|
from dotenv import load_dotenv |
|
from langchain.text_splitter import RecursiveCharacterTextSplitter |
|
from langchain_community.vectorstores import FAISS |
|
|
|
from langchain_community.embeddings import HuggingFaceEmbeddings |
|
import os |
|
import pandas as pd |
|
|
|
import gradio as gr |
|
from openai import OpenAI |
|
|
|
load_dotenv(override=True) |
|
client = OpenAI() |
|
DB_FAISS_PATH = "./vectorstore/db_faiss_50k" |
|
data_file_path = "./data/132_webmd_vogon_urlsContent_cleaned.tsv" |
|
|
|
|
|
|
|
|
|
CHUNK_SIZE = 512 |
|
CHUNK_OVERLAP = 128 |
|
|
|
embedding_model_hf = "BAAI/bge-m3" |
|
|
|
qa_model_name = "gpt-3.5-turbo" |
|
bestReformulationPrompt = "Given a chat history and the latest user question, which may reference context from the chat history, you must formulate a standalone question that can be understood without the chat history. You are strictly forbidden from using any outside knowledge. Do not, under any circumstances, answer the question. Reformulate it if necessary; otherwise, return it as is." |
|
bestSystemPrompt = "You're an assistant for question-answering tasks. Under absolutely no circumstances should you use external knowledge or go beyond the provided preknowledge. Your approach must be systematic and meticulous. First, identify CLUES such as keywords, phrases, contextual information, semantic relations, tones, and references that aid in determining the context of the input. Second, construct a concise diagnostic REASONING process (limiting to 130 words) based on premises supporting the INPUT relevance within the provided context. Third, utilizing the identified clues, reasoning, and input, furnish the pertinent answer for the question. Remember, you are required to use ONLY the provided context to answer the questions. If the question does not align with the preknowledge or if the preknowledge is absent, state that you don't know the answer. External knowledge is strictly prohibited. Failure to adhere will result in incorrect answers. The preknowledge is as follows:" |
|
|
|
|
|
embeddings_hf = HuggingFaceEmbeddings(model_name = embedding_model_hf, show_progress = True) |
|
|
|
def setupDb(data_path): |
|
df = pd.read_csv(data_path, sep="\t") |
|
relevant_content = df["url"].values |
|
text_splitter = RecursiveCharacterTextSplitter( |
|
chunk_size=CHUNK_SIZE, |
|
chunk_overlap=CHUNK_OVERLAP, |
|
) |
|
|
|
if not os.path.exists(DB_FAISS_PATH): |
|
split_docs = text_splitter.create_documents( |
|
df["url_content"].tolist(), |
|
metadatas=[ |
|
{"title": row["url_title"], "url": row["url"]} |
|
for _, row in df.iterrows() |
|
], |
|
) |
|
print(f"Documents are split into {len(split_docs)} passages") |
|
|
|
db = FAISS.from_documents(split_docs, embeddings_hf) |
|
print(f"Document saved in db") |
|
db.save_local(DB_FAISS_PATH + "/index_1") |
|
else: |
|
print(f"Db already exists") |
|
db = FAISS.load_local( |
|
DB_FAISS_PATH, embeddings_hf, allow_dangerous_deserialization=True |
|
) |
|
return db, relevant_content |
|
|
|
def reformulate_question(chat_history, latest_question, reformulationPrompt): |
|
system_message = { |
|
"role": "system", |
|
"content": reformulationPrompt |
|
} |
|
|
|
formatted_history = [] |
|
for i, chat in enumerate(chat_history): |
|
formatted_history.append({"role": "user", "content": chat[0]}) |
|
formatted_history.append({"role": "assistant", "content": chat[1]}) |
|
|
|
|
|
formatted_history.append({"role": "user", "content": latest_question}) |
|
response = client.chat.completions.create( |
|
model="gpt-3.5-turbo", |
|
messages=[system_message] + formatted_history, |
|
temperature=0 |
|
) |
|
|
|
reformulated_question = response.choices[0].message.content |
|
return reformulated_question |
|
|
|
def getQuestionAnswerOnTheBasisOfContext(question, context, systemPrompt): |
|
system_message = { |
|
"role": "system", |
|
"content": systemPrompt + context |
|
} |
|
|
|
response = client.chat.completions.create( |
|
model=qa_model_name, |
|
messages=[system_message] + [{"role": "user", "content": question}], |
|
temperature=0 |
|
) |
|
answer = response.choices[0].message.content |
|
return answer |
|
|
|
|
|
def chatWithRag(reformulationPrompt, QAPrompt, question): |
|
global curr_question_no, chat_history |
|
curr_question_prompt = bestSystemPrompt |
|
if QAPrompt != None or len(QAPrompt): |
|
curr_question_prompt = QAPrompt |
|
|
|
|
|
reformulated_query = question |
|
retreived_documents = [doc for doc in db.similarity_search_with_score(reformulated_query) if doc[1] < 1.3] |
|
answer = getQuestionAnswerOnTheBasisOfContext(reformulated_query, '. '.join([doc[0].page_content for doc in retreived_documents]), curr_question_prompt) |
|
chat_history.append((question, answer)) |
|
curr_question_no += 1 |
|
docs_info = "\n\n".join([ |
|
f"Title: {doc[0].metadata['title']}\nUrl: {doc[0].metadata['url']}\nContent: {doc[0].page_content}\nValue: {doc[1]}" for doc in retreived_documents |
|
]) |
|
full_response = f"Answer: {answer}\n\nReformulated question: {reformulated_query}\nRetrieved Documents:\n{docs_info}" |
|
|
|
return full_response |
|
|
|
db, relevant_content = setupDb(data_file_path) |
|
chat_history = [] |
|
curr_question_no = 1 |
|
|
|
with gr.Blocks() as demo: |
|
gr.Markdown("# RAG on webmd") |
|
with gr.Row(): |
|
reformulationPrompt = gr.Textbox(bestReformulationPrompt, lines=1, placeholder="Enter the system prompt for reformulation of query", label="Reformulation System prompt") |
|
QAPrompt = gr.Textbox(bestSystemPrompt, lines=1, placeholder="Enter the system prompt for QA.", label="QA System prompt") |
|
question = gr.Textbox(lines=1, placeholder="Enter the question asked", label="Question") |
|
output = gr.Textbox(label="Output") |
|
submit_btn = gr.Button("Submit") |
|
submit_btn.click(chatWithRag, inputs=[reformulationPrompt, QAPrompt, question], outputs=output) |
|
question.submit(chatWithRag, [reformulationPrompt, QAPrompt, question], [output]) |
|
with gr.Accordion("Urls", open=False): |
|
gr.Markdown(', '.join(relevant_content)) |
|
|
|
gr.close_all() |
|
demo.launch() |