Small AI You Can Run at Home Turns Messy Reports Into Clean Data
Companies can now organize their messy documents without sending secrets to Big Tech.
Most companies are sitting on piles of written reports that nobody can search properly. Think of maintenance logs, safety incident write-ups, customer complaints — thousands of documents full of useful facts locked inside paragraphs. This team, publishing at a knowledge-graph conference in Bangkok, built a pipeline that uses AI to read those documents and automatically convert them into a "knowledge graph." That's just a fancy term for a smart map of facts: who did what, where, when, and how it connects to everything else. Instead of a pile of papers, you get a database you can actually ask questions of.
Their test case was unusual and refreshingly practical: 80 handwritten-style incident reports from a French power-grid company. From raw paragraphs, the AI extracted the key people, equipment, and events, then mapped how they relate to each other. The researchers also had the AI build its own rulebook — what they call an ontology — for how those facts should be categorized. That rulebook is the important part. Once the AI knows the rules, it stops guessing wildly and starts filing facts consistently, the way a good librarian uses the same shelf labels every time.
The clever bit is what they ran it on. Instead of renting time on giant cloud servers, they used small open-source AI models — the kind you can download for free and run on hardware a mid-size company might already own. They tested sizes from 7 billion to 32 billion parameters (a rough measure of how much an AI 'knows'), including versions squeezed down to run faster and cheaper. The squeezed versions were slower to think but far cheaper to run, which the researchers call a reasonable trade-off.
Why does this matter outside the lab? Two reasons. First, privacy: power grids, hospitals, banks, and law firms can't casually upload internal documents to a public AI service. Running the AI in-house solves that. Second, cost: you don't pay per document forever. The remaining catch is that the AI still needed human-annotated examples to check its work, and the setup was tuned for one specific domain. Swap in a totally different industry and the rulebook likely needs rebuilding.
- The AI turns ordinary written reports into a searchable fact database — like giving your company a librarian who never sleeps
- It ran on free, open-source AI models small enough for a normal company server, not a giant cloud account
- Tested on 80 real French power-grid incident reports, with the smaller, cheaper AI versions performing acceptably
Why It Matters
Your employer could soon search decades of internal reports in seconds — privately, without leaking documents to outside AI companies.