Observed facts: The Real AI SA author says the analyzer's knowledge base contains about 5,000 concepts and that development has cost 8 million rubles. In a reported pilot extracting assignments from administrative orders, document handling reportedly fell from 7–10 minutes to around 15 seconds. These project figures and the pilot account come from the author.
Mechanism: The system identifies concepts and relations in text and represents them as a graph in JSON, organized around actions, properties and objects. Rather than relying on a large language model to infer outputs, it uses a knowledge base and structured analysis; the author says it can run offline and needs modest hardware. That architecture can make outputs more repeatable, but shifts effort to vocabulary coverage and handling variation in actual documents.
Why now: The update gives a concrete, if limited, example of a non-LLM text-analysis approach alongside current LLM-centered tooling, and reports both operating economics and a real workflow trial.
Implications: For a constrained, clean-text workflow, a deterministic parser may be worth testing where privacy, cost or repeatability rules out an LLM. The pilot's reported time reduction is not evidence that the approach generalizes across organizations or document types.
Evidence boundary: The numbers, capability descriptions and pilot outcome are claims in the project's author-written Habr article; the source provides no independent accuracy evaluation or corroborating customer evidence.
Watch: Independent measurements of extraction accuracy on varied real-world documents would show whether the reported speed gain survives beyond the single described pilot.