RAG Institutional Assistant
The problem
The school support team was overwhelmed answering the same questions about procedures and regulations to thousands of students daily.
The solution
24/7 virtual assistant using RAG (Retrieval-Augmented Generation) architecture connected to institutional knowledge bases and regulations.
Architecture
Ingest of regulations and procedures into a vector store. LangChain runs retrieve-then-generate: it retrieves passages, conditions the model, and cites the institutional source. It does not invent policy. Chat hits that index, not the open web.
How it was measured
Figure reported by the sole assistant developer at UVP, Feb 2024–Jun 2025. The 65% is the share of L1 queries (hours, procedures, repeated policy) the bot closed without escalating to a human, on traffic observed in that period. No published sample size; no university audit on this page.
Who built UVP's AI assistant and what does it resolve?
Franco Sanchez, as AI developer and tech mentor at Universidad del Valle de Puebla, Feb 2024–Jun 2025. A RAG over institutional policy answers L1 procedure questions 24/7 and leaves exceptions to the human team.
The work
Universidad del Valle de Puebla had a support team answering the same procedure questions every day. Hours, documents, policy that was already written. That is L1 traffic: not academic advice, lookup. From February 2024 to June 2025 I was the AI developer of the assistant and, in parallel, a technical mentor to students. Sixteen months, not a hackathon.
The architecture is RAG, not a model asked to “be the university.” Ingest of knowledge bases and regulations into vectors. LangChain retrieves passages, conditions generation, and keeps the answer tied to those passages. If the index does not have the procedure, the bot must not invent a graduation date. The channel is 24/7 because students do not only ask at the window. What it does not resolve as L1 escalates: the bot is not the coordinator.
The 65% is the share of L1 the operator (me) saw close without escalation. I did not publish a ticket count or a UVP report. An evaluator should treat that figure as self-reported. What is on the employment record is the role, the dates, and that the assistant existed to cut repeated load, not to replace a coordinator. If UVP ever signs an attestation, it will go in sameAs. It is not invented in the JSON-LD today.
I also mentored. That is not the product, but it is why the system is written so a student can keep the index: institutional documents change. A RAG that is not re-ingested is a rumor a month later. Mentoring and the bot borrowed the same rule: do not leave dead text in production.
This page is not a “what is RAG” tutorial. It is one university's system, with a name, a date range, and a percentage that states its method and its limits. Anyone looking for a LangChain course should leave this URL.
Screenshots
