Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval

2026-01-26 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A novel system introduces "Executable Schema Contracts" to automate question answering over heterogeneous, multi-source data like tables, documents, and JSON files. This system automatically discovers a unified, executable schema from raw data, which then acts as a shared contract for knowledge graph construction and query-time retrieval. It employs a closed-world field catalog to constrain LLM-based schema discovery and uses deterministic structural analysis to infer identity and foreign keys. The schema guides data extraction, deduplication, and cross-source linking into a provenance-aware Neo4j knowledge graph. At query time, the schema conditions a multi-tool agent, routing queries across structured lookup, graph traversal, and vector search. In controlled zero-shot comparisons, the system improved Exact Match (EM) by +2.6 to +19.8 points over baselines across four QA benchmarks, including BlendQA, HybridQA, TAT-QA, and ComplexTR, showing consistent gains across GPT-4.1, Claude Haiku, and Llama 3.3 70B models.

Key takeaway

For AI Engineers and Architects building robust question-answering systems over diverse, evolving data, you should prioritize adopting schema-guided approaches. This system demonstrates that automatically induced, executable schemas significantly improve accuracy for cross-source joins and multi-hop retrieval by providing a unified contract for ingestion and query routing. Implement a closed-world field catalog to ground LLM-discovered schemas and leverage statistical analysis for structural intelligence. While query-time schema extension can add ~3.4 seconds latency, the overall gains in accuracy and auditable provenance make this a compelling architecture for complex data environments.

Key insights

An automatically induced, executable schema can unify knowledge graph ingestion and multi-source query-time retrieval.

Principles

• Closed-world schema induction ensures LLM-discovered structure is executable.
• Prioritize statistical structural analysis before semantic extraction.
• Schema-conditioned routing significantly improves retrieval agent performance.

Method

The system couples schema profiling, LLM extraction, KG construction, and vector retrieval, using a single induced schema to constrain extraction, link entities, and guide agent routing.

In practice

• Automate schema discovery and maintenance for evolving data.
• Enhance cross-source joins and multi-hop QA with auditable provenance.
• Ground LLM schema outputs using a closed-world field catalog.

Topics

• Executable Schema Contracts
• Knowledge Graph Construction
• Multi-Source Retrieval
• LLM-based Schema Discovery
• Question Answering
• Retrieval-Augmented Generation

Code references

• snap-stanford/STaRK

Best for: Research Scientist, Machine Learning Engineer, NLP Engineer, AI Scientist, AI Engineer, AI Architect

Related on AIssential

• Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval — The system automatically derives an executable schema from multi-source data to unify knowledge graph construction and multi-tool query retrieval.
• SAGA: Schema-Aware Grounding for Agentic Text-to-SPARQL Generation — Schema-aware grounding significantly enhances Text-to-SPARQL generation by preventing type-blind query construction in LLM agents.
• 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.
• LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining — The LLM-INSTRUCT system combines retrieval, constrained decoding, and multi-agent debate for accurate, constrained argument mining.
• MARS: Multi-hop Adaptive Retrieval and SPARQL Generation for KGQA — MARS enables reliable KGQA by combining LLMs with structured, adaptive multi-hop retrieval over KGs without fine-tuning.
• Team Aurevia at CLPsych 2026: Local Healthcare NLP for Schema-Constrained Self-State Modeling — Constrained local LLM pipelines can achieve competitive performance in sensitive healthcare NLP tasks.

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 cs.AI updates on arXiv.org.