Description: Retrieval-augmented generation (RAG) is a common method utilized in generative AI to inject authoritative truth on relevant topics into model query prompts such that the large language model (LLM) need not rely on its learned understanding of the subject matter. This approach can help reduce model hallucinations, wherein the LLM provides incorrect or made-up responses to queries. It can also enable models to provide utility in domains outside the scope of their training data, which often consists of open data available on the internet. For both cases, RAG inference pipelines rely on the retrieval of high-quality source material for the LLM response; asking a model a question it has insufficient information to answer will increase the probability of LLM hallucination.

This lecture will cover in depth the underlying processes used in context retrieval within RAG pipelines and demonstrate how open-source Transformer encoder models can be fine-tuned for domain-specific retrieval tasks to maximize the relevance of the context provided to the LLM for specific domains. Additional methods of search will also be discussed as means of improving context retrieval beyond semantic search. Finally, the effect of context retrieval on text generation will be demonstrated and discussed.

Presenter(s): Dr. Matthew Boyer, GDIT/PET Location: Webcast Date & Time: May 13 & 15, 2025, 2:00p - 3:30p ET

Controlled by: DoD HPCMP Controlled by: PET Program CUI Category: OPSEC Limited Dissemination Control: FEDCON POC: Mr. Ronald Hedgepeth, pet@hpc.mil

---

CUI