app.py

import gradio as gr
import pandas as pd
import pixeltable as pxt
from pixeltable.iterators import DocumentSplitter
import numpy as np
from pixeltable.functions.huggingface import sentence_transformer
from pixeltable.functions import openai
import os

# Ensure a clean slate for the demo
pxt.drop_dir('rag_demo', force=True)
pxt.create_dir('rag_demo')

# Set up embedding function
@pxt.expr_udf
def e5_embed(text: str) -> np.ndarray:
    return sentence_transformer(text, model_id='intfloat/e5-large-v2')

# Create prompt function
@pxt.udf
def create_prompt(top_k_list: list[dict], question: str) -> str:
    concat_top_k = '\n\n'.join(
        elt['text'] for elt in reversed(top_k_list)
    )
    return f'''
    PASSAGES:

    {concat_top_k}

    QUESTION:

    {question}'''

def process_files(ground_truth_file, pdf_files):
    # Process ground truth file
    if ground_truth_file.name.endswith('.csv'):
        df = pd.read_csv(ground_truth_file.name)
    else:
        df = pd.read_excel(ground_truth_file.name)
    
    queries_t = pxt.create_table('rag_demo.queries', df)

    # Process PDF files
    documents_t = pxt.create_table(
        'rag_demo.documents',
        {'document': pxt.DocumentType()}
    )
    
    for pdf_file in pdf_files:
        documents_t.insert({'document': pdf_file.name})

    # Create chunks view
    chunks_t = pxt.create_view(
        'rag_demo.chunks',
        documents_t,
        iterator=DocumentSplitter.create(
            document=documents_t.document,
            separators='token_limit',
            limit=300
        )
    )

    # Add embedding index
    chunks_t.add_embedding_index('text', string_embed=e5_embed)

    # Create top_k query
    @chunks_t.query
    def top_k(query_text: str):
        sim = chunks_t.text.similarity(query_text)
        return (
            chunks_t.order_by(sim, asc=False)
                .select(chunks_t.text, sim=sim)
                .limit(5)
        )

    # Add computed columns to queries_t
    queries_t['question_context'] = chunks_t.top_k(queries_t.Question)
    queries_t['prompt'] = create_prompt(
        queries_t.question_context, queries_t.Question
    )

    # Prepare messages for OpenAI
    messages = [
        {
            'role': 'system',
            'content': 'Please read the following passages and answer the question based on their contents.'
        },
        {
            'role': 'user',
            'content': queries_t.prompt
        }
    ]

    # Add OpenAI response column
    queries_t['response'] = openai.chat_completions(
        model='gpt-4-0125-preview', messages=messages
    )
    queries_t['answer'] = queries_t.response.choices[0].message.content

    return "Files processed successfully!"

def query_llm(question):
    queries_t = pxt.get_table('rag_demo.queries')
    chunks_t = pxt.get_table('rag_demo.chunks')

    # Perform top-k lookup
    context = chunks_t.top_k(question).collect()

    # Create prompt
    prompt = create_prompt(context, question)

    # Prepare messages for OpenAI
    messages = [
        {
            'role': 'system',
            'content': 'Please read the following passages and answer the question based on their contents.'
        },
        {
            'role': 'user',
            'content': prompt
        }
    ]

    # Get LLM response
    response = openai.chat_completions(model='gpt-4-0125-preview', messages=messages)
    answer = response.choices[0].message.content

    # Add new row to queries_t
    new_row = {'Question': question, 'answer': answer}
    queries_t.insert([new_row])

    # Return updated dataframe
    return queries_t.select(queries_t.Question, queries_t.answer).collect()

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# RAG Demo App")
    
    with gr.Row():
        ground_truth_file = gr.File(label="Upload Ground Truth (CSV or XLSX)")
        pdf_files = gr.File(label="Upload PDF Documents", file_count="multiple")
    
    process_button = gr.Button("Process Files")
    process_output = gr.Textbox(label="Processing Output")
    
    question_input = gr.Textbox(label="Enter your question")
    query_button = gr.Button("Query LLM")
    
    output_dataframe = gr.Dataframe(label="LLM Outputs")

    process_button.click(process_files, inputs=[ground_truth_file, pdf_files], outputs=process_output)
    query_button.click(query_llm, inputs=question_input, outputs=output_dataframe)

if __name__ == "__main__":
    demo.launch()