SemanticMesh — Marc'Antonio Lopez

Context

Metadata management rarely fails because a database is missing. It fails because the same entity is called three different things in a glossary, a technical document and a DDL file. My M.Sc. thesis, developed as a graduate intern at Data Reply, explored whether production-oriented Generative AI could build and query that alignment automatically.

Problem

An LLM asked to link a business term to a schema column can produce a plausible link that does not exist. The thesis treats that as the central problem: the system has to retrieve evidence, validate the alignment and grade its own answer before returning it. The target was a working pipeline over real relational metadata, not a benchmark-only prototype.

Contribution

I authored the thesis and built the system: the builder graph that extracts triplets, resolves entities, validates with an actor–critic loop and heals failing Cypher; the query graph that combines dense search, BM25 and graph traversal, reranks with a cross-encoder and grades the answer for grounding; and the evaluation protocol across seven datasets.

Approach

Two LangGraph pipelines share one Neo4j ontology. The first writes and repairs the alignment; the second reads it and decides whether an answer is supported.

System architecture Politecnico di Torino · 2026

Builder graph

documents + DDL

triplet extraction

entity resolution

actor–critic validation

Cypher healing

Neo4j KG ontology

Query graph

question

dense + BM25 + graph

cross-encoder rerank

hallucination grading

grounded answer

Fig. 1 The builder writes the schema–document alignment into Neo4j; the query graph retrieves over it, reranks and grades every answer.

Builder graph

Triplet extraction. Documents and DDL are turned into candidate entities and relations.
Entity resolution. Duplicate terms and columns are merged before they reach the graph.
Actor–critic validation. A second agent checks proposed links and rejects the ones it cannot support.
Cypher healing. Failed writes are diagnosed and retried rather than dropped.

Query graph

Hybrid recall. Dense retrieval, BM25 and graph traversal each return candidates over the aligned ontology.
Cross-encoder reranking. Candidates are reordered by a cross-encoder before generation.
Hallucination grading. The final answer is checked against the retrieved evidence and graded.

Evaluation & results

The system was evaluated over real relational metadata: seven datasets, 111 tables and 210 questions, with ablations for the retrieval and validation stages.

7: Datasets
111: Tables
210: Questions
210/210: Grounded answers

0 hallucinations · 4.31/5 mean score · 21 ablation studies

Every answer in the evaluation was graded as grounded, with no hallucinations. The mean score was 4.31 out of 5, and 21 ablation studies tested the contribution of individual stages.

Figures are as reported in the thesis evaluation; no additional runs were performed for this page. Limitations were not documented in a form I can publish, so they are not listed here.

Code and thesis material on GitHub ↗