tanmaydesh5189

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.

Many enterprises are worried that if they use pre-trained models to improve their organization's productivity, the training data used may be copyrighted. The other options could be using your own data/code reops to train these models, but if you don't have rich high-quality data/code repos then the model trained will produce suboptimal results anyways. Is there any better solution?