Our Team

Vectorial consists of several software, AI, and ML engineers. For over 2 years, we have implemented advanced LLM and ML solutions (research, features,  and fully fledged products) for our customers.

We are currently working on DeepFunding RFPs for “Evolving DNN Architectures” and “Utilize LLMs for modeling within MOSES” and have made considerable early progress. Our team has additional bandwidth and is excited to take on this RFP as well.

Company Name (if applicable)

Project details

The goal of this project is to create a framework that can construct and expand knowledge graphs on the Hyperon platform and, secondly, utilize these knowledge graphs during large language model (LLM) inference to improve factual grounding during responses. We plan to construct a knowledge graph from an initial knowledge base, such as Wikipedia, and expand upon it via an iterative intelligent search into more detailed source materials, such as textbooks and research papers. The expanded graph will serve as a source of reliable truth to reduce the rate of LLM hallucinations during AGI reasoning tasks requiring multi-hop retrieval and other forms of semantically complex question answering.

We plan to begin with narrower domains, such as artificial intelligence and machine learning, to prepare for later expansion into broader categories. While the overall technique will remain domain independent for testing and initial development, we will meet memory pressure and capacity constraints using narrow example domains as proofs of concept.

For example, we can begin at the Wikipedia page for Machine Learning, using it to instantiate a representative node to serve as the initial anchor for a knowledge graph within MORK. Then, we create a series of edges, annotated with perceived semantic relationships within the anchoring document (e.g. links to other Wikipedia pages that exist within the Machine Learning page). Non-machine learning related information can be filtered out by asking an LLM whether or not the content contains relevant information to machine learning. This process will be performed recursively until all information on Wikipedia that relates to machine learning is captured.

After this first stage, we can begin incorporating less structured knowledge bases, such as arXiv papers or textbook datasets. By chunking and vector embedding these knowledge bases, we can use Agentic RAG (allowing LLMs to dynamically search for information) with ranked retrieval to find knowledge in the external data sources. All information identified as new will be dynamically added to the knowledge graph. We will use LLMs to determine if there are any facts present in the queried information sources which are not present in the current knowledge graph, by allowing the LLM agent to “walk” the existing graph in search of existing information. We intend to implement this walk using MeTTa to allow the agent to query the knowledge representation database within MORK programmatically, and then use the reasoning capacity of the LLM to interpret relationships and identify missing facts. This is also the interface by which the LLM will be able to update and reason with the knowledge graph.

Next, to demonstrate the effectiveness of Knowledge Graphs in improving factual grounding of LLM responses, we will conduct a series of experiments using common Q&A benchmarks, including the OpenAI SimpleQA benchmark, high-level standardized testing formats such as FE and GRE science exams, and multi-hop Q&A benchmarks. Our intention is to use benchmarks which cover a diverse set of topics and difficulty levels. We will compare performance across at least three model treatments: without access to a Knowledge Graph, with access to an initial Knowledge Graph, and with access to the complete Knowledge Graph enriched with external sources using our method. We plan to demonstrate the capability of our approach by generating a knowledge graph based on each topic in the benchmarks, and asking each LLM-knowledge-graph pair questions from its respective topic.

Finally, we can perform Reinforcement Learning to improve the LLM’s ability to search the knowledge graph, rewarding successful fact retrieval and subsequent reasoning. We can perform additional experiments such as fine-tuning the LLM with SFT or LoRA for domain-specific improvements. We will also experiment with a multi-agent approach that seeks to repeat the process across multiple nondeterministic walks, democratizing the response to achieve improved confidence in truthfulness by way of majority consensus.

Open Source Licensing

Background & Experience

Our expertise originates from several advanced degrees in Data Science, Computer Science, Mathematics, and Machine Learning. Our team also has working domain experience at companies such as AWS, Oracle, Home Depot, and Fortune 50 financial corporations. Team members have completed a series of relevant projects regarding document retrieval, factual grounding, named entity & relationship extraction, multi-hop question answering, and graph/topology algorithms including search, reachability, metagraph application, and network flow.

At Vectorial, our team has collaborated for over 2 years to build several AI agent systems for our clients using advanced generative AI techniques. We are well-acquainted with the state-of-the-art methods used to improve LLMs and plan to use this experience to create a powerful knowledge graph framework that will scale along with AGI systems on the Hyperon platform.

Links and references

https://github.com/PatrickNercessian

https://github.com/PatrickNercessian/evolving-dnn

https://pjnercessian.wordpress.com/2023/11/07/drift-correction-robotic-computer-vision/

https://www.linkedin.com/feed/update/urn:li:activity:6755149565814480896/