In preparation

Reliability by engineering: a RAG over Costa Rica's official gazette

Working title · Project: Gaceta AI

Hypothesis

In a RAG over an official gazette, most of the reliability comes from verifiable engineering decisions rather than more sophisticated retrieval: chunking by notice, checking citations in code and filtering personal data before indexing. It is measured with an ablation over a new benchmark, and the results may refute it.

Contributions

• An open benchmark of questions on Costa Rican regulation, each with an answer and a verifiable citation (gazette, date and page).
• A retrieval ablation over a Spanish legal corpus: vectors, hybrid search, BM25, a reranker and a graph of regulations.
• Citation checking in code: how many citations the generator invents and what it costs to remove them.
• Privacy before indexing in an official gazette that publishes personal data.
• Prompt injection from the corpus: paid notices as an attack vector.
• A comparison with general-purpose assistants with web search, on the same task and metrics.

Where it would be published

1. arXiv preprint (cs.CL and cs.IR)
2. NLLP 2027 workshop, co-located with EMNLP
3. ACL Rolling Review as a resource paper, if the results hold up

Written in English once the results of phases 1 and 2 are measured.