Giving a Language Model the Freedom to Interpret Without the Authority to Be Wrong

2026-07-12 · Source: Artificial Intelligence on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

A natural language search system has been developed for a fraud case management tool used by banks and NBFCs. This system allows users to query dynamic, reconfigurable schemas, often containing 300-400 fields per tenant, without direct database interaction. Unlike typical text-to-SQL solutions, which risk invented column names or faulty joins, this approach separates probabilistic LLM interpretation from deterministic execution. The LLM generates a structured "intent" object, which is then rigorously validated against real schema metadata. Key techniques include retrieval-augmented generation (RAG) using PostgreSQL vector embeddings for efficient field discovery, narrowing 300-400 fields to about 15 candidates, and trigram similarity for accurate value lookups like user names or groups across multiple tables. A two-layer validation process, both pre- and post-generation, ensures that only confirmed fields and values are used, preventing LLM hallucinations from impacting query execution and maintaining compliance.

Key takeaway

For AI Engineers building natural language interfaces over dynamic or sensitive enterprise data, directly generating SQL from LLMs is a critical compliance risk. Instead, you should architect systems that strictly separate probabilistic interpretation from deterministic execution. Have your LLM produce a structured "intent" object, then validate every field and value against pre-retrieved, confirmed schema metadata and real data. This ensures trustworthiness and prevents silent, incorrect results in regulated environments.

Key insights

Separate probabilistic LLM interpretation from deterministic execution to ensure trustworthy natural language search over dynamic schemas.

Principles

• Separate interpretation from execution.
• Match tool to sub-problem shape.
• Validate LLM output deterministically.

Method

The system uses retrieval-augmented generation (RAG) with vector embeddings for field discovery and trigram similarity for value lookups, followed by a two-layer validation of the LLM's structured intent.

In practice

• Embed field names with context.
• Use trigram similarity for value lookups.
• Implement two-layer output validation.

Topics

• Natural Language Search
• Dynamic Schema
• Retrieval-Augmented Generation
• LLM Hallucination Prevention
• Text-to-Intent
• Enterprise Software

Best for: AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• AI Is Changing Schema Matching in Ways Rule-Based Systems Couldn't — LLMs enhance schema matching by enabling semantic reasoning and transformation logic generation, overcoming rule-based system limitations.
• Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments — Constrained LLM generation can achieve semantic correctness by dynamically adapting grammars to runtime environments.
• To Describe or Not to Describe? Benchmarking Database Representations for Schema Linking in Text-to-SQL — LLM-generated natural language descriptions enhance Text-to-SQL schema linking more effectively than DDL/DML.
• Why NL-to-SQL Systems Fail: Lessons from building a Semantic SQL validation pipeline — Semantic correctness and robust validation are paramount for reliable NL-to-SQL systems, beyond mere syntax.
• Code, Chaos, and the Invisible Logic of LLMs — LLMs predict words probabilistically, not with human intention, leading to potential misunderstandings in code generation.
• Natural Language to SQL Over an Unknown Schema — A multi-stage NL2SQL pipeline ensures safety and accuracy for operational databases by constraining model output and validating execution.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence on Medium.