 'score'])

                docs_content = doc_df['page_content'].astype(str)
                docs_url = doc_df['meta_url']

                return docs_keep_as_doc, docs_content, docs_url
            
            # Check for if more docs are removed than the desired output
            if out_passages > docs_keep_length: 
                out_passages = docs_keep_length
                k_val = docs_keep_length
                     
            vec_rank = [*range(1, docs_keep_length+1)]
            vec_score = [(docs_keep_length/x)*vec_weight for x in vec_rank]

            # 2nd level check on retrieved docs with BM25

            content_keep=[]
            for item in docs_keep:
                content_keep.append(item[0].page_content)

            corpus = corpus = [doc.lower().split() for doc in content_keep]
            dictionary = Dictionary(corpus)
            bm25_model = OkapiBM25Model(dictionary=dictionary)
            bm25_corpus = bm25_model[list(map(dictionary.doc2bow, corpus))]
            bm25_index = SparseMatrixSimilarity(bm25_corpus, num_docs=len(corpus), num_terms=len(dictionary),
                                   normalize_queries=False, normalize_documents=False)
            query = new_question_kworded.lower().split()
            tfidf_model = TfidfModel(dictionary=dictionary, smartirs='bnn')  # Enforce binary weighting of queries
            tfidf_query = tfidf_model[dictionary.doc2bow(query)]
            similarities = np.array(bm25_index[tfidf_query])
            #print(similarities)
            temp = similarities.argsort()
            ranks = np.arange(len(similarities))[temp.argsort()][::-1]

            # Pair each index with its corresponding value
            pairs = list(zip(ranks, docs_keep_as_doc))
            # Sort the pairs by the indices
            pairs.sort()
            # Extract the values in the new order
            bm25_result = [value for ranks, value in pairs]
            
            bm25_rank=[]
            bm25_score = []

            for vec_item in docs_keep:
                x = 0
                for bm25_item in bm25_result:
                    x = x + 1
                    if bm25_item.page_content == vec_item[0].page_content:
                        bm25_rank.append(x)
                        bm25_score.append((docs_keep_length/x)*bm25_weight)

            # 3rd level check on retrieved docs with SVM retriever
            svm_retriever = SVMRetriever.from_texts(content_keep, embeddings, k = k_val)
            svm_result = svm_retriever.get_relevant_documents(new_question_kworded)

         
            svm_rank=[]
            svm_score = []

            for vec_item in docs_keep:
                x = 0
                for svm_item in svm_result:
                    x = x + 1
                    if svm_item.page_content == vec_item[0].page_content:
                        svm_rank.append(x)
                        svm_score.append((docs_keep_length/x)*svm_weight)

        
            ## Calculate final score based on three ranking methods
            final_score = [a  + b + c for a, b, c in zip(vec_score, bm25_score, svm_score)]
            final_rank = [sorted(final_score, reverse=True).index(x)+1 for x in final_score]
            # Force final_rank to increment by 1 each time
            final_rank = list(pd.Series(final_rank).rank(method='first'))

            #print("final rank: " + str(final_rank))
            #print("out_passages: " + str(out_passages))

            best_rank_index_pos = []

            for x in range(1,out_passages+1):
                try:
                    best_rank_index_pos.append(final_rank.index(x))
                except IndexError: # catch the error
                    pass

            # Adjust best_rank_index_pos to 

            best_rank_pos_series = pd.Series(best_rank_index_pos)


            docs_keep_out = [docs_keep[i] for i in best_rank_index_pos]
        
            # Keep only 'best' options
            docs_keep_as_doc = [x[0] for x in docs_keep_out]
                               
            # Make df of best options
            doc_df = create_doc_df(docs_keep_out)

            return docs_keep_as_doc, doc_df, docs_keep_out

def get_expanded_passages(vectorstore, docs, width):

    """
    Extracts expanded passages based on given documents and a width for context.
    
    Parameters:
    - vectorstore: The primary data source.
    - docs: List of documents to be expanded.
    - width: Number of documents to expand around a given document for context.
    
    Returns:
    - expanded_docs: List of expanded Document objects.
    - doc_df: DataFrame representation of expanded_docs.
    """

    from collections import defaultdict
    
    def get_docs_from_vstore(vectorstore):
        vector = vectorstore.docstore._dict
        return list(vector.items())

    def extract_details(docs_list):
        docs_list_out = [tup[1] for tup in docs_list]
        content = [doc.page_content for doc in docs_list_out]
        meta = [doc.metadata for doc in docs_list_out]
        return ''.join(content), meta[0], meta[-1]
    
    def get_parent_content_and_meta(vstore_docs, width, target):
        target_range = range(max(0, target - width), min(len(vstore_docs), target + width + 1))
        parent_vstore_out = [vstore_docs[i] for i in target_range]
        
        content_str_out, meta_first_out, meta_last_out = [], [], []
        for _ in parent_vstore_out:
            content_str, meta_first, meta_last = extract_details(parent_vstore_out)
            content_str_out.append(content_str)
            meta_first_out.append(meta_first)
            meta_last_out.append(meta_last)
        return content_str_out, meta_first_out, meta_last_out

    def merge_dicts_except_source(d1, d2):
            merged = {}
            for key in d1:
                if key != "source":
                    merged[key] = str(d1[key]) + " to " + str(d2[key])
                else:
                    merged[key] = d1[key]  # or d2[key], based on preference
            return merged

    def merge_two_lists_of_dicts(list1, list2):
        return [merge_dicts_except_source(d1, d2) for d1, d2 in zip(list1, list2)]

    # Step 1: Filter vstore_docs
    vstore_docs = get_docs_from_vstore(vectorstore)
    print("Inside get_expanded_passages")
    print("Docs:", docs)
    print("Type of Docs:", type(docs))
    print("Type of first element in Docs:", type(docs[0]))
    print("Length of first tuple in Docs:", len(docs[0]))

    doc_sources = {doc.metadata['source'] for doc, _ in docs}
    vstore_docs = [(k, v) for k, v in vstore_docs if v.metadata.get('source') in doc_sources]

    # Step 2: Group by source and proceed
    vstore_by_source = defaultdict(list)
    for k, v in vstore_docs:
        vstore_by_source[v.metadata['source']].append((k, v))
        
    expanded_docs = []
    for doc, score in docs:
        search_source = doc.metadata['source']
        search_section = doc.metadata['page_section']
        parent_vstore_meta_section = [doc.metadata['page_section'] for _, doc in vstore_by_source[search_source]]
        search_index = parent_vstore_meta_section.index(search_section) if search_section in parent_vstore_meta_section else -1

        content_str, meta_first, meta_last = get_parent_content_and_meta(vstore_by_source[search_source], width, search_index)
        meta_full = merge_two_lists_of_dicts(meta_first, meta_last)

        expanded_doc = (Document(page_content=content_str[0], metadata=meta_full[0]), score)
        expanded_docs.append(expanded_doc)

    doc_df = create_doc_df(expanded_docs)  # Assuming you've defined the 'create_doc_df' function elsewhere

    return expanded_docs, doc_df

def create_final_prompt(inputs: Dict[str, str], instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings): # , 
        
        question =  inputs["question"]
        chat_history = inputs["chat_history"]
        

        new_question_kworded = adapt_q_from_chat_history(question, chat_history, extracted_memory) # new_question_keywords, 
        

        #print("The question passed to the vector search is:")
        #print(new_question_kworded)
        
        docs_keep_as_doc, doc_df, docs_keep_out = hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val = 5, out_passages = 2,
                                                                          vec_score_cut_off = 1, vec_weight = 1, bm25_weight = 1, svm_weight = 1)#,
                                                                          #vectorstore=globals()["vectorstore"], embeddings=globals()["embeddings"])
        
        # Expand the found passages to the neighbouring context
        docs_keep_as_doc, doc_df = get_expanded_passages(vectorstore, docs_keep_out, width=1)

        if docs_keep_as_doc == []:
            {"answer": "I'm sorry, I couldn't find a relevant answer to this question.", "sources":"I'm sorry, I couldn't find a relevant source for this question."}
        
        #new_inputs = inputs.copy()
        #new_inputs["question"] = new_question
        #new_inputs["chat_history"] = chat_history_str
        
        #print(docs_url)
        #print(doc_df['metadata'])

        # Build up sources content to add to user display

        doc_df['meta_clean'] = [f"<b>{'  '.join(f'{k}: {v}' for k, v in d.items() if k != 'page_section')}</b>" for d in doc_df['metadata']]
        doc_df['content_meta'] = doc_df['meta_clean'].astype(str) + ".<br><br>" + doc_df['page_content'].astype(str)

        modified_page_content = [f" SOURCE {i+1} - {word}" for i, word in enumerate(doc_df['page_content'])]
        docs_content_string = ''.join(modified_page_content)

        #docs_content_string = '<br><br>\n\n SOURCE '.join(doc_df['page_content'])#.replace("  "," ")#.strip()
        sources_docs_content_string = '<br><br>'.join(doc_df['content_meta'])#.replace("  "," ")#.strip()
        #sources_docs_content_tup = [(sources_docs_content,None)]
        #print("The draft instruction prompt is:")
        #print(instruction_prompt)
        
        instruction_prompt_out = instruction_prompt.format(question=new_question_kworded, summaries=docs_content_string)
        #print("The final instruction prompt:")
        #print(instruction_prompt_out)
        
        print('Final prompt is: ')
        print(instruction_prompt_out)
                
        return instruction_prompt_out, sources_docs_content_string, new_question_kworded

def get_history_sources_final_input_prompt(user_input, history, extracted_memory, vectorstore, embeddings):#):
    
    #if chain_agent is None:
    #    history.append((user_input, "Please click the button to submit the Huggingface API key before using the chatbot (top right)"))
    #    return history, history, "", ""
    print("\n==== date/time: " + str(datetime.datetime.now()) + " ====")
    print("User input: " + user_input)
    
    history = history or []
    

    
    # Create instruction prompt
    instruction_prompt, content_prompt = create_prompt_templates()
    instruction_prompt_out, docs_content_string, new_question_kworded =\
                create_final_prompt({"question": user_input, "chat_history": history}, #vectorstore,
                                    instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings)
    
  
    history.append(user_input)
    
    print("Output history is:")
    print(history)

    #print("The output prompt is:")
    #print(instruction_prompt_out)
    
    return history, docs_content_string, instruction_prompt_out

def highlight_found_text_single(search_text:str, full_text:str, hlt_chunk_size:int=hlt_chunk_size, hlt_strat:List=hlt_strat, hlt_overlap:int=hlt_overlap) -> str:
    """
    Highlights occurrences of search_text within full_text.
    
    Parameters:
    - search_text (str): The text to be searched for within full_text.
    - full_text (str): The text within which search_text occurrences will be highlighted.
    
    Returns:
    - str: A string with occurrences of search_text highlighted.
    
    Example:
    >>> highlight_found_text("world", "Hello, world! This is a test. Another world awaits.")
    'Hello, <mark style="color:black;">world</mark>! This is a test. Another world awaits.'
    """

    def extract_text_from_input(text,i=0):
        if isinstance(text, str):
            return text.replace("  ", " ").strip()#.replace("\r", " ").replace("\n", " ")
        elif isinstance(text, list):
            return text[i][0].replace("  ", " ").strip()#.replace("\r", " ").replace("\n", " ")
        else:
            return ""
        
    def extract_search_text_from_input(text):
        if isinstance(text, str):
            return text.replace("  ", " ").strip()#.replace("\r", " ").replace("\n", " ").replace("  ", " ").strip()
        elif isinstance(text, list):
            return text[-1][1].replace("  ", " ").strip()#.replace("\r", " ").replace("\n", " ").replace("  ", " ").strip()
        else:
            return ""
            
    full_text = extract_text_from_input(full_text)
    search_text = extract_search_text_from_input(search_text)

    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=hlt_chunk_size,
        separators=hlt_strat,
        chunk_overlap=hlt_overlap,
    )
    sections = text_splitter.split_text(search_text)

    #print(sections)

    found_positions = {}
    for x in sections:
        text_start_pos = full_text.find(x)
        
        if text_start_pos != -1:
            found_positions[text_start_pos] = text_start_pos + len(x)

    # Combine overlapping or adjacent positions
    sorted_starts = sorted(found_positions.keys())
    combined_positions = []
    if sorted_starts:
        current_start, current_end = sorted_starts[0], found_positions[sorted_starts[0]]
        for start in sorted_starts[1:]:
            if start <= (current_end + 1):
                current_end = max(current_end, found_positions[start])
            else:
                combined_positions.append((current_start, current_end))
                current_start, current_end = start, found_positions[start]
        combined_positions.append((current_start, current_end))

    # Construct pos_tokens
    pos_tokens = []
    prev_end = 0
    for start, end in combined_positions:
        pos_tokens.append(full_text[prev_end:start]) # ((full_text[prev_end:start], None))
        pos_tokens.append('<mark style="color:black;">' + full_text[start:end] + '</mark>')# ("<mark>" + full_text[start:end] + "</mark>",'found')
        prev_end = end
    pos_tokens.append(full_text[prev_end:])

    return "".join(pos_tokens)

def highlight_found_text(search_text: str, full_text: str, hlt_chunk_size:int=hlt_chunk_size, hlt_strat:List=hlt_strat, hlt_overlap:int=hlt_overlap) -> str:
    """
    Highlights occurrences of search_text within full_text.
    
    Parameters:
    - search_text (str): The text to be searched for within full_text.
    - full_text (str): The text within which search_text occurrences will be highlighted.
    
    Returns:
    - str: A string with occurrences of search_text highlighted.
    
    Example:
    >>> highlight_found_text("world", "Hello, world! This is a test. Another world awaits.")
    'Hello, <mark style="color:black;">world</mark>! This is a test. Another <mark style="color:black;">world</mark> awaits.'
    """

    def extract_text_from_input(text, i=0):
        if isinstance(text, str):
            return text.replace("  ", " ").strip()
        elif isinstance(text, list):
            return text[i][0].replace("  ", " ").strip()
        else:
            return ""

    def extract_search_text_from_input(text):
        if isinstance(text, str):
            return text.replace("  ", " ").strip()
        elif isinstance(text, list):
            return text[-1][1].replace("  ", " ").strip()
        else:
            return ""

    full_text = extract_text_from_input(full_text)
    search_text = extract_search_text_from_input(search_text)



    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=hlt_chunk_size,
        separators=hlt_strat,
        chunk_overlap=hlt_overlap,
    )
    sections = text_splitter.split_text(search_text)

    found_positions = {}
    for x in sections:
        text_start_pos = 0
        while text_start_pos != -1:
            text_start_pos = full_text.find(x, text_start_pos)
            if text_start_pos != -1:
                found_positions[text_start_pos] = text_start_pos + len(x)
                text_start_pos += 1

    # Combine overlapping or adjacent positions
    sorted_starts = sorted(found_positions.keys())
    combined_positions = []
    if sorted_starts:
        current_start, current_end = sorted_starts[0], found_positions[sorted_starts[0]]
        for start in sorted_starts[1:]:
            if start <= (current_end + 10):
                current_end = max(current_end, found_positions[start])
            else:
                combined_positions.append((current_start, current_end))
                current_start, current_end = start, found_positions[start]
        combined_positions.append((current_start, current_end))

    # Construct pos_tokens
    pos_tokens = []
    prev_end = 0
    for start, end in combined_positions:
        pos_tokens.append(full_text[prev_end:start])
        pos_tokens.append('<mark style="color:black;">' + full_text[start:end] + '</mark>')
        prev_end = end
    pos_tokens.append(full_text[prev_end:])

    return "".join(pos_tokens)

# # Chat functions
def produce_streaming_answer_chatbot_hf(history, full_prompt): 
    
    #print("The question is: ")
    #print(full_prompt)
    
    # Get the model and tokenizer, and tokenize the user text.
    model_inputs = tokenizer(text=full_prompt, return_tensors="pt", return_attention_mask=False).to(torch_device) # return_attention_mask=False was added

    # Start generation on a separate thread, so that we don't block the UI. The text is pulled from the streamer
    # in the main thread. Adds timeout to the streamer to handle exceptions in the generation thread.
    streamer = TextIteratorStreamer(tokenizer, timeout=120., skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=sample,
        repetition_penalty=flan_alpaca_repetition_penalty,
        top_p=top_p,
        temperature=temperature,
        top_k=top_k
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    # Pull the generated text from the streamer, and update the model output.
    import time
    start = time.time()
    NUM_TOKENS=0
    print('-'*4+'Start Generation'+'-'*4)

    history[-1][1] = ""
    for new_text in streamer:
        if new_text == None: new_text = ""
        history[-1][1] += new_text
        NUM_TOKENS+=1
        yield history
        
    time_generate = time.time() - start
    print('\n')
    print('-'*4+'End Generation'+'-'*4)
    print(f'Num of generated tokens: {NUM_TOKENS}')
    print(f'Time for complete generation: {time_generate}s')
    print(f'Tokens per secound: {NUM_TOKENS/time_generate}')
    print(f'Time per token: {(time_generate/NUM_TOKENS)*1000}ms')

def produce_streaming_answer_chatbot_ctrans(history, full_prompt): 
    
    print("The question is: ")
    print(full_prompt)

    tokens = ctrans_llm.tokenize(full_prompt)
    
    #config = GenerationConfig(reset=True)

    # Pull the generated text from the streamer, and update the model output.
    import time
    start = time.time()
    NUM_TOKENS=0
    print('-'*4+'Start Generation'+'-'*4)

    history[-1][1] = ""
    for new_text in ctrans_llm.generate(tokens, top_k=top_k, temperature=temperature, repetition_penalty=repetition_penalty): #ctrans_generate(prompt=tokens, config=config):
        if new_text == None: new_text =  ""
        history[-1][1] += ctrans_llm.detokenize(new_text) #new_text
        NUM_TOKENS+=1
        yield history
    
    time_generate = time.time() - start
    print('\n')
    print('-'*4+'End Generation'+'-'*4)
    print(f'Num of generated tokens: {NUM_TOKENS}')
    print(f'Time for complete generation: {time_generate}s')
    print(f'Tokens per secound: {NUM_TOKENS/time_generate}')
    print(f'Time per token: {(time_generate/NUM_TOKENS)*1000}ms')


def ctrans_generate(
    prompt: str,
    llm=ctrans_llm,
    config: GenerationConfig = GenerationConfig(),
):
    """Run model inference, will return a Generator if streaming is true."""

    return llm(
        prompt,
        **asdict(config),
    )

def turn_off_interactivity(user_message, history):
        return gr.update(value="", interactive=False), history + [[user_message, None]]

def update_message(dropdown_value):
        return gr.Textbox.update(value=dropdown_value)

def hide_block():
        return gr.Radio.update(visible=False)

# # Chat history functions

def clear_chat(chat_history_state, sources, chat_message, current_topic):
    chat_history_state = []
    sources = ''
    chat_message = ''
    current_topic = ''

    return chat_history_state, sources, chat_message, current_topic

def _get_chat_history(chat_history: List[Tuple[str, str]], max_memory_length:int = max_memory_length): # Limit to last x interactions only

    if (not chat_history) | (max_memory_length == 0):
        chat_history = []

    if len(chat_history) > max_memory_length:
        chat_history = chat_history[-max_memory_length:]
        
    #print(chat_history)

    first_q = ""
    first_ans = ""
    for human_s, ai_s in chat_history:
        first_q = human_s
        first_ans = ai_s

        #print("Text to keyword extract: " + first_q + " " + first_ans)
        break

    conversation = ""
    for human_s, ai_s in chat_history:
        human = f"Human: " + human_s
        ai = f"Assistant: " + ai_s
        conversation += "\n" + "\n".join([human, ai])

    return conversation, first_q, first_ans, max_memory_length

def add_inputs_answer_to_history(user_message, history, current_topic):
    
    if history is None:
        history = [("","")]

    #history.append((user_message, [-1]))

    chat_history_str, chat_history_first_q, chat_history_first_ans, max_memory_length = _get_chat_history(history)


    # Only get the keywords for the first question and response, or do it every time if over 'max_memory_length' responses in the conversation
    if (len(history) == 1) | (len(history) > max_memory_length):
        
        #print("History after appending is:")
        #print(history)

        first_q_and_first_ans = str(chat_history_first_q) + " " + str(chat_history_first_ans)
        #ner_memory = remove_q_ner_extractor(first_q_and_first_ans)
        keywords = keybert_keywords(first_q_and_first_ans, n = 8, kw_model=kw_model)
        #keywords.append(ner_memory)

        # Remove duplicate words while preserving order
        ordered_tokens = set()
        result = []
        for word in keywords:
                if word not in ordered_tokens:
                        ordered_tokens.add(word)
                        result.append(word)

        extracted_memory = ' '.join(result)

    else: extracted_memory=current_topic
    
    print("Extracted memory is:")
    print(extracted_memory)
    
    
    return history, extracted_memory

def remove_q_stopwords(question): # Remove stopwords from question. Not used at the moment 
    # Prepare keywords from question by removing stopwords
    text = question.lower()

    # Remove numbers
    text = re.sub('[0-9]', '', text)

    tokenizer = RegexpTokenizer(r'\w+')
    text_tokens = tokenizer.tokenize(text)
    #text_tokens = word_tokenize(text)
    tokens_without_sw = [word for word in text_tokens if not word in stopwords]

    # Remove duplicate words while preserving order
    ordered_tokens = set()
    result = []
    for word in tokens_without_sw:
        if word not in ordered_tokens:
            ordered_tokens.add(word)
            result.append(word)
     


    new_question_keywords = ' '.join(result)
    return new_question_keywords

def remove_q_ner_extractor(question):
    
    predict_out = ner_model.predict(question)



    predict_tokens = [' '.join(v for k, v in d.items() if k == 'span') for d in predict_out]

    # Remove duplicate words while preserving order
    ordered_tokens = set()
    result = []
    for word in predict_tokens:
        if word not in ordered_tokens:
            ordered_tokens.add(word)
            result.append(word)
     


    new_question_keywords = ' '.join(result).lower()
    return new_question_keywords

def apply_lemmatize(text, wnl=WordNetLemmatizer()):

    def prep_for_lemma(text):

        # Remove numbers
        text = re.sub('[0-9]', '', text)
        print(text)

        tokenizer = RegexpTokenizer(r'\w+')
        text_tokens = tokenizer.tokenize(text)
        #text_tokens = word_tokenize(text)

        return text_tokens

    tokens = prep_for_lemma(text)

    def lem_word(word):
    
        if len(word) > 3: out_word = wnl.lemmatize(word)
        else: out_word = word

        return out_word

    return [lem_word(token) for token in tokens]

def keybert_keywords(text, n, kw_model):
    tokens_lemma = apply_lemmatize(text)
    lemmatised_text = ' '.join(tokens_lemma)

    keywords_text = keybert.KeyBERT(model=kw_model).extract_keywords(lemmatised_text, stop_words='english', top_n=n, 
                                                   keyphrase_ngram_range=(1, 1))
    keywords_list = [item[0] for item in keywords_text]

    return keywords_list