In November 2022, OpenAI unveiled ChatGPT, an AI chatbot that impressed many with its human-like conversation skills. Roughly one year later, on November 6, 2023, OpenAI held its first ever keynote event. The announcements made during the keynote generated a similar level of excitement as ChatGPT's launch. The most interesting of the announcements are OpenAI's retrieval tool and GPT-4’s 128k context length update. Before getting into OpenAI’s retrieval, we should discuss Retrieval Augmentation Generation (RAG).
What is Retrieval Augmented Generation (RAG)?
Retrieval augmentation is a technique that enhances the ability of a large language model to generate relevant and accurate responses by using an external source of information. For example, if you ask a question that requires some factual knowledge, such as “What is the state of AI after Biden’s Executive Order”, a large language model might not have the correct answer stored in its parameters, or it might have outdated information. The advantage of RAG is that we can update the models knowledge without having to finetune the model. Examples of RAG pipelines are using vector databases such as Qdrant Pinecone, and ChromaDB. With the use of embedding model, data such as audio, text, and images are converted into numerical representations of data. After the data conversion, the data is then stored in vector stores.
One implementation of RAG is key word searches. Key word searches are a common way of using an information retrieval system to find relevant documents. A key word search is a query that consists of one or more words that describe the topic or information that the user is interested in. The information retrieval system then matches the key words with the documents in its database and returns the most relevant ones to the user based on the semantic meaning. When a user makes a semantic query, the search system works to interpret their intent and context.
OpenAI's Answer to Retrieval Augmented Generation (RAG)
OpenAI’s retrieval system has tech startups scrambling because of the ease of integration without relying on external databases. Retrieval is a tool that allows Assistants to access external sources of information, such as documents, websites, or databases, and use them to generate more relevant and accurate responses. Furthermore, the retrieval tool system has a built-in chunking system. Based on the document, the retrieval tool will select the optimal chunking system-based on the document. In addition, users can select GPT-4 128k to assist with the document analysis.
A high context-length GPT-4 and optimization algorithm is ideal on paper. However, after experimenting with OpenAI’s retrieval tool, several issues were spotted. The system is rather dull and uninspiring. It can parse data from individual files, but it does not support multiple data sources, which is a limitation for many users. Another issue is the data limitations. Users are restricted to their own data and may need other sources to optimize their research. The retrieval tool doesn’t have access to external databases or the internet. Users have the option of using OpenAI function to integrate external tools, but this is not ideal for research purposes. Research should focus on the core problems, not on adjusting inadequate research tools.
Furthermore, the document parsing system spreads itself too thin. A proper RAG system is unique to the data and its domain. GPT-4’s output may not meet the demands of users and may require changes. The retrieval doesn’t offer modifications unless users ask GPT-4 to try again. However, that requires token cost, and specific prompt engineering. However, GPT-4's output may still not meet the demand of users. In addition, the Assistant's file management system isn't user friendly. Documents such as PDFs are lumped in with the fine-tuning dataset. Users will get overwhelmed if they need assistance on a wide range of reports.
Silatus' Answer to Retrieval Augmented Generation (RAG)
Here at Silatus, we optimize the research experience. If researchers don’t have enough information on a topic, Silatus can help them. Silatus has access to external databases and the web to cover all our users’ questions. Furthermore, Silatus provides a file management system for our users. This system organizes the users’ documents and the documents from the generated search engine. Users can also interact with the integrated database and use features such as question and answering, summarization, and document generation.
In addition, Silatus has a variety of templates that users can choose from to create different types of documentation, such as reports, proposals, summaries, or reviews. These templates are designed to suit the needs and preferences of different users and domains. Users can browse through the templates and select the one that best matches their purpose and audience. Users can also customize the templates by adding, deleting, or modifying the sections, headings, and content. Silatus makes it easy for users to edit their documents without worrying about the cost of tokens. Users can use the editing features, such as spelling and grammar check, formatting, and style, without spending any tokens. Silatus helps users create high-quality documentation with minimal effort and expense.
While OpenAI's retrieval tool and upgraded GPT-4 model show promise, they have some limitations that prevent them from fully meeting researchers' needs. Silatus offers a more robust research platform with access to external data, customizable templates, and unlimited editing features. By leveraging Silatus' strengths, researchers can conduct comprehensive analysis and efficiently generate high-quality documentation tailored to their goals. With continued innovation, AI will become an increasingly asset for streamlining research workflows. However, choosing the right platform is crucial to maximize productivity. Silatus stands out for its versatility, customizability, and focus on optimizing the research experience.
