Spaces:
Sleeping
Sleeping
all done
Browse files
app.py
CHANGED
@@ -22,15 +22,6 @@ import tqdm
|
|
22 |
import accelerate
|
23 |
import re
|
24 |
|
25 |
-
|
26 |
-
import torch
|
27 |
-
from sacrebleu import corpus_bleu
|
28 |
-
from rouge_score import rouge_scorer
|
29 |
-
from bert_score import score
|
30 |
-
from transformers import GPT2LMHeadModel, GPT2Tokenizer, pipeline
|
31 |
-
import nltk
|
32 |
-
from nltk.util import ngrams
|
33 |
-
|
34 |
api_key = os.getenv('API_KEY')
|
35 |
|
36 |
|
@@ -87,6 +78,25 @@ def load_db():
|
|
87 |
# Initialize langchain LLM chain
|
88 |
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
89 |
progress(0.1, desc="Initializing HF tokenizer...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
|
91 |
# HuggingFaceHub uses HF inference endpoints
|
92 |
progress(0.5, desc="Initializing HF Hub...")
|
@@ -237,138 +247,32 @@ def format_chat_history(message, chat_history):
|
|
237 |
formatted_chat_history.append(f"User: {user_message}")
|
238 |
formatted_chat_history.append(f"Assistant: {bot_message}")
|
239 |
return formatted_chat_history
|
|
|
240 |
|
241 |
-
|
242 |
-
def load_gpt2_model():
|
243 |
-
model = GPT2LMHeadModel.from_pretrained('gpt2')
|
244 |
-
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
|
245 |
-
return model, tokenizer
|
246 |
-
|
247 |
-
gpt2_model, gpt2_tokenizer = load_gpt2_model()
|
248 |
-
bias_pipeline = pipeline("zero-shot-classification", model="Hate-speech-CNERG/dehatebert-mono-english")
|
249 |
-
|
250 |
-
def evaluate_bleu_rouge(candidates, references):
|
251 |
-
bleu_score = corpus_bleu(candidates, [references]).score
|
252 |
-
scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)
|
253 |
-
rouge_scores = [scorer.score(ref, cand) for ref, cand in zip(references, candidates)]
|
254 |
-
rouge1 = sum([score['rouge1'].fmeasure for score in rouge_scores]) / len(rouge_scores)
|
255 |
-
return bleu_score, rouge1
|
256 |
-
|
257 |
-
def evaluate_bert_score(candidates, references):
|
258 |
-
P, R, F1 = score(candidates, references, lang="en", model_type='bert-base-multilingual-cased')
|
259 |
-
return P.mean().item(), R.mean().item(), F1.mean().item()
|
260 |
-
|
261 |
-
def evaluate_perplexity(text, model, tokenizer):
|
262 |
-
encodings = tokenizer(text, return_tensors='pt')
|
263 |
-
max_length = model.config.n_positions
|
264 |
-
stride = 512
|
265 |
-
lls = []
|
266 |
-
for i in range(0, encodings.input_ids.size(1), stride):
|
267 |
-
begin_loc = max(i + stride - max_length, 0)
|
268 |
-
end_loc = min(i + stride, encodings.input_ids.size(1))
|
269 |
-
trg_len = end_loc - i
|
270 |
-
input_ids = encodings.input_ids[:, begin_loc:end_loc]
|
271 |
-
target_ids = input_ids.clone()
|
272 |
-
target_ids[:, :-trg_len] = -100
|
273 |
-
with torch.no_grad():
|
274 |
-
outputs = model(input_ids, labels=target_ids)
|
275 |
-
log_likelihood = outputs[0] * trg_len
|
276 |
-
lls.append(log_likelihood)
|
277 |
-
ppl = torch.exp(torch.stack(lls).sum() / end_loc)
|
278 |
-
return ppl.item()
|
279 |
-
|
280 |
-
def evaluate_diversity(texts):
|
281 |
-
all_tokens = [tok for text in texts for tok in text.split()]
|
282 |
-
unique_bigrams = set(ngrams(all_tokens, 2))
|
283 |
-
diversity_score = len(unique_bigrams) / len(all_tokens) if all_tokens else 0
|
284 |
-
return diversity_score
|
285 |
-
|
286 |
-
def evaluate_racial_bias(text, pipeline):
|
287 |
-
results = pipeline([text], candidate_labels=["hate speech", "not hate speech"])
|
288 |
-
bias_score = results[0]['scores'][results[0]['labels'].index('hate speech')]
|
289 |
-
return bias_score
|
290 |
-
|
291 |
-
def evaluate_all(question, response, reference, gpt2_model, gpt2_tokenizer, bias_pipeline):
|
292 |
-
candidates = [response]
|
293 |
-
references = [reference]
|
294 |
-
bleu, rouge1 = evaluate_bleu_rouge(candidates, references)
|
295 |
-
bert_p, bert_r, bert_f1 = evaluate_bert_score(candidates, references)
|
296 |
-
perplexity = evaluate_perplexity(response, gpt2_model, gpt2_tokenizer)
|
297 |
-
diversity = evaluate_diversity(candidates)
|
298 |
-
racial_bias = evaluate_racial_bias(response, bias_pipeline)
|
299 |
-
return {
|
300 |
-
"BLEU": bleu,
|
301 |
-
"ROUGE-1": rouge1,
|
302 |
-
"BERT P": bert_p,
|
303 |
-
"BERT R": bert_r,
|
304 |
-
"BERT F1": bert_f1,
|
305 |
-
"Perplexity": perplexity,
|
306 |
-
"Diversity": diversity,
|
307 |
-
"Racial Bias": racial_bias
|
308 |
-
}
|
309 |
-
|
310 |
-
#---------------------------------------------------------------------------------
|
311 |
-
|
312 |
-
def display_metrics(metrics):
|
313 |
-
result = ""
|
314 |
-
for k, v in metrics.items():
|
315 |
-
if k == 'BLEU':
|
316 |
-
result += f"BLEU measures the overlap between the generated output and reference text based on n-grams. Higher scores indicate better match. Score obtained: {v}\n\n"
|
317 |
-
elif k == "ROUGE-1":
|
318 |
-
result += f"ROUGE-1 measures the overlap of unigrams between the generated output and reference text. Higher scores indicate better match. Score obtained: {v}\n\n"
|
319 |
-
elif k == 'BERT P':
|
320 |
-
result += "BERTScore evaluates the semantic similarity between the generated output and reference text using BERT embeddings.\n\n"
|
321 |
-
result += f"**BERT Precision**: {metrics['BERT P']}\n"
|
322 |
-
result += f"**BERT Recall**: {metrics['BERT R']}\n"
|
323 |
-
result += f"**BERT F1 Score**: {metrics['BERT F1']}\n\n"
|
324 |
-
elif k == 'Perplexity':
|
325 |
-
result += f"Perplexity measures how well a language model predicts the text. Lower values indicate better fluency and coherence. Score obtained: {v}\n\n"
|
326 |
-
elif k == 'Diversity':
|
327 |
-
result += f"Diversity measures the uniqueness of bigrams in the generated output. Higher values indicate more diverse and varied output. Score obtained: {v}\n\n"
|
328 |
-
elif k == 'Racial Bias':
|
329 |
-
result += f"Racial Bias score indicates the presence of biased language in the generated output. Higher scores indicate more bias. Score obtained: {v}\n\n"
|
330 |
-
return result
|
331 |
-
#---------------------------------------------------------------------------------------------------------------------------------------------------
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
def conversation(qa_chain, message, history, gpt2_model, gpt2_tokenizer, bias_pipeline):
|
339 |
formatted_chat_history = format_chat_history(message, history)
|
340 |
-
|
341 |
-
|
|
|
342 |
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
343 |
response_answer = response["answer"]
|
344 |
-
answer_of_question = response["answer"]
|
345 |
if response_answer.find("Helpful Answer:") != -1:
|
346 |
response_answer = response_answer.split("Helpful Answer:")[-1]
|
347 |
response_sources = response["source_documents"]
|
348 |
-
context = " ".join([d.page_content for d in response_sources])
|
349 |
-
|
350 |
response_source1 = response_sources[0].page_content.strip()
|
351 |
response_source2 = response_sources[1].page_content.strip()
|
352 |
response_source3 = response_sources[2].page_content.strip()
|
353 |
-
|
354 |
response_source1_page = response_sources[0].metadata["page"] + 1
|
355 |
response_source2_page = response_sources[1].metadata["page"] + 1
|
356 |
response_source3_page = response_sources[2].metadata["page"] + 1
|
357 |
-
|
358 |
-
|
359 |
|
360 |
-
#
|
361 |
-
|
362 |
-
|
363 |
-
|
364 |
-
return (qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page,
|
365 |
-
response_source2, response_source2_page, response_source3, response_source3_page,
|
366 |
-
evaluation_metrics)
|
367 |
-
|
368 |
-
|
369 |
-
# def interact(qa_chain, message, history):
|
370 |
-
# return conversation(qa_chain, message, history, evaluator)
|
371 |
-
|
372 |
|
373 |
|
374 |
def upload_file(file_obj):
|
@@ -380,21 +284,19 @@ def upload_file(file_obj):
|
|
380 |
# initialize_database(file_path, progress)
|
381 |
return list_file_path
|
382 |
|
383 |
-
####################################
|
384 |
|
385 |
def demo():
|
386 |
with gr.Blocks(theme="base") as demo:
|
387 |
vector_db = gr.State()
|
388 |
qa_chain = gr.State()
|
389 |
collection_name = gr.State()
|
390 |
-
history = gr.State([]) # Initialize history as an empty list
|
391 |
|
392 |
gr.Markdown(
|
393 |
"""<center><h2>PDF-based chatbot</center></h2>
|
394 |
<h3>Ask any questions about your PDF documents</h3>""")
|
395 |
gr.Markdown(
|
396 |
"""<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
|
397 |
-
The user interface
|
398 |
This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
|
399 |
<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 a reply.
|
400 |
""")
|
@@ -402,15 +304,16 @@ def demo():
|
|
402 |
with gr.Tab("Step 1 - Upload PDF"):
|
403 |
with gr.Row():
|
404 |
document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
|
|
|
405 |
|
406 |
with gr.Tab("Step 2 - Process document"):
|
407 |
with gr.Row():
|
408 |
-
db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value="ChromaDB", type="index", info="Choose your vector database")
|
409 |
with gr.Accordion("Advanced options - Document text splitter", open=False):
|
410 |
with gr.Row():
|
411 |
-
slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
|
412 |
with gr.Row():
|
413 |
-
slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
414 |
with gr.Row():
|
415 |
db_progress = gr.Textbox(label="Vector database initialization", value="None")
|
416 |
with gr.Row():
|
@@ -418,16 +321,17 @@ def demo():
|
|
418 |
|
419 |
with gr.Tab("Step 3 - Initialize QA chain"):
|
420 |
with gr.Row():
|
421 |
-
llm_btn = gr.Radio(list_llm_simple,
|
|
|
422 |
with gr.Accordion("Advanced options - LLM model", open=False):
|
423 |
with gr.Row():
|
424 |
-
slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
|
425 |
with gr.Row():
|
426 |
-
slider_maxtokens = gr.Slider(minimum=224, maximum=4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
|
427 |
with gr.Row():
|
428 |
-
slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
|
429 |
with gr.Row():
|
430 |
-
llm_progress = gr.Textbox(value="None",
|
431 |
with gr.Row():
|
432 |
qachain_btn = gr.Button("Initialize Question Answering chain")
|
433 |
|
@@ -448,153 +352,34 @@ def demo():
|
|
448 |
with gr.Row():
|
449 |
submit_btn = gr.Button("Submit message")
|
450 |
clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
|
451 |
-
|
452 |
-
metrics_output = gr.Textbox(lines=10, label="Evaluation Metrics")
|
453 |
-
|
454 |
# Preprocessing events
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
qachain_btn.click(initialize_LLM,
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
|
466 |
# Chatbot events
|
467 |
msg.submit(conversation, \
|
468 |
-
inputs=[qa_chain, msg, chatbot
|
469 |
-
outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
470 |
queue=False)
|
471 |
-
|
472 |
-
submit_btn.click(conversation,
|
473 |
-
inputs=[qa_chain, msg, history, gpt2_model, gpt2_tokenizer, bias_pipeline],
|
474 |
-
outputs=[qa_chain, chatbot, history, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page, metrics_output])
|
475 |
-
|
476 |
-
clear_btn.click(lambda: [None, "", 0, "", 0, "", 0],
|
477 |
-
inputs=None,
|
478 |
-
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page],
|
479 |
-
queue=False)
|
480 |
-
|
481 |
demo.queue().launch(debug=True)
|
482 |
|
483 |
-
if __name__ == "__main__":
|
484 |
-
demo()
|
485 |
-
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
|
490 |
-
|
491 |
-
|
492 |
-
# with gr.Blocks(theme="base") as demo:
|
493 |
-
# vector_db = gr.State()
|
494 |
-
# qa_chain = gr.State()
|
495 |
-
# collection_name = gr.State()
|
496 |
-
# history = gr.State()
|
497 |
-
|
498 |
-
# gr.Markdown(
|
499 |
-
# """<center><h2>PDF-based chatbot</center></h2>
|
500 |
-
# <h3>Ask any questions about your PDF documents</h3>""")
|
501 |
-
# gr.Markdown(
|
502 |
-
# """<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
|
503 |
-
# The user interface explicitely shows multiple steps to help understand the RAG workflow.
|
504 |
-
# This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
|
505 |
-
# <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 a reply.
|
506 |
-
# """)
|
507 |
-
|
508 |
-
# with gr.Tab("Step 1 - Upload PDF"):
|
509 |
-
# with gr.Row():
|
510 |
-
# document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
|
511 |
-
# # upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
|
512 |
-
|
513 |
-
# with gr.Tab("Step 2 - Process document"):
|
514 |
-
# with gr.Row():
|
515 |
-
# db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
|
516 |
-
# with gr.Accordion("Advanced options - Document text splitter", open=False):
|
517 |
-
# with gr.Row():
|
518 |
-
# slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
|
519 |
-
# with gr.Row():
|
520 |
-
# slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
521 |
-
# with gr.Row():
|
522 |
-
# db_progress = gr.Textbox(label="Vector database initialization", value="None")
|
523 |
-
# with gr.Row():
|
524 |
-
# db_btn = gr.Button("Generate vector database")
|
525 |
-
|
526 |
-
# with gr.Tab("Step 3 - Initialize QA chain"):
|
527 |
-
# with gr.Row():
|
528 |
-
# llm_btn = gr.Radio(list_llm_simple, \
|
529 |
-
# label="LLM models", value = list_llm_simple[0], type="index", info="Choose your LLM model")
|
530 |
-
# with gr.Accordion("Advanced options - LLM model", open=False):
|
531 |
-
# with gr.Row():
|
532 |
-
# slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
|
533 |
-
# with gr.Row():
|
534 |
-
# slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
|
535 |
-
# with gr.Row():
|
536 |
-
# slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
|
537 |
-
# with gr.Row():
|
538 |
-
# llm_progress = gr.Textbox(value="None",label="QA chain initialization")
|
539 |
-
# with gr.Row():
|
540 |
-
# qachain_btn = gr.Button("Initialize Question Answering chain")
|
541 |
-
|
542 |
-
# with gr.Tab("Step 4 - Chatbot"):
|
543 |
-
# chatbot = gr.Chatbot(height=300)
|
544 |
-
# with gr.Accordion("Advanced - Document references", open=False):
|
545 |
-
# with gr.Row():
|
546 |
-
# doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
|
547 |
-
# source1_page = gr.Number(label="Page", scale=1)
|
548 |
-
# with gr.Row():
|
549 |
-
# doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
|
550 |
-
# source2_page = gr.Number(label="Page", scale=1)
|
551 |
-
# with gr.Row():
|
552 |
-
# doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
|
553 |
-
# source3_page = gr.Number(label="Page", scale=1)
|
554 |
-
# with gr.Row():
|
555 |
-
# msg = gr.Textbox(placeholder="Type message (e.g. 'What is this document about?')", container=True)
|
556 |
-
# with gr.Row():
|
557 |
-
# submit_btn = gr.Button("Submit message")
|
558 |
-
# clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
|
559 |
-
# with gr.Row("Metrics"):
|
560 |
-
# metrics_output = gr.Textbox(lines=10, label="Evaluation Metrics")
|
561 |
-
|
562 |
-
|
563 |
-
# # Preprocessing events
|
564 |
-
# #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
|
565 |
-
# db_btn.click(initialize_database, \
|
566 |
-
# inputs=[document, slider_chunk_size, slider_chunk_overlap], \
|
567 |
-
# outputs=[vector_db, collection_name, db_progress])
|
568 |
-
# qachain_btn.click(initialize_LLM, \
|
569 |
-
# inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
|
570 |
-
# outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
|
571 |
-
# inputs=None, \
|
572 |
-
# outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
573 |
-
# queue=False)
|
574 |
-
|
575 |
-
# Chatbot events
|
576 |
-
# msg.submit(interact, inputs=[gr.State(),qa_chain, msg, history], outputs=[
|
577 |
-
# gr.State(), chatbot, history, response_source1, response_source1_page,
|
578 |
-
# response_source2, response_source2_page, response_source3, response_source3_page,
|
579 |
-
# None, None, None, metrics_output
|
580 |
-
# ],queue=False)
|
581 |
-
|
582 |
-
|
583 |
-
# submit_btn.click(conversation, \
|
584 |
-
# inputs=[qa_chain, msg, chatbot], \
|
585 |
-
# outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
586 |
-
# queue=False)
|
587 |
-
# clear_btn.click(lambda:[None,"",0,"",0,"",0], \
|
588 |
-
# inputs=None, \
|
589 |
-
# outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
590 |
-
# queue=False)
|
591 |
-
|
592 |
-
|
593 |
-
|
594 |
-
|
595 |
-
|
596 |
-
# demo.queue().launch(debug=True)
|
597 |
-
|
598 |
-
|
599 |
-
# if __name__ == "__main__":
|
600 |
-
# demo()
|
|
|
22 |
import accelerate
|
23 |
import re
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
api_key = os.getenv('API_KEY')
|
26 |
|
27 |
|
|
|
78 |
# Initialize langchain LLM chain
|
79 |
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
80 |
progress(0.1, desc="Initializing HF tokenizer...")
|
81 |
+
# HuggingFacePipeline uses local model
|
82 |
+
# Note: it will download model locally...
|
83 |
+
# tokenizer=AutoTokenizer.from_pretrained(llm_model)
|
84 |
+
# progress(0.5, desc="Initializing HF pipeline...")
|
85 |
+
# pipeline=transformers.pipeline(
|
86 |
+
# "text-generation",
|
87 |
+
# model=llm_model,
|
88 |
+
# tokenizer=tokenizer,
|
89 |
+
# torch_dtype=torch.bfloat16,
|
90 |
+
# trust_remote_code=True,
|
91 |
+
# device_map="auto",
|
92 |
+
# # max_length=1024,
|
93 |
+
# max_new_tokens=max_tokens,
|
94 |
+
# do_sample=True,
|
95 |
+
# top_k=top_k,
|
96 |
+
# num_return_sequences=1,
|
97 |
+
# eos_token_id=tokenizer.eos_token_id
|
98 |
+
# )
|
99 |
+
# llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
|
100 |
|
101 |
# HuggingFaceHub uses HF inference endpoints
|
102 |
progress(0.5, desc="Initializing HF Hub...")
|
|
|
247 |
formatted_chat_history.append(f"User: {user_message}")
|
248 |
formatted_chat_history.append(f"Assistant: {bot_message}")
|
249 |
return formatted_chat_history
|
250 |
+
|
251 |
|
252 |
+
def conversation(qa_chain, message, history):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
formatted_chat_history = format_chat_history(message, history)
|
254 |
+
#print("formatted_chat_history",formatted_chat_history)
|
255 |
+
|
256 |
+
# Generate response using QA chain
|
257 |
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
258 |
response_answer = response["answer"]
|
|
|
259 |
if response_answer.find("Helpful Answer:") != -1:
|
260 |
response_answer = response_answer.split("Helpful Answer:")[-1]
|
261 |
response_sources = response["source_documents"]
|
|
|
|
|
262 |
response_source1 = response_sources[0].page_content.strip()
|
263 |
response_source2 = response_sources[1].page_content.strip()
|
264 |
response_source3 = response_sources[2].page_content.strip()
|
265 |
+
# Langchain sources are zero-based
|
266 |
response_source1_page = response_sources[0].metadata["page"] + 1
|
267 |
response_source2_page = response_sources[1].metadata["page"] + 1
|
268 |
response_source3_page = response_sources[2].metadata["page"] + 1
|
269 |
+
# print ('chat response: ', response_answer)
|
270 |
+
# print('DB source', response_sources)
|
271 |
|
272 |
+
# Append user message and response to chat history
|
273 |
+
new_history = history + [(message, response_answer)]
|
274 |
+
# return gr.update(value=""), new_history, response_sources[0], response_sources[1]
|
275 |
+
return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
276 |
|
277 |
|
278 |
def upload_file(file_obj):
|
|
|
284 |
# initialize_database(file_path, progress)
|
285 |
return list_file_path
|
286 |
|
|
|
287 |
|
288 |
def demo():
|
289 |
with gr.Blocks(theme="base") as demo:
|
290 |
vector_db = gr.State()
|
291 |
qa_chain = gr.State()
|
292 |
collection_name = gr.State()
|
|
|
293 |
|
294 |
gr.Markdown(
|
295 |
"""<center><h2>PDF-based chatbot</center></h2>
|
296 |
<h3>Ask any questions about your PDF documents</h3>""")
|
297 |
gr.Markdown(
|
298 |
"""<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
|
299 |
+
The user interface explicitely shows multiple steps to help understand the RAG workflow.
|
300 |
This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
|
301 |
<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 a reply.
|
302 |
""")
|
|
|
304 |
with gr.Tab("Step 1 - Upload PDF"):
|
305 |
with gr.Row():
|
306 |
document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
|
307 |
+
# upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
|
308 |
|
309 |
with gr.Tab("Step 2 - Process document"):
|
310 |
with gr.Row():
|
311 |
+
db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
|
312 |
with gr.Accordion("Advanced options - Document text splitter", open=False):
|
313 |
with gr.Row():
|
314 |
+
slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
|
315 |
with gr.Row():
|
316 |
+
slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
317 |
with gr.Row():
|
318 |
db_progress = gr.Textbox(label="Vector database initialization", value="None")
|
319 |
with gr.Row():
|
|
|
321 |
|
322 |
with gr.Tab("Step 3 - Initialize QA chain"):
|
323 |
with gr.Row():
|
324 |
+
llm_btn = gr.Radio(list_llm_simple, \
|
325 |
+
label="LLM models", value = list_llm_simple[0], type="index", info="Choose your LLM model")
|
326 |
with gr.Accordion("Advanced options - LLM model", open=False):
|
327 |
with gr.Row():
|
328 |
+
slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
|
329 |
with gr.Row():
|
330 |
+
slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
|
331 |
with gr.Row():
|
332 |
+
slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
|
333 |
with gr.Row():
|
334 |
+
llm_progress = gr.Textbox(value="None",label="QA chain initialization")
|
335 |
with gr.Row():
|
336 |
qachain_btn = gr.Button("Initialize Question Answering chain")
|
337 |
|
|
|
352 |
with gr.Row():
|
353 |
submit_btn = gr.Button("Submit message")
|
354 |
clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
|
355 |
+
|
|
|
|
|
356 |
# Preprocessing events
|
357 |
+
#upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
|
358 |
+
db_btn.click(initialize_database, \
|
359 |
+
inputs=[document, slider_chunk_size, slider_chunk_overlap], \
|
360 |
+
outputs=[vector_db, collection_name, db_progress])
|
361 |
+
qachain_btn.click(initialize_LLM, \
|
362 |
+
inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
|
363 |
+
outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
|
364 |
+
inputs=None, \
|
365 |
+
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
366 |
+
queue=False)
|
367 |
|
368 |
# Chatbot events
|
369 |
msg.submit(conversation, \
|
370 |
+
inputs=[qa_chain, msg, chatbot], \
|
371 |
+
outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
372 |
+
queue=False)
|
373 |
+
submit_btn.click(conversation, \
|
374 |
+
inputs=[qa_chain, msg, chatbot], \
|
375 |
+
outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
376 |
+
queue=False)
|
377 |
+
clear_btn.click(lambda:[None,"",0,"",0,"",0], \
|
378 |
+
inputs=None, \
|
379 |
+
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
|
380 |
queue=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
381 |
demo.queue().launch(debug=True)
|
382 |
|
|
|
|
|
|
|
|
|
|
|
|
|
383 |
|
384 |
+
if __name__ == "__main__":
|
385 |
+
demo()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|