100% Hallucination-Free AI
Hallucinations in large language models (LLMs) remains a critical barrier to the adoption of AI in enterprise and other high-stakes applications.
Despite advancements in RAG, state-of-the-art methods fail to achieve more than 80% accuracy, even when provided with relevant and accurate context.
Acurai introduces a groundbreaking solution that eliminates 100% of hallucinations in LLM-generated outputs.
The Problem: RAG
Retrieval-Augmented Generation (RAG) was introduced in early 2021 to improve accuracy. RAG was a major breakthrough. RAG achieved a 44.5% accuracy on the Natural Questions benchmark.
Nvidia researchers released OP-RAG in September 2024 to "achieve higher answer quality." OP-RAG achieves an F1 score of 47.25%. However, it requires sending almost 50,000 characters of text to the LLM along with each query. In other words, to achieve a 50-50 chance of getting a correct answer, the LLM must process 50,000 characters for every question asked.
Outside Acurai, the highest performing RAG involves using OpenAI's o1 model. This RAG implementation achieves just over 80% accuracy. However, the system requires sending over 64,000 characters of context along with each query. Sending less results in lower accuracy. For example, sending only 2,000 characters of context results in 58% accuracy (just barely better than flipping a coin). OpenAI's o1 model is extremely slow. Having to process 64,000 characters per question on o1 is unbearably slow. In return for this slowness, companies can get four out of five questions answered correctly.
The most widely touted form of RAG today is called "Advanced RAG." Advanced RAG consists of combining vanilla RAG with dozens of additional methods. In July 2024, Fudan University researchers published "Searching for Best Practices in Retrieval-Augmented Generation." They tested various combinations of Advanced RAG methods to find the optimal implementation. The optimal Advanced RAG combination achieved a 44.6% score — basically the same score as the original RAG introduced in early 2021!
Summary: RAG's retrieval has not improved over the last four years. In other words, there hasn't been any significant achievements on pinpointing the precise chunks that contain the answer. On the contrary, the improvements in accuracy have come from sending an increasing numbers of chunks, due to LLMs' ability to process longer amounts of context. In fact, virtually all the "improvement" comes from sending tens of thousands of context characters along with each query, tasking the LLM to find the correct needles in the haystack. What's been missing is the ability to locate the exact information needed, and only sending that precise information to the LLM. That's the missing key to creating 100% accurate RAG-based chatbots.
Key Challenges:
-
Relevancy: RAG systems struggle to find the relevant information to send to the LLM. Notice that the "improvements" in RAG come from sending more information to the LLM (so that the LLM can try to find the right answer). Yet, even when sending 64,000 characters of context, the relevant information is often still not included.
-
Accuracy: LLMs can hallucinate, even when relevant context is provided along with the query. (See video above.) Thus, accuracy is an additional challenge beyond finding the relevant content.
-
Speed: Outside Acurai, RAG vendors commonly promote "Advanced RAG" — RAG combined with dozens of methods. However, each additional method introduces additional slowness. In fact, the Fudan University researchers noted that the mere combination of BM-25 + Hyde took 11.71 seconds per query.
The Solution: AcuRAG
RAG essentially consists of two components: (R)etrieval and (G)eneration.
On July 30, 2024, Acurai released the first 100% accurate Generator. The 100% accuracy was validated on four datasets from the RAGTruth Corpus. (See paper.)
Acurai, our 100% accurate Generator, ensures that the answers match the provided the context without any deviation whatsoever. Acurai "sticks to the script" as we like to say. By ensuring that responses match the context, hallucinations are fully eliminated — as long as the provided context itself is truthful. Which brings us to our Retrieval breakthrough.
Acurai has recently filed a patent application on the first 100% relevant Retriever. We call this AcuRAG. We are currently validating our innovative Retriever on a popular RAG dataset. We shall publish the resulting 100% accuracy on this large RAG dataset. We will also be hiring third-party auditors to conduct the same assessment to provide independent verification — independent verification that the issues of 100% accuracy and 100% relevancy are finally solved.
Importantly, we have also recently had breakthroughs in regards to speed. As stated above, Advanced RAG implementations can be extremely slow. Not only has Acurai solved for 100% accuracy and 100% relevancy, but we have solved the issue of speed as well. Once again, we are going to empirically demonstrate the accomplishment. We will be soon be releasing our free AI Wikipedia service called Wikipedia Portl. Wikipedia Portl demonstrates instantaneous, 100% accurate, 100% relevant responses on over 6 million documents. Unlike our competitors who hype non-solutions, our real-world solutions speak for themselves.
Acurai's AcuRAG is the long-awaited solution that companies have been searching for. No hype. Just empircal, documented proof.
Key Innovations:
-
1.Relevancy: Traditional RAG struggles to find the relevant information. Traditional RAG attempts to compensate by sending literally hundreds of chunks of information to the LLM, hoping the relevant information is contained somewhere in the tens of thousands of characters. In stark contrast, AcuRAG pinpoints the precise relevant statements found within millions of documents. Therefore, the Generator only receives the precise information it needs to answer the query.
-
2.Accuracy: The Acurai Generator receives the precise information from AcuRAG. It's patent-pending methodology ensures that the response matches the provided information with 100% accuracy — 100% accuracy already demonstrated on four datasets from the RAGTruth Corpus.
-
3.Speed: Traditional RAG typically involves a full search using a vector database combined with BM-25 search. Knowledge Graphs are someimes added in as well. Acurai is different. We do not use BM-25 nor Knowledge Graphs. The heart of our search is accomplished through applied set theory on in-memory (i.e. in RAM) sets — resulting in instantaneous results. We do use vector embeddings, but in a unique way. Vector embeddings are used after the search is completed. This means that usually less than 0.01% of the existing vector embeddings are even consulted during any given query. We do not use vector embeddings as a gross search mechanism, but solely as a refinement to the in-memory set-theory-based search — resulting in unparalled speed.
Breakthrough Results
RAGTruth Corpus
Partner with us
If you have a need for hallucination-free AI, please fill out the short form below.
As you might imagine, we currently have a waiting list. We raised a $2M pre-seed round in October 2024. We are releasing empirical documentation of our results in early 2025, as well as releasing Wikipedia Portl which demonstrates all three levels of innovation applied to 6 million documents.
The same code used for Wikipedia Portl can be applied to creating RAG-based chatbots for your organization — chatbots with unparalleled accuracy, relevancy, and speed. In other words, it's everything you need.
In the meantime, we would be very interested in hearing about your project and the problems you are experiencing with hallucinations, as your project may dovetail with our own early-access development timelines.
Also, we will add you to our newsletter list & send updates as we more publish papers, etc.
Welcome to the revolution!
Michael & Adam
Co-Founders, Acurai