Every company has its own words: acronyms on a purchase order screen, a short form a planner types every day, a table name that means one thing in purchasing and another in logistics.
Stages one, two and four belong to your team; stage three belongs to Juno and runs in the background.
Stage one sets the scope. Your data lead picks the registered tables that people ask about and marks the columns in daily use. Juno reads the description of each table and profiles the values in the marked columns, so it learns which column holds supplier numbers and which holds plant codes.
Stage two is the words themselves. Four kinds of input go in: a vocabulary file of company terms, a list of acronyms with their expansions, a list of abbreviations, and custom aliases that give a table or a column the business name your people use. Each entry keeps the name of the person who added it.
Stage three runs the training task. Juno rebuilds the knowledge index: every table and column, the values in each column, and the names in business language with your synonyms attached. People carry on asking questions while the task runs, and the new version takes over when it finishes.
Stage four closes the loop. The result downloads as a working file, your team corrects what reads oddly, and the corrected file goes back in. Each published version is dated, carries an author and can be restored, so a vocabulary change is as reviewable as a rule change.
The four inputs arrive as spreadsheets or through the vocabulary screen; both routes end in the same list.
Company terms with a short meaning for each: “push-out”, “allocation-bound”, “firm zone”.
The short forms your people type, each with the words it stands for: NCNR, OTIF, DOS, ASN.
Shortened words that appear in field names and in typed questions, with their expansions.
A business name for a table or a column, so a question can use the word your team uses.
| Term | What it means in this company | Kind | Added by | When |
|---|---|---|---|---|
| NCNR | Non-cancellable, non-returnable: a line the supplier holds to the agreed date | Acronym | D. Okafor | Thu 09:00 |
| LT fence | Lead-time fence: the date inside which the plan stays firm | Abbreviation | L. Tan | Thu 09:20 |
| “strat supplier” | The strategic class on the supplier record | Alias | P. Sharma | Thu 09:40 |
| DOS | Days of supply: stock divided by average daily demand | Acronym | L. Tan | Thu 09:48 |
| “cxl window” | The days in which a purchase order line may still be cancelled | Alias | D. Okafor | Thu 10:04 |
| “push-out” | A customer moving an agreed delivery date later | Vocabulary | J. Ridley | Fri 08:30 |
Six of the 96 entries in the showcase estate’s first vocabulary file, which P. Sharma owns and reviews. Showcase estate.
The task runs in the background, so people keep asking questions while a new version is prepared.
| Stage | Step | What happens | Who |
|---|---|---|---|
| Choose the scope | 01 Pick the tables | The registered tables people ask about join the training set | Data lead |
| 02 Mark the columns | Columns in daily use are marked; the rest stay in the catalogue | Data lead | |
| 03 Profile the values | Juno reads a sample and records the pattern of each column | Juno | |
| Give the words | 04 Load the vocabulary | Company terms with their meanings enter the list | Domain owner |
| 05 Load the acronyms | Each short form is stored with the words it stands for | Domain owner | |
| 06 Load the abbreviations | Shortened words used in fields and questions are expanded | Domain owner | |
| 07 Add aliases | Tables and columns receive the business names your team uses | Domain owner | |
| Run the task | 08 Train | The knowledge index is rebuilt and matched to the ontology | Juno |
| Correct and repeat | 09 Download the result | The proposed meanings arrive as one working file | Data lead |
| 10 Correct and upload | The corrected file goes back in and training runs again | Data lead |
Steps three and eight belong to Juno. The other eight are your team’s, and they are the reason the result sounds like your company. A first pass over a purchasing estate takes an afternoon; later passes take minutes, because only the changed tables need a fresh run.
A term taught once is live for every person and every agent on the next question.
| Supplier | Lines | Value |
|---|---|---|
| Delta Fasteners Ltd | 6 | $54,200 |
| Corvid Plastics | 3 | $28,900 |
| Two others | 2 | $13,300 |
Showcase estate.
On Wednesday L. Tan asked which open purchase order lines were still inside their cancellation window that week: 31 lines worth $348,000. Two follow-up answers came back wider than expected, because the questions used words Juno had yet to learn: the NCNR flag D. Okafor writes on a supplier note, and the “LT fence” L. Tan works to every week.
On Thursday D. Okafor added NCNR to the acronym list at 09:00, L. Tan added the lead-time fence at 09:20, and P. Sharma added the alias for the strategic supplier class at 09:40. P. Sharma owns the vocabulary file and reviewed all three entries. The training task ran in the background at 10:30, over four purchasing tables, in six minutes. By 10:40 the same question returned the 31 lines with the 11 NCNR lines marked, and the cancellation policy written on Friday read the supplier class through the same vocabulary.
What a thumbs-up records, what a correction changes, and how a verified pair becomes the answer everyone sees.
The work moves from writing a model to teaching the words your company already uses.
Most tools that answer questions over data ask for a semantic model and a business glossary written by hand before the first question. Practitioners describe the preparation for a large estate in quarters, and every new source starts the exercise again. The glossary then drifts, because the people who own the words work far from the people who maintain the model.
Juno starts with the supply chain ontology already in place, so training adds your words to a map that exists. The people who own a term add it themselves, in a file or on a screen, and see the effect on the next question.
The people who use the words: a planner, a buyer, a master-data steward. Your data lead collects the files, and each entry keeps the name of the person who added it.
Minutes for a few dozen tables. In the showcase estate 23 tables took 18 minutes as a background task, and people carried on asking questions throughout.
You add the entry and run training again for the tables it touches. The word is live from the next question, and the earlier version stays available for restore.
Table names, column names and your vocabulary may reach the external model at design time. Row data stays inside your estate. The trust page states the boundary rule in full.
Yes. A term can carry a scope, so purchasing and logistics each keep their own word for the same ontology attribute, and both questions are answered correctly.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.