Context:
-
We are developing a Retrieval-Augmented Generation (RAG) based application exclusively using Python. The system incorporates various models, including GPT-4, GPT-4 Turbo, and GPT-4o, for tasks such as web scraping and model training. Users can upload sources in the form of documents and websites, which are then stored in a vector database pinecone. We utilize models like text-embedding-large-003 and text-embedding-003-small to generate embeddings from these data sources. From a vectore stores via langchain similarity seach most for similar chunks having size of 1000 are passed to OPEN AI with system prompt & custom prompts.
-
The primary focus of our application is on the e-commerce sector, specifically supporting an e-shop with a wide range of product categories. Each category contains multiple products. The system is designed to provide users with accurate and detailed responses to their queries about products and categories, including relevant purchase or reference links.
-
A default system prompt is bound to each user query to guide the response generation.
Users have the flexibility to set custom prompts, tailoring the response to their specific needs.
Data Specificity:
- The data managed and queried in the system is specific to the e-commerce domain, covering various categories and products offered by the e-shop.
Requirements:
Accurate Query Responses:
- For product-specific inquiries, the response should include detailed information about the product, its price, and a single link to the product page.
- For category-specific questions, the response should provide an overview of the category and a single link to the category page.
- For general or greeting-related questions, responses should not include any links.
Embedding and Data Management:
- Efficiently generate and manage embeddings to ensure quick and accurate data retrieval from the vector database.
- Offer flexibility in embedding sizes (large or small) to cater to different data requirements and optimize performance.
Challenges:
- Ensuring the responses are relevant and accurate, particularly with dynamic data such as product details and availability.
- Balancing the use of different GPT models to maintain efficiency and performance, especially when handling complex queries or high loads.
- Maintaining an updated and accurate database to provide the latest information to users.
Objective:
- To refine the RAG-based application, ensuring it delivers precise and relevant information to users about products and categories in the e-commerce domain. This includes enhancing the system’s ability to provide accurate details, manage data effectively, and generate appropriate responses without irrelevant links.
Request for Feedback:
We are seeking guidance for:
- Best practices for optimizing the accuracy and relevance of query responses.
- Efficiently managing and utilizing embedding models for various data sizes.
- Techniques to enhance the overall user experience and ensure the system’s performance remains robust under varying conditions.
Any insights or recommendations to address these challenges and improve our application would be greatly appreciated.