Ask HN: What data modeling approaches work for convs. AI systems?
I’ve been working on data modeling for conversational AI / LLM-driven systems and wanted feedback from people building or operating these systems in practice.
Based on recent work, a few approaches that seem to help (and their limits):
Semantic-first models: treating intent, entities, and relationships as first-class objects rather than forcing everything through star schemas
Hybrid structured + retrieval layers: combining strict schemas for facts with embeddings for discovery, at the cost of more complex orchestration
Query mediation layers: translating natural language into constrained query plans instead of free-form SQL or retrieval
Explicit conversational state: modeling context and history as data, not just prompt text
Evaluation beyond accuracy: measuring conversational drift, ambiguity resolution, and recovery paths
I’ve written up these ideas, trade-offs, and examples here (this is a Medium Friend Link, so it should open fully without a paywall):
https://medium.com/data-science-collective/how-to-build-data-models-that-actually-work-for-conversational-ai-in-2026-67d16f261344?sk=8f0f64875ec5e4c26493f6fb207938ec
What I’m hoping to learn from this community:
Which of these approaches hold up in production, and which fall apart?
Are there modeling patterns you’ve found simpler or more robust?
What failure modes show up only at scale or with real users?
Anything here that feels over-engineered or missing entirely?
Looking for concrete experiences, counter-examples, and corrections.
No comments yet.