---
title: "One Auto-Generated Schema Beats RAG by 51 EM Points on Multi-Source AI Queries"
subtitle: "Researchers propose using an auto-discovered schema as a shared contract for both knowledge graph ingestion and multi-source QA retrieval, beating RAG and agentic baselines across four benchmarks."
url: "https://alphasignal.ai/news/one-auto-generated-schema-beats-rag-by-51-em-points-on-multi-source-ai-queries"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-06-03T00:00:00.000Z"
updated: "2026-06-11T05:41:54.632Z"
type: "Paper"
topics:
  - "Retrieval"
  - "Agents"
subtopics:
  - "Rag"
  - "Document Understanding"
  - "Agent Frameworks"
upvotes: 1434
source_url: "https://arxiv.org/abs/2606.05415"
comments: 0
truncated: true
---

# One Auto-Generated Schema Beats RAG by 51 EM Points on Multi-Source AI Queries

> Researchers propose using an auto-discovered schema as a shared contract for both knowledge graph ingestion and multi-source QA retrieval, beating RAG and agentic baselines across four benchmarks.

Asking a question that spans a company's internal spreadsheets, PDFs, and JSON logs sounds simple. In practice, it is one of the hardest problems in applied AI. [A new research paper](https://arxiv.org/abs/2606.05415) from Jonnalagedda, Yao, Gao, Hasson, and Das proposes a principled fix: automatically discover a unified schema from raw heterogeneous data, then use that same schema as a binding contract for both building a knowledge graph and routing queries at inference time.

## The problem RAG quietly ignores

Standard RAG works well when your data lives in one place and questions are self-contained. Real-world data spans tables, documents, and semi-structured files with implicit semantics, and querying it requires integrating evidence across inconsistent schemas and formats -- yet existing approaches either demand costly manual engineering or bypass structure entirely. The result is a class of failures that pure vector retrieval cannot fix: cross-source joins, multi-hop reasoning, and answers that need traceable provenance back to specific source records.

Four recurring bottlenecks motivate the work. Domain experts must predefine entity types and extraction rules per dataset, repeating this as sources evolve. Static schemas go stale while fully open-ended approaches sacrifice correctness. Answers must be traceable to source records. And multi-source questions require choosing retrieval strategies and bridging sources via joinable identifiers -- without explicit structure, routing is brittle.

## One schema to rule them all

The system automatically discovers an executable schema from raw multi-source data and uses it as a shared contract for knowledge graph construction and query-time retrieval. The key insight is that the same artifact -- a YAML schema file -- simultaneously constrains what the LLM can reference during ingestion, determines how entities are extracted and linked, and guides where the agent looks at query time.

![End-to-end pipeline architecture showing heterogeneous data sources feeding into auto schema generation, knowledge graph ingestion, and schema-aware QA agent](https://arxiv.org/html/2606.05415v1/x1.png) 

A closed-world field catalog constrains LLM-based schema discovery to attested fields; deterministic structural analysis infers identity keys, foreign keys, and source hierarchy; and the resulting schema drives extraction, deduplication, and cross-source linking into a provenance-aware knowledge graph. The "closed-world" constraint is critical: before any LLM call, the system enumerates every observable field in each source and assigns it a stable hash-based ID. The LLM may only reference IDs that exist in this catalog -- hallucinated field references are rejected or repaired before execution.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval (arXiv)](https://arxiv.org/abs/2606.05415)
- [Executable Schema Contracts: Full HTML Paper](https://arxiv.org/html/2606.05415)
- [Executable Schema Contracts: PDF](https://arxiv.org/pdf/2606.05415)