LlamaIndex와 W&B Weave로 구축하는 RAG 기반 디지털 레스토랑 메뉴

    with open(pdf_path, 'rb') as file:
        reader = PyPDF2.PdfReader(file)
        for page_num in range(len(reader.pages)):
            page_text = reader.pages[page_num].extract_text()
            text_chunks = split_text(page_text)
            for chunk in text_chunks:

                prompt_text = ("I'm going to give you some text that I want you to convert into a JSON object (list), "
                            "each item with a title and description. For example, if the text describes various dishes at a restaurant, "
                            "the JSON output might look like: [{'title': 'Chicken Special', 'description': 'A delicious chicken dish seasoned with herbs and spices.', 'keywords': 'chicken'}, "
                            "{'title': 'Seafood Platter', 'description': 'An assortment of fresh seafood, including shrimp, scallops, and lobster.', 'keywords': 'seafood, shrimp,platter'}]"
                            "NOTE: THE ITEM TITLES WILL NOT BE A CATEGORY OF ITEMS, RATHER THEY ARE SPECIFIC DISHES/ENTRES/APETIZERS/DRINKS ETC. -> THEY ARE ONLY SPECIC ITEMS, NOT CATEGORIES"
                            "NOTE: Keywords should be the a 2-4 categories/common search terms describing the item, for example (but not limited to) -> dessert, side, chicken, sandwhich, cocktail, drink, burger, salad, etc etc -> Try to use multiple categories/descriptors"
                            "THIS DATA IS CHUNKED SO IF THE DATA AT THE BEGINNING OR END (FOR TITLE OR DESCRIPTION) SEEMS INCOMPLETE COMPARED TO THE REST OF THE ENTRIES, JUST SKIP IT"
                            "ONLY RESPOND WITH THE JSON OBJECT AND NOTHING ELSE. CONVERT THE FULL DATA. DO NOT TRUNCATE OR STOP EARLY. HERE IS THE DATA I WANT YOU TO CONVERT: " + chunk)

                response = client.chat.completions.create(
                    model="gpt-4-1106-preview",
                    messages=[
                        {"role": "system", "content": "You are a helpful assistant."},
                        {"role": "user", "content": prompt_text}
                    ]
                )


    {
        "title": "Roasted Half Chicken",
        "description": "Slow-roasted with housemade herb butter. Served with choice of two sides.",
        "keywords": "chicken, roasted, herb butter",
        "page": 1
    },
    {
        "title": "Parmesan Crusted Chicken Pasta",
        "description": "Crispy hand breaded parmesan chicken breast with melted mozzarella and marinara sauce over linguine.",
        "keywords": "chicken, pasta, parmesan",
        "page": 1
    },

from flask import Flask, render_template, request, jsonify
import json
import os
from llama_index import VectorStoreIndex, ServiceContext, StorageContext, load_index_from_storage
from llama_index.schema import Document
from llama_index.embeddings import HuggingFaceEmbedding
from llama_index.retrievers import VectorIndexRetriever
from llama_index.query_engine import RetrieverQueryEngine
from llama_index.postprocessor import SimilarityPostprocessor
from llama_index import (VectorStoreIndex, get_response_synthesizer)

def create_document_from_json_item(json_item):
    ti = json_item['title'] 
    des = json_item['description']
    keys = json_item['keywords']

    # if des:  # depending on the menu/embedding model, this could be helpful 
    #     ti += des

    if keys: 
        ti += "keywords:" + keys
    document = Document(text=ti, metadata=json_item)
    return document

def generate_embeddings_for_document(document, model_name="BAAI/bge-small-en-v1.5"):
    embed_model = HuggingFaceEmbedding(model_name=model_name)
    embeddings = embed_model.get_text_embedding(document.text)
    return embeddings

file_path = "./gpt4_menu_data_v2.json"
index = None 

if not os.path.exists("./index"):        
    with open(file_path, 'r', encoding='utf-8') as file:
        json_data = json.load(file)

    documents = []
    for item in json_data:
        document = create_document_from_json_item(item)
        document_embeddings = generate_embeddings_for_document(document)
        document.embedding = document_embeddings
        documents.append(document)

    service_context = ServiceContext.from_defaults(llm=None, embed_model='local')
    index = VectorStoreIndex.from_documents(documents, service_context=service_context)
    index.storage_context.persist(persist_dir="./index")
else:
    storage_context = StorageContext.from_defaults(persist_dir="./index")
    service_context = ServiceContext.from_defaults(llm=None, embed_model='local')
    index = load_index_from_storage(storage_context, service_context=service_context)


retriever = VectorIndexRetriever(index=index, similarity_top_k=10)
service_context = ServiceContext.from_defaults(llm=None, embed_model='local')
response_synthesizer = get_response_synthesizer(service_context=service_context)
query_engine = RetrieverQueryEngine(retriever=retriever, response_synthesizer=response_synthesizer, node_postprocessors=[SimilarityPostprocessor(similarity_cutoff=0.6)])

def parse_response_to_json(response_str):
    items = response_str.split("title: ")[1:]  # Split the response and ignore the first empty chunk
    json_list = []

    for item in items:
        lines = item.strip().split('\n')
        item_json = {
            "title": lines[0].strip(),
            "description": lines[1].replace("description: ", "").strip(),
            "keywords": lines[2].replace("keywords: ", "").strip(),
            "page": int(lines[3].replace("page: ", "").strip())
        }
        json_list.append(item_json)

    return json_list

def query_index(query):
    response = query_engine.query(query)
    # Original results parsed to JSON
    return parse_response_to_json(str(response))

# Login to W&B
wandb.login()
# Define constants for the StreamTable
WB_ENTITY = ""  # Set your W&B entity name here, or leave it empty to use the current logged-in entity
WB_PROJECT = "ai_menu"
STREAM_TABLE_NAME = "usage_data"
# Define a StreamTable
st = StreamTable(f"{WB_ENTITY}/{WB_PROJECT}/{STREAM_TABLE_NAME}")

app = Flask(__name__)

client = openai.OpenAI(api_key='sk-YOUR_API_KEY')

@app.route('/chat')
def index():
    return render_template('chat.html')

@app.route('/')
def menu():
    return render_template('index.html')

def describe_items(json_list):
    description_str = "Some possible items you might be interested in include the following:<br><br>"
    for item in json_list:
        description_str += f"<strong>{item['title']}</strong> - {item['description']}<br><br>"
    return description_str

def generate_response_gpt(query, original_res):
    # Generating prompt for GPT
    prompt = f"This is a user at a restaurant searching for items to order. Given these initial results {original_res} for the following user query '{query}', return the JSON object for the items that make sense to include as a response (e.g., remove only items that are not at all relevant to the query='{query}') -- keep in mind that they may all be relevant and its perfectly fine to not remove any items. YOU MUST RETURN THE RESULT IN JSON FORM"
    
    response = client.chat.completions.create(
        model="gpt-3.5-turbo-1106",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": prompt}
        ]
    )
    
    if response.choices:
        reply = response.choices[0].message.content
        filtered_res_json_str = re.search(r"```json(.+?)```", reply, re.DOTALL)

        print(filtered_res_json_str)
        if filtered_res_json_str:
            filtered_res_json = json.loads(filtered_res_json_str.group(1))
            if not len(filtered_res_json): 
                return original_res
        else:
            filtered_res_json = original_res
        
        
        return filtered_res_json
    else:
        return original_res


@app.route('/search', methods=['POST'])
def search():
    query = request.form.get('query')
    original_res = query_index(query)
    filtered_res_json = generate_response_gpt(query, original_res)
    st.log({"query": query, "results": describe_items(filtered_res_json)})
    return jsonify({'res': describe_items(filtered_res_json)})


if __name__ == '__main__':
    app.run(host='0.0.0.0', port=5000, debug=True)