Imani is an Enterprise Data & AI manager at a major manufacturing company. Her company is at the forefront of integrating AI into its processes and systems. She successfully led a project to build and deploy RAG document search with detailed provenance and sophisticated data parsing, and she iterated on the solution until it accurately answered many workers’ questions. Today, however, Miles, a maintenance planner, asked the RAG: “Why did this machine break for the third time this month?” The system responded with some generic troubleshooting guides. His feedback: “I thought this fancy search system was supposed to be AI?! But it really has no idea what I mean and gives useless answers to my questions.”

Why couldn’t RAG search answer Miles’ question?

Imani’s team definitely built a powerful AI search system. However, Miles' question goes beyond the capabilities of what most RAG systems can do. Here are the reasons why the AI could not answer his question:

• No understanding of user context: The AI has no idea what Miles means by “this machine”. It does not understand who he is or what his current context is, so it cannot answer with anything other than the most generic information.
• No access to related work records: Even if the AI understood context, it does not have access to a structured list of work records that might help answer Miles' question. These work records might be stored somewhere in the unfathomable depths of SAP, and are not easy to access or integrate with.
• No understanding of ontological relationships: Even if the AI had access to work records, it has no understanding of how the records relate to machines or people in the manufacturing company. A record might contain an ID reference, but this structured data is invisible to a RAG system that is only designed to search unstructured information.
• No understanding of time: Miles references “this month”. The RAG system doesn’t know what the current time is in relation to the documents in the database powering the RAG.
• No ability to conduct analysis: Analysis requires a series of reasoning steps. That is beyond the capabilities of the RAG system which merely answers questions based on retrieved document chunks.

How does Datch solve this?

Datch’s Asset Insights AI addresses these challenges with a comprehensive approach, which includes the following steps:

• Ontology design: We work with customers to build a custom ontology for their domain. This ontology gives the AI an understanding of the users' context and nomenclature. That is, using the ontology, the Large Language Model (LLM) can understand the user’s terminology and how different concepts relate to each other.

‍

• Ingestion: Our ingestion data pipeline integrates with multiple third-party systems and ingests information into a knowledge graph. The ingestion system finds explicit and implicit information from structured (e.g. CSVs, Database dumps) and unstructured (e.g. PDFs, Word Documents, Tables, Images) sources, and connects it all together into a graph structure.
• Enrichment: We run custom classifiers to enrich records in the knowledge graph with specific codes and standards. This greatly helps the AI perform aggregation and analysis of the data.
• Ask anything: Workers like Miles can ask the LLM complex questions focused on a specific context and get meaningful answers. The questions do not need to match to specific hand-coded pathways.
• LLM agent: A LLM agent can use the knowledge graph as a tool to help answer the worker's questions. The agent has an understanding of the user’s domain and context, and can use that knowledge to traverse the knowledge graph, collecting relevant information along the way. Then it combines the collected information into a comprehensive answer.
• Provenance: LLMs can suffer from hallucinations. These hallucinations are greatly reduced when the LLM bases its answer on specific authorized knowledge, but users must have the ability to verify answers for themselves. Datch’s Asset Insights AI therefore always references all the sources in its chain of reasoning, and allows the human to retrace the AI’s steps.
• Optimization loop: Sometimes a user will ask a question for which the system cannot answer because it lacks the required source information, or because is lacks the correct tool to answer. In that case, Datch works with customers to monitor the active system, identify any failure cases, fix the problem, and continually verify that the AI’s answer remains accurate over time.

GraphRAG vs Datch Asset Insights

Microsoft has a GraphRAG research project, and Meta’s LlamaIndex also supports GraphRAG. These are both useful systems but they are not the same as Datch’s Knowledge Graph AI.

GraphRAG in both Microsoft and Meta’s technologies pre-processes unstructured documents to build a knowledge graph. When a user asks a question the GraphRAG can pull in additional related context from that knowledge graph to help answer the question. That additional information, however, is still sourced from the same original unstructured data. Just from a different place in that data.

Datch’s Asset Insight product - powered by what we’ve coined as DatchRAG technology - is differentiated in 3 key ways:

1. It integrates structured and unstructured information together
2. It enriches data further upon ingestion, creating new, related datapoints/nodes on the graph, and
3. It offers all the added benefits of the customer-specific solution outlined earlier in this article.

‍

‍

Interested in seeing a demo of how this works for yourself? Drop your email in the box below and we’ll be in touch ASAP.