Snippets in employee_handbook_qna_gpt3.ipynb

step1_get_corpus_embeddings.py

from sentence_transformers import SentenceTransformer, util
import openai

with open('HKIHRMEmployeeHandbook.txt') as f:
    full_text = f.read()

passages = [line.strip() for line in full_text.split('\n \n') if len(line.strip()) > 0]

model_name = 'multi-qa-mpnet-base-cos-v1'
bi_encoder = SentenceTransformer(model_name)
corpus_embeddings = bi_encoder.encode(passages, convert_to_tensor=True, show_progress_bar=True)

step2_search_relevant_passages.py

query = 'what should I do if I worked overtime?'

question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
hits = [hit for hit in hits[0]]

hits = sorted([hit['corpus_id'] for hit in hits])
context = "\n".join([passages[hit] for hit in hits])

step3_create_prompt.py

template = """Context:
<<context>>

Answer the following question:
Q: <<query>>
A: 
"""
prompt = template.replace('<<context>>', context).replace('<<query>>', query)

step4_query_gpt3.py

import openai

openai.api_key = input("OpenAI API Key: ")
response = openai.Completion.create(engine="text-davinci-003", prompt=prompt, max_tokens=256, temperature=0.2)
print(response['choices'][0]['text'])

step5_answer_function.py

def answer(query):
    # Encode the query using the bi-encoder and find potentially relevant passages
    question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
    hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
    hits = [hit for hit in hits[0]]

    hits = sorted([hit['corpus_id'] for hit in hits])
    context = "\n".join([passages[hit] for hit in hits])

    template = """Context:
<<context>>

Answer the following question by paraphrasing it and then elaborate the answer:
Q: <<query>>
A: 
"""
    prompt = template.replace('<<context>>', context).replace('<<query>>', query)
    prompt_length = len(tokenizer(prompt)['input_ids'])
    response = openai.Completion.create(engine="text-davinci-003", prompt=prompt, max_tokens=4096-prompt_length, temperature=0.2)
    return response['choices'][0]['text']

step6_qa_loop.py

while True:
    query = input("Q: ")
    if query == "xxx": break
    ans = answer(query)
    print(f"A: {ans}")
    print("="*70)
    print()

step7_gradio_ui.py

import gradio as gr
examples = [
  ["can I carry forward annual leave?"],
  ["do I entitle to compensation leave?"],
  ["how many days of annual leave do I have?"],
  ["how to reimburse the company expenses?"],
  ["do I get paid if I got sick?"],
  ["what should I do if I got sick?"],
]

title = "Q&A Demo"

def inference(text):
  return answer(text)

io = gr.Interface(
  inference,
  gr.Textbox(lines=3),
  outputs=[
    gr.Textbox(lines=3, label="GPT 3.5")
  ],
  title=title,
  examples=examples
)
io.launch()