Nutrition Data Infrastructure for the AI Era: Operationalizing FAIR for Agent-Mediated Research

Summary

The paper presents Nutrition Data Service (NDS), source-preserving infrastructure that operationalizes FAIR (findable, accessible, interoperable, reusable) for automated, agent-mediated nutrition research. The central argument is that agent-mediated nutrition research requires a new data infrastructure for data identity, search, and crosswalk.

Problem and Motivation

AI agents can retrieve literature, operate research tools, write and execute analysis code, and combine evidence across sources, but their analyses inherit the identity, semantic, and release ambiguities of the underlying data. In nutrition research, a single study may span food-composition tables, dietary surveys, branded-product catalogs, prices, biomarkers, and health outcomes. An agent must not only retrieve a plausible nutrient value; it must select the intended food record, preserve its source and release, interpret analytical bases, and make cross-source joins that remain auditable.

The present food-data landscape makes failures likely: An assessment of 101 food-composition databases covering 110 countries found that only 32% offered an API and only 17 satisfied all 13 evaluated criteria of the FAIR principles. The review also reports that nutrient-intake estimates for an identical diet can vary by 20–45% with the database selected. Food sources rarely share stable identifiers, while names vary with geography, species, preparation, edible portion, processing, brand, and database purpose. A string join can silently conflate nutritionally distinct foods; flattening sources into one table can erase the very release history and analytical semantics needed to detect the mistake.

System Architecture

NDS separates authoritative source storage, a rebuildable retrieval index, and access for AI agents and applications. Authoritative values remain in the source store; the retrieval index is a rebuildable aid to discovery, not scientific evidence. NDS imports heterogeneous sources into DynamoDB while preserving records that appear to describe the same food. Each food receives the deterministic key food uid = uuid5(source id, source record id), so re-importing a release reproduces the same identity. The record also carries its source system, dataset, and release as explicit fields.

Food matching uses PostgreSQL with pgvector as a separate retrieval index. The English-language path has three stages: parsing, high-recall retrieval, and precision-oriented reranking. Query parsing uses an LLM constrained by a structured schema to decompose descriptions into normalized base food and closed-vocabulary facets. Hybrid retrieval uses two channels: a semantic channel (cosine similarity adjusted for facet agreement, weighted 0.70 · cosine + 0.30 · facet) and a lexical channel (full-text search of normalized names), combined via reciprocal rank fusion. An LLM-based listwise reranker compares candidates and returns a verdict, score, and ordering. If no candidate is defensible, NDS returns an explicit unsupported result rather than silently substituting the nearest food.

NDS provides a REST API for applications, named MCP (Model Context Protocol) operations for agents, and Parquet exports. MCP returns structured records rather than asking agents to extract values from prose.

Crosswalk

A crosswalk is "a versioned set of directed edges between records that lack shared identifiers. It relates records without merging them or asserting universal equivalence: both endpoints remain addressable, and each edge states the relation supported for a declared use and pair of endpoint releases. Typed relations are directed from source to target: exact (same food), broad (target more general), narrow (target more specific), close (related but not subsuming), and no-match. A mapping release is immutable and records its source and target releases, construction policy, and content identity. A workflow pins a watermark, an opaque committed snapshot, and resolution returns the concrete mapping release and policy selected at that snapshot. Later releases cannot change a pinned analysis."

Evaluation Results

Record-level resolution: On 1,000 held-out meal descriptions generated from NHANES recalls (3,597 reference foods), NDS achieves strict identity F1 of 0.875, with recall@5 of 0.942. An equivalence-aware scoring that credits siblings with matching energy and macronutrients raises F1 to 0.914. The remaining error is therefore driven more by ranking among near-duplicate records than by failure to retrieve a plausible record.

End-to-end estimation on NutriBench: On all 11,857 NutriBench v1 queries, NDS answers 96.4% of queries; among those answers, 84.6% are within 7.5 g and MAE is 4.3 g, compared with the best published GPT-4o result of 66.8% and 8.6 g. The unsupported 3.6%—mostly foreign meals—exposes a remaining source-coverage limitation.

External crosswalk benchmark (NHANES-to-DFG2): On the 1,304 labeled foods from Lemay et al., NDS achieves overall accuracy of 0.688 versus the published system's 0.654. NDS gains 15.8 accuracy points on the 611 no-match foods (0.624 vs 0.466) and loses 7.7 on the 693 matchable foods (0.745 vs 0.822). Of its 177 errors on matchable foods, 139 are abstentions; emitted targets agree with the benchmark 93.1% of the time. Most lost match accuracy therefore comes from refusing a link, not selecting the wrong target.

FNDDS-to-GI crosswalk audit: A blinded audit of 500 served mappings (125 per relation) using an independent LLM judge found 96.2% defensible and 77.0% receiving the asserted relation. Most disagreement concerns typing rather than whether a mapping exists: 37 of 125 narrow edges are judged close. Among 125 actual abstentions, 60 (48%) have no defensible target, 55 (44%) have only a dominant-component proxy that the contract intentionally rejects, and 10 (8%) have a missed close or narrow target; none has a missed exact target.

Reproducible agent-mediated research: In a person-level glycemic load analysis of 50 adults from NHANES 2017–2020 (830 records, 422 distinct foods), the NDS MCP arm produced per-person GL coefficient of variation of 0.000 across 12 runs (4 models × 3 repetitions), with 100% of people stable within ±10% and 100% in the same GL tertile in every run. The DIY web arm produced CV of 0.293 (worst person 0.816), with only 2.0% of people stable within ±10% and 32.0% in the same tertile. All 12 NDS runs return the same 207 food-to-GI assignments across models and repetitions, so every person's GL is invariant. NDS also had fewer false abstentions (4% versus 14%) and higher mean per-person carbohydrate coverage (83.1% versus 74.0%).

Infrastructure validation: Source reconciliation of 14,590 foods across four USDA FoodData Central releases (Foundation, FNDDS 2021–2023, SR Legacy, and a 1,000-record Branded sample) found zero failures across 3,501,054 source checks for nutrient amounts, units, bases, portions, and absent-versus-zero values.

Limitations

The paper acknowledges: geographic and source coverage outside the United States remains incomplete; NDS does not yet mediate licensed, subscription, or protected clinical data; evaluation covers selected tasks and some comparisons rely on published aggregate results or LLM-adjudicated labels rather than expert adjudication; and NDS makes sources, versions, and mapping decisions explicit, but it does not guarantee that the underlying evidence is clinically valid or appropriate for every analysis.

Conclusion

"FAIR principles provide the foundation, but agent use raises the operational standard. Nutrition infrastructure must do more than return plausible values: it must make identity, search scope, mapping semantics, uncertainty, and unsupported operations machine-actionable. This does not replace scientific judgment. It gives agents a reliable evidence layer so that researchers can inspect and reuse the same data choices instead of reconstructing them for every analysis."

Improvements to AI Systems:

1. Source-Aware Retrieval with Explicit Identity Preservation

• Implement deterministic record identity (e.g., UUID5 from source + record ID) so AI agents can cite exact data versions and re-imports without duplication.

• Force every retrieved value to carry source system, dataset, and release metadata as first-class fields, preventing silent conflation of nutritionally distinct foods.

2. Structured Abstention Instead of Silent Substitution

• When no defensible match exists, return an explicit “unsupported” result with reason codes (e.g., no-match, dominant-component-only, coverage gap) rather than nearest-neighbor fallback.

• This reduces hallucinated answers and enables agents to trigger human review or alternative data sources.

3. Hybrid Retrieval with Facet-Aware Semantic Matching

• Combine semantic embeddings (cosine similarity) with lexical full-text search and closed-vocabulary facet agreement (e.g., preparation, edible portion, species) via weighted fusion and reciprocal rank fusion.

• Improves recall on ambiguous food descriptions while maintaining precision on near-duplicate records.

4. LLM-Based Listwise Reranking with Verdict Output

• Use an LLM to compare candidate records and return a verdict (exact, broad, narrow, close, no-match) plus a confidence score, not just a ranked list.

• Enables downstream agents to reason about mapping semantics and auditability.

5. Versioned, Immutable Crosswalks with Watermark Pinning

• Represent cross-source mappings as directed, typed edges between specific release versions, never merging records or asserting universal equivalence.

• Allow agents to pin a watermark snapshot so later data releases cannot alter a completed analysis, ensuring reproducibility.

6. Agent-Oriented API with Structured Records (MCP)

• Expose data via Model Context Protocol operations that return structured JSON records (with identity, facets, and mappings) instead of prose.

• Eliminates extraction errors and lets agents directly consume machine-actionable evidence.

7. Equivalence-Aware Evaluation and Error Attribution

• Score retrieval not only on strict identity but also on nutritional equivalence (e.g., matching energy and macronutrients) to distinguish ranking errors from retrieval failures.

• Use abstention analysis to separate “wrong target” errors from “refused link” errors, guiding system tuning.

8. Coverage-Aware Abstention for Unsupported Queries

• Detect when a query falls outside source coverage (e.g., foreign meals) and return an explicit unsupported flag with a reason, rather than forcing a low-confidence answer.

• This improves trust and enables agents to route to supplementary datasets.

9. Reproducible Agent Workflows via Deterministic Mappings

• Ensure that repeated runs across different LLM models produce identical food-to-nutrient assignments by using deterministic keys and pinned crosswalks.

• Achieves zero variance in downstream calculations (e.g., glycemic load) across model changes.

10. Source Reconciliation with Automated Validation

• Implement continuous checks across all nutrient fields (amounts, units, bases, portions, absent-vs-zero) to guarantee zero silent data corruption during ingestion.

• This gives agents a verified evidence layer that does not require per-analysis revalidation.

---

What the Improved AI System Can Do:

• Answer nutrition queries with auditable, version-stamped evidence, citing exact food records and release dates.

• Refuse to answer when no defensible match exists, explaining why (e.g., coverage gap, dominant-component proxy).

• Join heterogeneous food datasets across surveys, composition tables, and branded products without losing source semantics.

• Reproduce identical analytical results across different LLM backends and repeated runs, enabling scientific replication.

• Detect and report mapping uncertainty (e.g., “close” vs “exact”) so researchers can adjust confidence thresholds.

• Operate autonomously on multi-source nutrition pipelines while maintaining full traceability for human review.