Problem
Employees of a large corporation need an LLM assistant for documents, analytics, summaries and presentations. Security rules forbid sending documents to public chats such as ChatGPT. Nornickel built its own platform inside the perimeter. The pilot on stock Open WebUI handled chat and simple RAG but failed at data analysis: users wanted pivot tables and HTML infographics from Excel files with 100 sheets. Given a bare "here is a file, solve it" prompt, the model invented fields and looped on broken code.
Example from the article: a user uploads a complex multi-sheet Excel workbook and asks for a pivot table. The model invents column names, writes failing code and burns tokens on retries.
Solution in brief
Keep Open WebUI as the UI and put a LangGraph orchestrator behind it that calls specialised agents over MCP. The data analytics agent profiles the Excel file with a deterministic script first, uses schema-constrained generation for decisions, and runs code step by step in a Firecracker sandbox. LiteLLM routes all model calls.
Questions this case answers
- How to make an LLM agent reliably analyze large multi-sheet spreadsheets without inventing fields or looping on broken code
- How to let an LLM choose among options without rationalizing a choice it has already made, using schema-constrained generation and a sandbox
- How to build an in-perimeter corporate AI assistant when employees cannot send documents to public LLM chats
- How to make an LLM agent reliably analyze complex multi-sheet Excel files
Tools and methods
- Open WebUI — chat UI and simple RAG
- LangGraph orchestrator over MCP — ReAct agent that calls agent services
- Deterministic Excel profiling — a script gives the LLM a JSON description of the file
- Schema-constrained generation — Pydantic schema with ordered fields
- Firecracker sandbox — one microVM per chat session
- LiteLLM gateway — one model catalog with fallbacks and tiered routing
- Mem0 — chat and per-user memory
Google or a one-time email code. No password.
