Building a Recipe Assistant with AI

Imagine having a virtual assistant that understands user preferences, context, and queries with precision—providing relevant and tailored responses in seconds. This is the promise of an AI-powered chatbot, a tool designed to make interactions more intuitive, efficient, and personalized using OpenAI API and Retrieval-Augmented Generation (RAG) frameworks such as LangChain.

Problem

The internet is flooded with information, forums, and resources, making it difficult to find precise and relevant answers. While this abundance is useful, it also presents challenges:

• Time-Consuming Searches: Finding accurate and relevant information often requires sifting through multiple sources, making the process inefficient.
• Information Overload: Many sources include unnecessary details or lengthy explanations, making it hard to extract the key insights users need quickly.
• Lack of Personalization: Generic responses often fail to consider individual preferences, past interactions, or contextual nuances, leading to less effective communication.
• Limited Context Awareness: Traditional search engines or static FAQ systems struggle to provide responses based on previous queries or user-specific data, requiring users to rephrase or search repeatedly.

These challenges highlight the need for a smarter, more efficient solution—one that leverages AI to deliver personalized, context-aware interactions in real-time. In the next sections, we'll explore how combining the OpenAI API with LangChain can address these issues and revolutionize AI-driven communication.

Architecture Overview

1. Key Components

The chatbot consists of the following core components:

• FastAPI Backend: Manages API endpoints for document ingestion and querying.
• OpenAI GPT Model: Generates responses based on retrieved context.
• Pinecone Vector Database: Stores document embeddings and enables similarity search.
• Text Processing Services: Extracts and preprocesses text from PDFs and text files.
• Logging & Error Handling: Ensures robust monitoring and fault tolerance.

2. Workflow

1. Document Ingestion:
   • Users upload files (PDFs or text) via the FastAPI endpoint.
   • Text is extracted and split into smaller chunks with metadata.
   • Chunks are converted into vector embeddings and stored in Pinecone.
2. Query Processing:
   • The user submits a query.
   • The system retrieves relevant document chunks from Pinecone.
   • The context is refined, and GPT generates a final response.
3. Response Generation:
   • The retrieved context is used to generate a well-informed answer.
   • The response includes relevant references to the source documents.

Setting Up the Environment with Docker

Step 1: Creating the Dockerfile

Create a Dockerfile for the FastAPI backend:

# Use an official Python runtime as the base image
FROM python:3.9-slim

# Set the working directory in the container
WORKDIR /app

# Copy the requirements file into the container
COPY requirements.txt ./

# Install the dependencies
RUN pip install --no-cache-dir -r requirements.txt

# Copy the application code into the container
COPY . ./

# Expose the port the app runs on
EXPOSE 8000

# Define the command to run the application
CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000", "--reload"]

Step 2: Setting Up Docker Compose

Create a docker-compose.yml file to manage the backend and frontend services:

services:
  backend:
    build:
      context: .
      dockerfile: Dockerfile
    container_name: chatbot-backend
    volumes:
      - .:/app
    env_file:
      - .env
    ports:
      - "8000:8000"
    command: uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

  frontend:
    build:
      context: ./chatbot-ui
      dockerfile: Dockerfile
    container_name: chatbot-frontend
    volumes:
      - ./chatbot-ui:/app
      - /app/node_modules
    ports:
      - "3000:3000"
    command: npm start

Step 3: Running the Containers

1. Build and start the containers:

docker-compose up --build

2. Verify that the backend is running at http://localhost:8000 and the frontend at http://localhost:3000.
3. To stop the containers, run

docker-compose up --build

Retrieval and Generation Pipeline

In this document, we will walk through the steps to build a Retrieval and Generation (RAG) pipeline, a key component of many advanced AI systems. This pipeline combines retrieval from a knowledge base (vector database) with generative AI to produce accurate and contextually relevant answers to user queries. Here you can find the repository with a sample of the project we are about to build: Sample AI Chatbot.

Step 1: Set Up the Knowledge Base

The knowledge base is the foundation of the RAG pipeline. It stores relevant information in a vectorized format for efficient retrieval.

• Select a Vector Database: Use a database Pinecone, Weaviate, or Redis for storing embeddings.
• Prepare Data: Collect the data like documents, articles, recipes to be used in the knowledge base and break it into smaller chunks for better indexing.
• Generate Embeddings: Use a pre-trained model like OpenAI’s text-embedding-ada-002 to generate embeddings for each chunk.
• Index the Data: Insert the chunks and their embeddings into the vector database.

index.upsert([ {"id": "1", "vector": embedding, "metadata": {"text": chunk_text}} ])

Step 2: Build the Retrieval Component

The retrieval component searches the vector database for the most relevant chunks based on a user query.

• Embed the Query: Convert the user query into an embedding vector.
• Search the Vector Database: Query the database to retrieve top matches for the query embedding.

results = index.query( vector=query_embedding, top_k=5, include_metadata=True ) 

• Refine the Retrieved Chunks: Optionally combine or summarize the retrieved results to optimize the context passed to the generative model.

Step 3: Connect with a Generative Model

Use a generative model like OpenAI’s GPT to create a final answer by combining the retrieved context and user query.

• Create a Prompt: Build a prompt that integrates the retrieved context and the user’s query. 

Context: <retrieved_context> Question: <user_query> Answer: 

• Call the Generative Model: Use OpenAI's API to send the prompt and generate a response. 

response = openai.ChatCompletion.create( model="gpt-4", messages=[{"role": "system", "content": prompt}] )

• Return the Generated Answer: Send the model’s response back to the user

Steps to Build the Indexing Pipeline

Step 1: Extracting Text from Documents

• For PDFs, use PyPDF2 to extract text while preserving page numbers.
• For text files, read and store content directly.

from PyPDF2 import PdfReader

def extract_text_from_pdf(content: bytes, filename: str) -> list:
    reader = PdfReader(io.BytesIO(content))
    return [
        {"text": page.extract_text(), "page_number": i + 1, "file_name": filename}
        for i, page in enumerate(reader.pages) if page.extract_text().strip()
    ]

Step 2: Splitting Text into Chunks

• Use CharacterTextSplitter to divide documents into manageable chunks.

from langchain.text_splitter import CharacterTextSplitter

def split_document_with_metadata(chunks, chunk_size=1000, overlap=200):
    splitter = CharacterTextSplitter(separator="\n", chunk_size=chunk_size, chunk_overlap=overlap)
    return [
        {"text": split, "page_number": chunk["page_number"], "file_name": chunk["file_name"]}
        for chunk in chunks for split in splitter.split_text(chunk["text"])
    ]

Step 3: Generating Embeddings

• Use OpenAI’s text-embedding-ada-002 model to convert text into embeddings.

import openai

def generate_embeddings(text):
    response = openai.OpenAI(api_key=OPENAI_API_KEY).embeddings.create(
        input=text, model="text-embedding-ada-002"
    )
    return response.data[0].embedding

Step 4: Storing Data in Pinecone

• Upload the generated embeddings along with metadata to Pinecone.

from pinecone import Pinecone
import uuid
 
pc = Pinecone(api_key=PINECONE_API_KEY)
index = pc.Index("knowledge-base")
 
def insert_data_to_pinecone(chunks):
    for chunk in chunks:
        unique_id = str(uuid.uuid4())
        embedding = generate_embeddings(chunk["text"])
        index.insert([(unique_id, embedding, chunk)])

Step 5: Retrieving Context from Pinecone

• Search for the most relevant document chunks using vector similarity.

def search_vector_database(query, top_k=3):
    query_embedding = generate_embeddings(query)
    results = index.query(vector=query_embedding, top_k=top_k, include_metadata=True)
    return [match["metadata"] for match in results["matches"] if match["score"] >= 0.85]

Conclusion

In conclusion, building an AI-powered chatbot like our Recipe Assistant demonstrates how AI can transform information retrieval and enhance user interactions. By leveraging OpenAI API, LangChain, and vector databases like Pinecone, we addressed common challenges such as information overload and lack of personalization. Our architecture, built with FastAPI and Docker, provides a scalable and efficient foundation for AI-driven applications.

While our example focused on a recipe assistant, the same principles can be applied to various domains—from customer support and education to research assistants and beyond. Whether you're building a chatbot for a different use case or enhancing an existing system, the tools and methodologies discussed here offer a flexible starting point. The possibilities are endless—it's up to you to tailor this approach to your unique needs!