New approach helps analysts explore databases through guided natural language queries

Natural language interfaces have opened up database exploration to analysts without SQL expertise, but translating a broad research goal into a coherent sequence of queries remains challenging—especially when working with unfamiliar, multi-table relational databases. A new research effort tackles this bottleneck by augmenting such interfaces with smarter query recommendations.

The Recommendation Problem#

Analysts often struggle to decompose complex questions into step-by-step queries that build coherently toward an insight. While large language models have made natural language access to data more accessible, they don't inherently guide users toward topically focused investigations or help them discover promising data subsets in an organized way.

A Richer Approach to Next-Query Suggestions#

Rather than simply flagging "interesting" patterns, the proposed system—called QRec-NLI—weaves together three complementary signals to recommend the next query: semantic relevance (whether the suggestion relates to the current analytical topic), data interestingness (whether it uncovers noteworthy patterns), and context coherence (whether it logically follows from prior queries). This balanced combination acts as analytical scaffolding, helping explorers stay focused while remaining open to discovery.

Validation and User Impact#

The team evaluated QRec-NLI across multiple dimensions: standard NL2SQL benchmarks, LLM-based validation of query descriptions, simulated agent comparisons, and a 12-person user study. In head-to-head trials against a baseline that relied only on interestingness rankings, QRec-NLI generated query sequences that users rated significantly higher for supporting insights and decision-making. Agentic tests confirmed it produces more topically aligned and locally coherent sequences than both interestingness-only and generic LLM-based prompting approaches.

Why It Matters#

As organizations increasingly adopt conversational interfaces for analytics, the ability to guide exploration intelligently—rather than leaving users to navigate alone—could accelerate discovery and reduce analytical dead ends. The work suggests that combining semantic and context signals, not just statistical novelty, is key to making database exploration feel purposeful and productive.