app.py

from haystack.document_stores.memory import InMemoryDocumentStore
from haystack.nodes import TfidfRetriever, FARMReader

import pickle

pickle_file = 'knowledge_graph.pickle'

# Load the knowledge graph from the pickle file
with open(pickle_file, 'rb') as f:
  knowledge_graph = pickle.load(f)

document_store = InMemoryDocumentStore()
node_sentences = {}
documents = []
nodes = [node for node in knowledge_graph.nodes() if node is not None]

for node in nodes:
  # Get all the edges related to the current node
  related_edges = [edge for edge in knowledge_graph.edges() if edge[0] == node or edge[1] == node]

  # Get the parents and grandparents of the current node
  parents = [edge[0] for edge in related_edges if edge[1] == node]
  grandparents = []
  for parent in parents:
    grandparents.extend([edge[0] for edge in related_edges if edge[1] == parent])

  # Get the children and grandchildren of the current node
  children = [edge[1] for edge in related_edges if edge[0] == node]
  grandchildren = []
  for child in children:
    grandchildren.extend([edge[1] for edge in related_edges if edge[0] == child])

  # Create the sentence by combining all the related nodes
  sentence_parts = grandparents + parents + [node] + children + grandchildren
  sentence = ' '.join(sentence_parts)

  # Store the sentence for the current node
  node_sentences[node] = sentence

  # Create the document with the node and the sentence as the content
  documents.append({'text': node, 'content': sentence})
document_store.write_documents(documents)

#Initialize the retriever
retriever = TfidfRetriever(document_store=document_store)

#Initialize the reader
model_name = "primasr/multilingualbert-for-eqa-finetuned"
reader = FARMReader(model_name_or_path=model_name, use_gpu=False)

#Create pipeline with the component of retriever and reader
from haystack.pipelines import Pipeline
pipeline = Pipeline()
pipeline.add_node(component=retriever, name="Retriever", inputs=["Query"])
pipeline.add_node(component=reader, name="Reader", inputs=["Retriever"])

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# Targeted to Translate English queries to Malay Language
# Initialize the tokenizer
en_id_tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-id")
# Initialize the model
en_id_model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-id")

# Targeted to Translate Malay Answer to English Language
# Initialize the tokenizer
id_en_tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-id-en")
# Initialize the model
id_en_model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-id-en")

#Defined some pairs for chatbot response
pairs = [
    [
        "your answer is wrong",
        "Sorry for providing wrong answer, here is the newest answer:\n\n",
        "I am sorry that I can't actually answer your question =("
    ],
    [
        "jawapan anda adalah salah",
        "Maaf sedangkan memberi jawapan yang salah. Berikut adalah jawapan yang baru:\n\n",
        "Minta Maaf, saya tidak boleh menemukan soalan anda =("
    ]]

#Function for checking reiterate providing answer for same question
def checkReiterateQuery(query,lang):
  if query in [pairs[0][0],pairs[1][0]]:
    if lang == 'en':
      j = 0
    else:
      j = 1
    return True, j

  else:
    return False , 3

import gradio as gr
from langdetect import detect
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

chat_history = []
answer_counter = 0

def chatbot_interface(message):
  global answer_counter
  global result

  # Append the current message to the chat history
  chat_history.append(message)
  lang = detect(message)
  reiterate, j = checkReiterateQuery(message, lang)

  #If user want to re-iterate the answer for same question
  if reiterate:
    answer_counter = answer_counter + 1
    if answer_counter < 5:
      retrieved_main_answer = pairs[j][1] + result['answers'][answer_counter].answer
      retrieved_main_context = result['answers'][answer_counter].context
    else:
      retrieved_main_answer = pairs[j][2]
      retrieved_main_context = ""
  else:
    answer_counter = 0
    #if language is english then convert it to malay language
    if lang == "en":
      tokenized_text = en_id_tokenizer.prepare_seq2seq_batch([message], return_tensors='pt')
      translation = en_id_model.generate(**tokenized_text)
      message = en_id_tokenizer.batch_decode(translation, skip_special_tokens=True)[0]

    result = pipeline.run(query=message.lower(), params={
            "Retriever": {"top_k": 5},
            "Reader": {"top_k": 5}})
    retrieved_main_answer = result['answers'][answer_counter].answer
    retrieved_main_context = result['answers'][answer_counter].context

  response = retrieved_main_answer + ", " + retrieved_main_context

  #Convert the response to english if user ask question in english
  if lang == "en":
    tokenized_text = id_en_tokenizer.prepare_seq2seq_batch([response.lower()], return_tensors='pt')
    translation = id_en_model.generate(**tokenized_text)
    response = id_en_tokenizer.batch_decode(translation, skip_special_tokens=True)[0]

  # Append the response to the chat history
  chat_history.append(response)

  # Join the chat history with newline characters
  chat_history_text = "\n\n".join(chat_history)

  return response, chat_history_text

# Create a Gradio interface
iface = gr.Interface(
    fn=chatbot_interface,
    inputs=gr.inputs.Textbox(label="Please Type Your Question Here: "),
    outputs=[gr.outputs.Textbox(label="Answers"), gr.outputs.Textbox(label="Chat History")],
    description="## Question Answering system\n\nIt supports **English** and **Bahasa Malaysia**.",
    allow_flagging = False
)

#Demo for the chatbot
io.launch(inline = False)