Redesigning a RAG and sleeping 10% fewer hallucinations
March 11, 2026 · RAG · IAMEX · LLM · Mexico
I worked on the conversational assistant that Mexico now knows as I’AM Chat, launched in November 2025. My role: Senior AI Engineer (RAG) at IAMEX, Mexico’s sovereign AI platform.
The goal was tough: a government assistant that must not invent, because people depend on its answers. And my specific task was to redesign the retrieval-augmented generation system so it hallucinates less.
What was breaking
Classic RAG retrieves documents, stuffs them into the prompt and prays. The typical problems:
- Poorly prioritized context: the model read 20 chunks when only 3 were relevant.
- Chunks cut mid-thought: retrieving half a paragraph is almost worse than retrieving nothing.
- No fidelity control: the model could “fill in” from its own priors instead of anchoring to the document.
The redesign
I redesigned the entire architecture:
- Semantic chunking instead of fixed-size slices.
- Aggressive re-ranking so only what survives the relevance filter enters the prompt.
- Citation anchoring: the answer must map to retrieved fragments, or the model says so explicitly.
- Confidence thresholds for saying “I don’t know” instead of inventing.
The result was a 10% reduction in the hallucination rate. In a sovereign assistant, that 10% isn’t a metric: it’s people getting correct information instead of made-up facts.
The takeaway
A good RAG isn’t “vector store + a nice prompt.” It’s systems engineering: where you cut the text, how you rank the evidence, and when the model should shut up. That, plus the fact that latency rules: elegance is nice, but the user doesn’t want an essay — they want an answer.