Deutsch

Lab

Every experiment starts with a question from day-to-day operations. We work it through with example data, show the models involved and write down where the limits are. Experiments that fail stay up too.

Experiment 1: Sorting enquiries

Started running

Question
Can the inbox be sorted so that a broken heating system does not wait behind catalogue requests?
Data
Four made-up enquiries from building services: outage, noise, quote, question.
Models
A language model reads the enquiry and fills in three fields. A rule tree with three questions decides. Which language model will be noted here after the first run.
The model reads, your rules decideExample with made-up data, field values written by hand

The heating in the new building has been down since this morning1, there is no hot water2. Please send someone as soon as possible3.

Our heat pump has been making loud noises since yesterday1 but still heats2. An appointment this week3 would be good.

Please send us a quote for 12 type 22 radiators1, delivery in week 463.

Do you have the installation guide for ET-47111? No rush3.

The language model reads
1Request
FaultFaultQuoteQuestion
2System
downrunningnot mentionednot mentioned
3Deadline
nowthis weekweek 46none
Your rules
Fault?
yes
System down?
yes
AEmergency service, call back within one hour
no
BService team, appointment next working day
no
Quote or order?
yes
BSales office, quote the same day
no
CAssistant answers from your documents
The model does not assign a priority. If you change a rule, it applies from the next enquiry on, without retraining.
Result
Open. The field values above are written by hand; the run with a language model is still to come.
Limits
The tree only knows what is in the fields. If an enquiry does not mention the system, the field stays open. Several requests in one email and attachments are not covered.
Transfer
You set the questions in the tree, for example by maintenance contract, system type or region. The language model stays the same.

Lab notebook

Dated entries on running, completed and discarded experiments.

  1. Experiment 1 started running

    Rule tree with three questions and four made-up enquiries from building services are in place. The field values are still written by hand; the run with a language model follows.