File size: 5,976 Bytes
f9149ca e8e039f f9149ca e8e039f f9149ca bc866d0 f9149ca e8e039f f9149ca e8e039f f9149ca e8e039f f9149ca e8e039f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 |
## Setup
# Import the necessary Libraries
import json
import tiktoken
import os
import pandas as pd
import uuid
import gradio as gr
from openai import OpenAI
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFDirectoryLoader
from langchain_community.embeddings.sentence_transformer import (
SentenceTransformerEmbeddings
)
from langchain_community.vectorstores import Chroma
from huggingface_hub import CommitScheduler
from pathlib import Path
# Create Client
client = OpenAI()
# Define the embedding model and the vectorstore
collection_name = 'project3_rag_db'
embedding_model_name = 'thenlper/gte-large'
embedding_model = SentenceTransformerEmbeddings(model_name=embedding_model_name)
persisted_vectordb_location = './project3_rag_db'
model_name = 'gpt-4o-mini'
# Load the persisted vectorDB
vectorstore_persisted = Chroma(
collection_name=collection_name,
persist_directory=persisted_vectordb_location,
embedding_function=embedding_model)
# Prepare the logging functionality
log_file = Path("logs/") / f"data_{uuid.uuid4()}.json"
log_folder = log_file.parent
scheduler = CommitScheduler(
repo_id="anirudhabokil/project3_rag_10K_chatbot_logs",
repo_type="dataset",
folder_path=log_folder,
path_in_repo="data",
every=2
)
# Define the Q&A system message
qna_system_message = """
You are an assistant to a Financial Analyst for a Fin tech company. Your task is to provide relevant information about analysis of key information from 10-K reports.
10-K reports are comprehensive annual reports filed by publicly traded companies in the United States with the Securities and Exchange Commission (SEC).
User input will include the necessary context for you to answer their questions. This context will begin with the token: ###Context.
The context contains references to specific portions of documents relevant to the user's query, along with source links.
The source for a context will begin with the token ###Source:
When crafting your response:
1. Select only context relevant to answer the question.
2. Include the source links in your response.
3. User questions will begin with the token: ###Question.
4. If the question is irrelevant to 10-K reports respond with - "I am an assistant to a Financial Analyst. I can only help you with questions related to 10-K reports"
Please adhere to the following guidelines:
- Your response should only be about the question asked and nothing else.
- Answer only using the context provided.
- Do not mention anything about the context in your final answer.
- If the answer is not found in the context, it is very very important for you to respond with "I don't know."
- Always quote the source when you use the context. Cite the relevant source at the end of your response under the section - Source:
- Do not make up sources. Use the links provided in the sources section of the context and nothing else. You are prohibited from providing other links/sources.
Here is an example of how to structure your response:
Answer:
[Answer]
Source:
[Use the ###Source provided in the context as it. Do not add https prefix]
"""
# Define the user message template
qna_user_message_template = """
###Context
Here are some 10-K reports and their source links that are relevant to the question mentioned below.
{context}
###Question
{question}
"""
# Define the predict function that runs when 'Submit' is clicked or when a API request is made
def predict(user_input,company):
filter = "/content/dataset/"+company+"-10-k-2023.pdf"
print(filter)
relevant_document_chunks = vectorstore_persisted.similarity_search(user_input, k=5, filter={"source":filter})
print(relevant_document_chunks)
# Create context_for_query
context_list = [d.page_content + "\n ###Source: " + d.metadata['source'] + '\n\n ' for d in relevant_document_chunks]
context_for_query = ". ".join(context_list)
print(context_for_query)
# Create messages
prompt = [
{'role': 'system', 'content': qna_system_message},
{'role': 'user', 'content': qna_user_message_template.format(
context=context_for_query,
question=user_input
)
}]
print(prompt)
# Get response from the LLM
try:
response = client.chat.completions.create(model=model_name,messages=prompt,temperature=0)
print(response)
answer = response.choices[0].message.content.strip()
# Handle errors using try-except
except Exception as e:
answer = f'Sorry, I encountered the following error: \n {e}'
# While the prediction is made, log both the inputs and outputs to a local log file
# While writing to the log file, ensure that the commit scheduler is locked to avoid parallel
# access
with scheduler.lock:
with log_file.open("a") as f:
f.write(json.dumps(
{
'user_input': user_input,
'retrieved_context': context_for_query,
'model_response': answer
}
))
f.write("\n")
return answer
# Set-up the Gradio UI
# Add text box and radio button to the interface
# The radio button is used to select the company 10k report in which the context needs to be retrieved.
user_input = gr.Textbox(label="Ask your question")
company = gr.Dropdown(['aws','google','IBM','Meta','msft'], label="Company")
answer = gr.Label(label="Answer")
# Create the interface
# For the inputs parameter of Interface provide [textbox,company]
demo = gr.Interface(fn=predict,
inputs=[user_input, company],
outputs=answer,
title="10-K Chatbot",
description="This API answers questions based on 10-k reports",
flagging_mode="auto",
concurrency_limit=8)
demo.queue()
demo.launch(share=True, debug=True)
|