Contoso’s legal department frequently updates contract templates and compliance playbooks. You propose a grounded legal copilot. One stakeholder…

The statement is incorrect and misaligned with business requirements for a dynamic domain. Large Language Models (LLMs) do not learn or permanently update their knowledge from user prompts during inference. The grounding data, stored in an index, must be regularly updated to reflect changes in the source documents (e.g., new legal templates). Failing to refresh the index means the copilot will provide answers based on outdated, non-compliant information, creating significant business risk. The core value of grounding is providing current, contextually relevant data at query time.

Assuming the options were A) Yes, this statement is aligned, and B) No, this statement is not aligned.

(Assuming A = “Yes, aligned”)

A. Treating a one-time grounding as sufficient ignores data freshness and raises compliance risk; not aligned with legal requirements.

1. Microsoft. (2024). Retrieval-Augmented Generation (RAG) design pattern. Azure Architecture Center. The architecture explicitly shows a "Data indexing" process. "The data indexing process needs to be run whenever the grounding data changes." (Section: Data ingestion and indexing).

2. Gao, Y., et al. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv preprint arXiv:2312.10997. "A crucial aspect of RAG is the maintenance of the external knowledge base. As information evolves, the knowledge base must be updated to ensure the generated responses remain accurate and relevant." (Section 5.2, RAG Evaluation).

3. Stanford University. (Course CS224N). Lecture on Pretraining and Transformers. Course materials explain that model weights are fixed after pretraining/fine-tuning and are not updated during standard inference, which contradicts the stakeholder's claim of "learning from prompts."