The day I started a RAG system for PEMEX

On June 25, 2026 I made the first commit to RAG_pemex, now called Xerebro. I’d been at PEMEX for a month and the problem was obvious: the refineries produce operational reports every single day (throughput, equipment failures, shutdowns, start-ups), and finding anything meant opening PDFs one by one.

The idea: you ask “what happened at the Tula plant yesterday?” and it answers citing the report, the refinery, the date and the shift. If it doesn’t cite the source, it’s useless.

Week one: learning the hard way

The early commits are gold. A few from June 25–30:

• “Fix timezone mismatch on cloud servers”. The server ran on UTC, so asking about “today” late at night gave you tomorrow. I had to pin it to Mexico time.
• “Implement refinery-balanced round-robin selection”. Ask about several refineries and the search returned chunks from whichever one had the most text, while the rest never showed up. Now results are spread across refineries.
• “Balanced round-robin date-based selection”. Same thing with dates: one wordy day would “eat” all the others.

None of this is in RAG tutorials. They teach you to drop documents into a vector database and call it a day; the hard part starts when the data is real and people ask questions the way they talk.

How it grew

At first I also had a separate version (Rag_pemex_local) on Hugging Face Spaces for testing. Over time everything merged into one project:

• Intent routing: “what is hydrodesulfurization?” doesn’t need the vector database, it comes from a technical knowledge base.
• An LLM cascade (Gemini, Groq, OpenRouter). If one runs out of quota, the next one takes over and you never notice.
• Natural-language dates (“yesterday”, “on Monday”, “August 15”) parsed without asking the model.

Today the repo has over 320 commits and it’s still in production, indexing a new report every day. And yes, I even made a loader where a drop falls and the PEMEX eagle appears. Priorities.