GraphRAG - Pilot program

Generally, a pilot program takes about 12 weeks, from the initial outlining of the project through development and deployment. It generally works as follows 👇

Weeks 1-4

We begin by defining the project needs, detailing the current problem, who it affects, and the outcomes that will constitute success. With this done, we work with you to collect all of the data sources that they wish to include in the completed interface.

Once that data is collected, our NLP technology can begin the hard work of properly parsing it, ensuring that unstructured data becomes structured in order for it to be correctly interpreted by a machine. This gives us the basis for an enriched ontology and the initial knowledge graph, using the client’s data as well as Lettria’s existing resources.

With structured data and the initial ontology and graph in place, we can move onto testing your custom GraphRAG solution.

‍Weeks 5-8

The first prototype begins to be tested by both you and Lettria’s team of experts, including data scientists, lexicographers, technical directors, and project managers. This happens in a series of feedback loops, as questions are posed and answers are evaluated according to a matrix (correct, partially correct, partially incorrect, incorrect). Each session lets us note various improvements that then go into the next iteration.

Throughout, the goal is to continue enriching the ontology and improving relationship graphs in order to confidently deploy our GraphRAG technology for a given use case. This second phase brings us to a solution that is ready for final testing and deployment.

‍Weeks 9-12

The final iteration is completed. Lettria’s engineers work to produce the appropriate API documentation, as well as to complete the interface within the client’s information system. We assist with the rollout and make sure users are comfortable using the new tool.

The end result is a verifiable, trustworthy AI application running within your company. There’s no black box results, no hallucinations, no misreading of charts and data. Query results come back quickly, and are displayed alongside their knowledge graph and the context from which they were found.