# Design a Grounded LLM Assistant for Case Preparation Design an LLM-based assistant that helps a user prepare a draft response from documents related to a legal dispute. The system may summarize documents, identify claims and supporting passages, organize a timeline, and draft text for review. It must not present itself as a lawyer or treat generated text as authoritative legal advice. Cover notebook-style interaction, ingestion, retrieval, grounding, prompt and tool boundaries, evaluation, privacy, and human review. The assistant can answer questions but must not autonomously file, send, or submit a response. ### Clarifying Questions to Ask - Which document types and jurisdictions are in scope? - Who is authorized to upload and view each matter? - Must every factual statement cite a source passage? - What retention, deletion, and audit requirements apply? - Which actions always require a qualified human reviewer? ### What a Strong Answer Covers - Matter-level access isolation and encrypted document storage - Parsing with provenance down to page or passage - Retrieval scoped to the current matter and resistant to prompt injection in documents - Draft generation that separates sourced facts, user assertions, and uncertainty - Citations that are mechanically checked against retrieved passages - Human approval, version history, and no autonomous external action - Evaluation for faithfulness, omission, citation quality, privacy, and unsafe overclaiming - Redaction, retention, deletion, incident response, and audit controls - Clear product warnings without relying on warnings as the only safeguard ### Follow-up Questions - How would you detect a citation that does not support the sentence? - What if two uploaded documents contradict each other? - How would you defend against malicious instructions embedded in evidence? - Which data should never be used for model training?