import requests import gradio as gr from ragatouille import RAGPretrainedModel import logging from pathlib import Path from time import perf_counter from sentence_transformers import CrossEncoder from huggingface_hub import InferenceClient from jinja2 import Environment, FileSystemLoader import numpy as np from os import getenv from backend.query_llm import generate_hf, generate_qwen from backend.semantic_search import table, retriever from huggingface_hub import InferenceClient # Bhashini API translation function api_key = getenv('API_KEY') user_id = getenv('USER_ID') def bhashini_translate(text: str, from_code: str = "en", to_code: str = "hi") -> dict: """Translates text from source language to target language using the Bhashini API.""" if not text.strip(): print('Input text is empty. Please provide valid text for translation.') return {"status_code": 400, "message": "Input text is empty", "translated_content": None, "speech_content": None} else: print('Input text - ',text) print(f'Starting translation process from {from_code} to {to_code}...') print(f'Starting translation process from {from_code} to {to_code}...') gr.Warning(f'Translating to {to_code}...') url = 'https://meity-auth.ulcacontrib.org/ulca/apis/v0/model/getModelsPipeline' headers = { "Content-Type": "application/json", "userID": user_id, "ulcaApiKey": api_key } payload = { "pipelineTasks": [{"taskType": "translation", "config": {"language": {"sourceLanguage": from_code, "targetLanguage": to_code}}}], "pipelineRequestConfig": {"pipelineId": "64392f96daac500b55c543cd"} } print('Sending initial request to get the pipeline...') response = requests.post(url, json=payload, headers=headers) if response.status_code != 200: print(f'Error in initial request: {response.status_code}') return {"status_code": response.status_code, "message": "Error in translation request", "translated_content": None} print('Initial request successful, processing response...') response_data = response.json() service_id = response_data["pipelineResponseConfig"][0]["config"][0]["serviceId"] callback_url = response_data["pipelineInferenceAPIEndPoint"]["callbackUrl"] print(f'Service ID: {service_id}, Callback URL: {callback_url}') headers2 = { "Content-Type": "application/json", response_data["pipelineInferenceAPIEndPoint"]["inferenceApiKey"]["name"]: response_data["pipelineInferenceAPIEndPoint"]["inferenceApiKey"]["value"] } compute_payload = { "pipelineTasks": [{"taskType": "translation", "config": {"language": {"sourceLanguage": from_code, "targetLanguage": to_code}, "serviceId": service_id}}], "inputData": {"input": [{"source": text}], "audio": [{"audioContent": None}]} } print(f'Sending translation request with text: "{text}"') compute_response = requests.post(callback_url, json=compute_payload, headers=headers2) if compute_response.status_code != 200: print(f'Error in translation request: {compute_response.status_code}') return {"status_code": compute_response.status_code, "message": "Error in translation", "translated_content": None} print('Translation request successful, processing translation...') compute_response_data = compute_response.json() translated_content = compute_response_data["pipelineResponse"][0]["output"][0]["target"] print(f'Translation successful. Translated content: "{translated_content}"') return {"status_code": 200, "message": "Translation successful", "translated_content": translated_content} # Existing chatbot functions VECTOR_COLUMN_NAME = "vector" TEXT_COLUMN_NAME = "text" HF_TOKEN = getenv("HUGGING_FACE_HUB_TOKEN") proj_dir = Path(__file__).parent logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1", token=HF_TOKEN) env = Environment(loader=FileSystemLoader(proj_dir / 'templates')) template = env.get_template('template.j2') template_html = env.get_template('template_html.j2') def add_text(history, text): history = [] if history is None else history history = history + [(text, None)] return history, gr.Textbox(value="", interactive=False) def bot(history, cross_encoder): top_rerank = 25 top_k_rank = 20 query = history[-1][0] if history else '' if not query: gr.Warning("Please submit a non-empty string as a prompt") raise ValueError("Empty string was submitted") logger.warning('Retrieving documents...') if cross_encoder == '(HIGH ACCURATE) ColBERT': gr.Warning('Retrieving using ColBERT.. First time query will take a minute for model to load..pls wait') RAG = RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0") RAG_db = RAG.from_index('.ragatouille/colbert/indexes/cbseclass10index') documents_full = RAG_db.search(query, k=top_k_rank) documents = [item['content'] for item in documents_full] prompt = template.render(documents=documents, query=query) prompt_html = template_html.render(documents=documents, query=query) generate_fn = generate_hf history[-1][1] = "" for character in generate_fn(prompt, history[:-1]): history[-1][1] = character yield history, prompt_html else: document_start = perf_counter() query_vec = retriever.encode(query) doc1 = table.search(query_vec, vector_column_name=VECTOR_COLUMN_NAME).limit(top_k_rank) documents = table.search(query_vec, vector_column_name=VECTOR_COLUMN_NAME).limit(top_rerank).to_list() documents = [doc[TEXT_COLUMN_NAME] for doc in documents] query_doc_pair = [[query, doc] for doc in documents] if cross_encoder == '(FAST) MiniLM-L6v2': cross_encoder1 = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2') elif cross_encoder == '(ACCURATE) BGE reranker': cross_encoder1 = CrossEncoder('BAAI/bge-reranker-base') cross_scores = cross_encoder1.predict(query_doc_pair) sim_scores_argsort = list(reversed(np.argsort(cross_scores))) documents = [documents[idx] for idx in sim_scores_argsort[:top_k_rank]] document_time = perf_counter() - document_start prompt = template.render(documents=documents, query=query) prompt_html = template_html.render(documents=documents, query=query) #generate_fn = generate_hf generate_fn=generate_qwen # Create a new history entry instead of modifying the tuple directly new_history = history[:-1] + [ (query, "") ] for character in generate_fn(prompt, history[:-1]): new_history[-1] = (query, character) # Update the last tuple with new text yield new_history, prompt_html # history[-1][1] = "" # for character in generate_fn(prompt, history[:-1]): # history[-1][1] = character # yield history, prompt_html #def translate_text(response_text, selected_language): def translate_text(selected_language,history): iso_language_codes = { "Hindi": "hi", "Gom": "gom", "Kannada": "kn", "Dogri": "doi", "Bodo": "brx", "Urdu": "ur", "Tamil": "ta", "Kashmiri": "ks", "Assamese": "as", "Bengali": "bn", "Marathi": "mr", "Sindhi": "sd", "Maithili": "mai", "Punjabi": "pa", "Malayalam": "ml", "Manipuri": "mni", "Telugu": "te", "Sanskrit": "sa", "Nepali": "ne", "Santali": "sat", "Gujarati": "gu", "Odia": "or" } to_code = iso_language_codes[selected_language] response_text = history[-1][1] if history else '' translation = bhashini_translate(response_text, to_code=to_code) return translation['translated_content'] # iso_language_codes = { # "Hindi": "hi", # "Gom": "gom", # "Kannada": "kn", # "Dogri": "doi", # "Bodo": "brx", # "Urdu": "ur", # "Tamil": "ta", # "Kashmiri": "ks", # "Assamese": "as", # "Bengali": "bn", # "Marathi": "mr", # "Sindhi": "sd", # "Maithili": "mai", # "Punjabi": "pa", # "Malayalam": "ml", # "Manipuri": "mni", # "Telugu": "te", # "Sanskrit": "sa", # "Nepali": "ne", # "Santali": "sat", # "Gujarati": "gu", # "Odia": "or" # } # to_code = iso_language_codes[selected_language] # translation = bhashini_translate(response_text, to_code=to_code) # return translation['translated_content'] # Gradio interface with gr.Blocks(theme='gradio/soft') as CHATBOT: history_state = gr.State([]) with gr.Row(): with gr.Column(scale=10): gr.HTML(value="""

ADWITIYA-

Custom Manual Chatbot and Quizbot

""") gr.HTML(value=f"""

Using GenAI for CBIC Capacity Building - A free chat bot developed by National Customs Targeting Center using Open source LLMs for CBIC Officers

""") gr.HTML(value=f"""

Developed by NCTC,Mumbai. Suggestions may be sent to nctc-admin@gov.in.

""") with gr.Column(scale=3): gr.Image(value='logo.png', height=200, width=200) chatbot = gr.Chatbot( [], elem_id="chatbot", avatar_images=('https://aui.atlassian.com/aui/8.8/docs/images/avatar-person.svg', 'https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg'), bubble_full_width=False, show_copy_button=True, show_share_button=True, ) with gr.Row(): txt = gr.Textbox( scale=3, show_label=False, placeholder="Enter text and press enter", container=False, ) txt_btn = gr.Button(value="Submit text", scale=1) cross_encoder = gr.Radio(choices=['(FAST) MiniLM-L6v2', '(ACCURATE) BGE reranker', '(HIGH ACCURATE) ColBERT'], value='(ACCURATE) BGE reranker', label="Embeddings", info="Only First query to Colbert may take little time)") language_dropdown = gr.Dropdown( choices=[ "Hindi", "Gom", "Kannada", "Dogri", "Bodo", "Urdu", "Tamil", "Kashmiri", "Assamese", "Bengali", "Marathi", "Sindhi", "Maithili", "Punjabi", "Malayalam", "Manipuri", "Telugu", "Sanskrit", "Nepali", "Santali", "Gujarati", "Odia" ], value="Hindi", # default to Hindi label="Select Language for Translation" ) prompt_html = gr.HTML() translated_textbox = gr.Textbox(label="Translated Response") def update_history_and_translate(txt, cross_encoder, history_state, language_dropdown): history = history_state history.append((txt, "")) #history_state.value=(history) # Call bot function bot_output = list(bot(history, cross_encoder)) history, prompt_html = bot_output[-1] # Update the history state history_state[:] = history # Translate text translated_text = translate_text(language_dropdown, history) return history, prompt_html, translated_text txt_msg = txt_btn.click(update_history_and_translate, [txt, cross_encoder, history_state, language_dropdown], [chatbot, prompt_html, translated_textbox]) txt_msg = txt.submit(update_history_and_translate, [txt, cross_encoder, history_state, language_dropdown], [chatbot, prompt_html, translated_textbox]) examples = ['My transhipment cargo is missing','can u explain and tabulate difference between b 17 bond and a warehousing bond', 'What are benefits of the AEO Scheme and eligibility criteria?', 'What are penalties for customs offences? ', 'what are penalties to customs officers misusing their powers under customs act?','What are eligibility criteria for exemption from cost recovery charges','list in detail what is procedure for obtaining new approval for openeing a CFS attached to an ICD'] gr.Examples(examples, txt) # txt_msg = txt_btn.click(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( # bot, [chatbot, cross_encoder], [chatbot, prompt_html]).then( # translate_text, [txt, language_dropdown], translated_textbox # ) # txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( # bot, [chatbot, cross_encoder], [chatbot, prompt_html]).then( # translate_text, [txt, language_dropdown], translated_textbox # ) # Launch the Gradio application CHATBOT.launch(share=True) # from ragatouille import RAGPretrainedModel # import subprocess # import json # import spaces # import firebase_admin # from firebase_admin import credentials, firestore # import logging # from pathlib import Path # from time import perf_counter # from datetime import datetime # import gradio as gr # from jinja2 import Environment, FileSystemLoader # import numpy as np # from sentence_transformers import CrossEncoder # from huggingface_hub import InferenceClient # from os import getenv # from backend.query_llm import generate_hf, generate_openai # from backend.semantic_search import table, retriever # from huggingface_hub import InferenceClient # VECTOR_COLUMN_NAME = "vector" # TEXT_COLUMN_NAME = "text" # HF_TOKEN = getenv("HUGGING_FACE_HUB_TOKEN") # proj_dir = Path(__file__).parent # # Setting up the logging # logging.basicConfig(level=logging.INFO) # logger = logging.getLogger(__name__) # client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1",token=HF_TOKEN) # # Set up the template environment with the templates directory # env = Environment(loader=FileSystemLoader(proj_dir / 'templates')) # # Load the templates directly from the environment # template = env.get_template('template.j2') # template_html = env.get_template('template_html.j2') # def add_text(history, text): # history = [] if history is None else history # history = history + [(text, None)] # return history, gr.Textbox(value="", interactive=False) # def bot(history, cross_encoder): # top_rerank = 25 # top_k_rank = 20 # query = history[-1][0] # if not query: # gr.Warning("Please submit a non-empty string as a prompt") # raise ValueError("Empty string was submitted") # logger.warning('Retrieving documents...') # # if COLBERT RAGATATOUILLE PROCEDURE : # if cross_encoder=='(HIGH ACCURATE) ColBERT': # gr.Warning('Retrieving using ColBERT.. First time query will take a minute for model to load..pls wait') # RAG= RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0") # RAG_db=RAG.from_index('.ragatouille/colbert/indexes/cbseclass10index') # documents_full=RAG_db.search(query,k=top_k_rank) # documents=[item['content'] for item in documents_full] # # Create Prompt # prompt = template.render(documents=documents, query=query) # prompt_html = template_html.render(documents=documents, query=query) # generate_fn = generate_hf # history[-1][1] = "" # for character in generate_fn(prompt, history[:-1]): # history[-1][1] = character # yield history, prompt_html # print('Final history is ',history) # #store_message(db,history[-1][0],history[-1][1],cross_encoder) # else: # # Retrieve documents relevant to query # document_start = perf_counter() # query_vec = retriever.encode(query) # logger.warning(f'Finished query vec') # doc1 = table.search(query_vec, vector_column_name=VECTOR_COLUMN_NAME).limit(top_k_rank) # logger.warning(f'Finished search') # documents = table.search(query_vec, vector_column_name=VECTOR_COLUMN_NAME).limit(top_rerank).to_list() # documents = [doc[TEXT_COLUMN_NAME] for doc in documents] # logger.warning(f'start cross encoder {len(documents)}') # # Retrieve documents relevant to query # query_doc_pair = [[query, doc] for doc in documents] # if cross_encoder=='(FAST) MiniLM-L6v2' : # cross_encoder1 = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2') # elif cross_encoder=='(ACCURATE) BGE reranker': # cross_encoder1 = CrossEncoder('BAAI/bge-reranker-base') # cross_scores = cross_encoder1.predict(query_doc_pair) # sim_scores_argsort = list(reversed(np.argsort(cross_scores))) # logger.warning(f'Finished cross encoder {len(documents)}') # documents = [documents[idx] for idx in sim_scores_argsort[:top_k_rank]] # logger.warning(f'num documents {len(documents)}') # document_time = perf_counter() - document_start # logger.warning(f'Finished Retrieving documents in {round(document_time, 2)} seconds...') # # Create Prompt # prompt = template.render(documents=documents, query=query) # prompt_html = template_html.render(documents=documents, query=query) # generate_fn = generate_hf # history[-1][1] = "" # for character in generate_fn(prompt, history[:-1]): # history[-1][1] = character # yield history, prompt_html # print('Final history is ',history) # #store_message(db,history[-1][0],history[-1][1],cross_encoder) # # def system_instructions(question_difficulty, topic,documents_str): # # return f""" [INST] Your are a great teacher and your task is to create 10 questions with 4 choices with a {question_difficulty} difficulty about topic request " {topic} " only from the below given documents, {documents_str} then create an answers. Index in JSON format, the questions as "Q#":"" to "Q#":"", the four choices as "Q#:C1":"" to "Q#:C4":"", and the answers as "A#":"Q#:C#" to "A#":"Q#:C#". [/INST]""" # RAG_db = gr.State() # # def load_model(): # # try: # # # Initialize the model # # RAG = RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0") # # # Load the RAG database # # RAG_db.value = RAG.from_index('.ragatouille/colbert/indexes/cbseclass10index') # # return 'Ready to Go!!' # # except Exception as e: # # return f"Error loading model: {e}" # # def generate_quiz(question_difficulty, topic): # # if not topic.strip(): # # return ['Please enter a valid topic.'] + [gr.Radio(visible=False) for _ in range(10)] # # top_k_rank = 10 # # # Load the model and database within the generate_quiz function # # try: # # RAG = RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0") # # RAG_db_ = RAG.from_index('.ragatouille/colbert/indexes/cbseclass10index') # # gr.Warning('Model loaded!') # # except Exception as e: # # return [f"Error loading model: {e}"] + [gr.Radio(visible=False) for _ in range(10)] # # RAG_db_ = RAG_db.value # # documents_full = RAG_db_.search(topic, k=top_k_rank) # # generate_kwargs = dict( # # temperature=0.2, # # max_new_tokens=4000, # # top_p=0.95, # # repetition_penalty=1.0, # # do_sample=True, # # seed=42, # # ) # # question_radio_list = [] # # count = 0 # # while count <= 3: # # try: # # documents = [item['content'] for item in documents_full] # # document_summaries = [f"[DOCUMENT {i+1}]: {summary}{count}" for i, summary in enumerate(documents)] # # documents_str = '\n'.join(document_summaries) # # formatted_prompt = system_instructions(question_difficulty, topic, documents_str) # # pre_prompt = [ # # {"role": "system", "content": formatted_prompt} # # ] # # response = client.text_generation( # # formatted_prompt, **generate_kwargs, stream=False, details=False, return_full_text=False, # # ) # # output_json = json.loads(f"{response}") # # global quiz_data # # quiz_data = output_json # # for question_num in range(1, 11): # # question_key = f"Q{question_num}" # # answer_key = f"A{question_num}" # # question = quiz_data.get(question_key) # # answer = quiz_data.get(quiz_data.get(answer_key)) # # if not question or not answer: # # continue # # choice_keys = [f"{question_key}:C{i}" for i in range(1, 5)] # # choice_list = [quiz_data.get(choice_key, "Choice not found") for choice_key in choice_keys] # # radio = gr.Radio(choices=choice_list, label=question, visible=True, interactive=True) # # question_radio_list.append(radio) # # if len(question_radio_list) == 10: # # break # # else: # # count += 1 # # continue # # except Exception as e: # # count += 1 # # if count == 3: # # return ['Sorry. Pls try with another topic!'] + [gr.Radio(visible=False) for _ in range(10)] # # continue # # return ['Quiz Generated!'] + question_radio_list # # def compare_answers(*user_answers): # # user_answer_list = user_answers # # answers_list = [quiz_data.get(quiz_data.get(f"A{question_num}")) for question_num in range(1, 11)] # # score = sum(1 for answer in user_answer_list if answer in answers_list) # # if score > 7: # # message = f"### Excellent! You got {score} out of 10!" # # elif score > 5: # # message = f"### Good! You got {score} out of 10!" # # else: # # message = f"### You got {score} out of 10! Don’t worry, you can prepare well and try better next time!" # # return message # #with gr.Blocks(theme='Insuz/SimpleIndigo') as demo: # with gr.Blocks(theme='NoCrypt/miku') as CHATBOT: # with gr.Row(): # with gr.Column(scale=10): # # gr.Markdown( # # """ # # # Theme preview: `paris` # # To use this theme, set `theme='earneleh/paris'` in `gr.Blocks()` or `gr.Interface()`. # # You can append an `@` and a semantic version expression, e.g. @>=1.0.0,<2.0.0 to pin to a given version # # of this theme. # # """ # # ) # gr.HTML(value="""

ADWITIYA-

Custom Manual Chatbot and Quizbot

#
""", elem_id='heading') # gr.HTML(value=f""" #

# Using GenAI for CBIC Capacity Building - A free chat bot developed by National Customs Targeting Center using Open source LLMs for CBIC Officers #

# """, elem_id='Sub-heading') # #usage_count = get_and_increment_value_count(db,collection_name, field_name) # gr.HTML(value=f"""

Developed by NCTC,Mumbai . Suggestions may be sent to ramyadevi1607@yahoo.com.

""", elem_id='Sub-heading1 ') # with gr.Column(scale=3): # gr.Image(value='logo.png',height=200,width=200) # chatbot = gr.Chatbot( # [], # elem_id="chatbot", # avatar_images=('https://aui.atlassian.com/aui/8.8/docs/images/avatar-person.svg', # 'https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg'), # bubble_full_width=False, # show_copy_button=True, # show_share_button=True, # ) # with gr.Row(): # txt = gr.Textbox( # scale=3, # show_label=False, # placeholder="Enter text and press enter", # container=False, # ) # txt_btn = gr.Button(value="Submit text", scale=1) # cross_encoder = gr.Radio(choices=['(FAST) MiniLM-L6v2','(ACCURATE) BGE reranker','(HIGH ACCURATE) ColBERT'], value='(ACCURATE) BGE reranker',label="Embeddings", info="Only First query to Colbert may take litte time)") # prompt_html = gr.HTML() # # Turn off interactivity while generating if you click # txt_msg = txt_btn.click(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( # bot, [chatbot, cross_encoder], [chatbot, prompt_html])#.then(update_count_html,[],[count_html]) # # Turn it back on # txt_msg.then(lambda: gr.Textbox(interactive=True), None, [txt], queue=False) # # Turn off interactivity while generating if you hit enter # txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( # bot, [chatbot, cross_encoder], [chatbot, prompt_html])#.then(update_count_html,[],[count_html]) # # Turn it back on # txt_msg.then(lambda: gr.Textbox(interactive=True), None, [txt], queue=False) # # Examples # gr.Examples(examples, txt) # # with gr.Blocks(title="Quiz Maker", theme=gr.themes.Default(primary_hue="green", secondary_hue="green"), css="style.css") as QUIZBOT: # # with gr.Column(scale=4): # # gr.HTML(""" # #
# #

ADWITIYA Customs Manual Quizbot

# #

Generative AI-powered Capacity building for Training Officers

# # ⚠️ NACIN Faculties create quiz from any topic dynamically for classroom evaluation after their sessions! ⚠️ # #
# # """) # # with gr.Column(scale=2): # # gr.HTML(""" # #
# #

Ready!

# #
# # """) # # # load_btn = gr.Button("Click to Load!🚀") # # # load_text = gr.Textbox() # # # load_btn.click(fn=load_model, outputs=load_text) # # topic = gr.Textbox(label="Enter the Topic for Quiz", placeholder="Write any topic/details from Customs Manual") # # with gr.Row(): # # radio = gr.Radio(["easy", "average", "hard"], label="How difficult should the quiz be?") # # generate_quiz_btn = gr.Button("Generate Quiz!🚀") # # quiz_msg = gr.Textbox() # # question_radios = [gr.Radio(visible=False) for _ in range(10)] # # generate_quiz_btn.click( # # fn=generate_quiz, # # inputs=[radio, topic], # # outputs=[quiz_msg] + question_radios # # ) # # check_button = gr.Button("Check Score") # # score_textbox = gr.Markdown() # # check_button.click( # # fn=compare_answers, # # inputs=question_radios, # # outputs=score_textbox # # ) # #demo = gr.TabbedInterface([CHATBOT, QUIZBOT], ["AI ChatBot", "AI Quizbot"]) # CHATBOT.queue() # CHATBOT.launch(debug=True)