Local models for service note analysis

Three documents, two visits, one leak

A service manager looks for machines where a leak was described at more than one visit. In the note from the first visit, the technician wrote „leak found”. After the second: „checked for leaks, no leak found”. The third document is a summary of the first visit and repeats its observation. All three contain the word „leak”, and they describe two visits, of which only one confirms a leak.

The assistant gathers this history into two entries. For the first visit it shows the observation and links to both documents that describe it. For the second it keeps the information that no leak was found. In this example there is no basis for saying the leak was seen again. A repeated description of a symptom does not establish the cause of a failure either.

What a local model for service notes costs and when it pays back

The count covers the time spent searching, reading and compiling the material. What stays is the time to check the output against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled. Anthropic, in Contextual Retrieval, reports that hybrid search with reranking cut failed retrievals by 67%. In notes written in varied language, searching for the word „leak” finds the documents, and their meaning still has to be read.

What runs locally and what needs agreeing

Local processing means the model runs on a named computer or server in the company, which suits notes about machines and customers' sites. The Ollama documentation distinguishes local and cloud models and describes switching off cloud features, so the name of a tool alone does not settle where processing happens. The app around the model needs its own agreement: copies of documents in backups, the manager's questions in the history and the connection to the service system.

When your current report is enough

If the reports have consistent observation fields and visit identifiers, an existing report or filter is enough. A model is worth testing when the information you need sits in free text and the manager reconstructs the history from scattered descriptions. Two errors tell you the most: confusing an observation with its negation, and counting a summary as a new visit. Writing a current note is a different job, covered in the article on a small model on a warehouse terminal.

Where to start

A good start is a part of the history the service team knows and can judge. The manager compares the summary with the documents and checks whether they still have to read everything from the beginning. We roll out local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free process review, a description of one case and the places where the material sits is enough. Confidential files stay in the company, and you get a written plan within 2 working days.