DataFocus · Journal
Teaching FocusGPT Your Business Terms with My Knowledge Base
How to add company terms and abbreviations to FocusGPT in DataFocus OT Flash through My Knowledge Base, share them system-wide through Admin > Knowledge, and check how a question was parsed.
You teach FocusGPT your company's terms by adding entries to My Knowledge Base, the window at the bottom left of the FocusGPT page in DataFocus OT Flash. Each entry holds the definition (content), the keywords that should trigger it, a type, a category and an enabled switch. Add entries one at a time or in a batch, and ask an administrator to publish the shared ones under Admin > Knowledge so every colleague gets the same reading. After that, you check the line FocusGPT prints about how it parsed your question and correct it if it is still wrong.
This post walks through that workflow for an analyst who keeps getting answers that are close to right but use the wrong meaning of a word.
Why does Smart Q&A misread company jargon?
FocusGPT, the Smart Q&A feature of OT Flash, takes a natural-language question, chooses a data table, parses the question into a search and returns a result. The manual notes that when it cannot tell which table to use it asks you to pick one, and that its reply includes a line describing how it parsed your input so you can judge whether it understood.
Company vocabulary is where parsing has the least to work with. Consider a hypothetical mid-size distributor (illustrative, not a real customer). Its sales team says "GM" for gross margin, finance says "net sales" for revenue after returns, and operations says "active customer" for a customer with an order in the last 90 days. A table might contain columns named gm_pct, rev_net and last_order_date. Nothing in the column names says that "active" means 90 days, or that "GM" is not a job title. A model reading the question has to guess, and two analysts asking the same question on different days can end up with different guesses.
The manual's own tip points the same way: include specific measures, dimensions and time ranges in the question, and answers are more accurate. A knowledge entry moves that precision out of each individual question and into a place FocusGPT consults every time.
What is in a knowledge entry?
In the manual's description of My Knowledge Base, each item has four pieces of content plus a switch:
| Field | What you put in it | Example (hypothetical) |
|---|---|---|
| Content | The definition, in plain words | "Active customer: a customer with at least one order in the last 90 days." |
| Keywords | The words or abbreviations that should bring this entry up | "active customer", "active accounts" |
| Type | The kind of knowledge the entry is | Choose from the types offered in the window |
| Category (shown as Stickers) | A label for grouping and finding entries | "Sales definitions" |
| Enabled | Whether FocusGPT uses the entry | On |
Three habits make entries work better.
- One term, one entry. Put "GM" and its definition in one entry rather than a paragraph covering ten abbreviations. Short entries are easier to find, switch off and correct.
- Write the definition the way you would explain it to a new colleague. State the measure, the table or column family it relates to, and any threshold, such as the 90 days above.
- List the spellings people actually use. If sales writes "GM" and finance writes "gross margin", put both in the keywords. The same idea applies to synonyms in general: the keywords field is where you record the different names for one concept.
The Enabled switch matters more than it looks. When a definition is under review, or a quarter-end rule no longer applies, you can turn the entry off without losing the text.
To find entries later, use the keyword box or the knowledge category box at the top right of the window. A consistent category scheme (by department, or by subject area such as sales, finance, inventory) is what makes that search useful once you have a few dozen entries.
Should I add entries one at a time or in a batch?
The window offers both. New knowledge adds a single item and is the right choice while you are learning how the fields behave: write one entry, ask a question that should use it, and see what changes.
Batch Import is for the moment you already have a glossary somewhere, such as a spreadsheet maintained by finance or a data dictionary exported from another tool. Batch Import brings in many terms at once; the manual does not describe the import file layout, so check the import window before you prepare a spreadsheet. A workable rule of thumb:
- Fewer than a handful of terms, or terms you are still debating: add them singly.
- An existing, agreed glossary: import it in a batch, then review the result and switch off anything you do not want FocusGPT to use yet.
Administrators have a bulk option too. Under Admin > Knowledge, the Bulk Knowledge Import tab imports knowledge in bulk and All bulk imports lets them review the import tasks.
How do administrators share knowledge with everyone?
My Knowledge Base is personal. A definition you add there is your own, which is useful for experiments and for terms only your team uses. Terms that the whole company must read the same way belong in system-wide knowledge.
The manual says system administrators manage knowledge shared across the whole system under Admin > Knowledge. That page holds several tabs:
- Personal Knowledge and System Knowledge Base, for managing personal knowledge and the knowledge shared across the system.
- FocusGPT's markings, which lists the mappings between natural-language questions and keyword statements that users have corrected or confirmed while searching. Administrators can view and maintain them by knowledge category. In the English interface this tab label leaves out the assistant name.
- Bulk Knowledge Import and All bulk imports, described above.
- Knowledge categories, where the category list is maintained.
The markings tab is worth a regular look. Every time someone corrects a parse and confirms it, that question-to-keyword mapping is recorded. An administrator who reviews them by category can see which phrases the organisation keeps having to correct, and turn the recurring ones into proper entries.
A division of labour that tends to work: analysts draft entries in their own knowledge base, and one named person (the data owner for that subject area) promotes the agreed ones to the system level and keeps the category list tidy. The manual describes the places and the roles; who approves what is a decision for your team.
How do I check what FocusGPT understood?
After you ask a question, read the parse line before you read the chart. The manual describes it as a line in the reply telling you how FocusGPT parsed your input. If the term you taught shows up in the interpretation, the entry was found. If FocusGPT still read the word the old way, work through this checklist:
- Is the entry enabled?
- Does the keywords field contain the exact wording you typed?
- Is the entry in your personal knowledge when you expected it to be system-wide, or the reverse?
- Does the content say what you mean in a sentence that can stand alone?
- Did you ask about the right table? FocusGPT chooses a table, and asks when unsure; a perfect definition cannot help if the question lands on a table without the relevant columns.
Responses usually take a minute or two, and you can ask follow-up questions in the same chat to narrow things down. The thumbs-up and thumbs-down icons under an answer tell FocusGPT whether it helped.
The manual also carries a caution that applies to every AI answer in the product: the page says content is AI-generated and should be checked carefully, and it recommends checking important conclusions in Data Search. Data Search shows the SQL and an analysis summary on the right, and when you type a full sentence there instead of keywords it asks FocusGPT to parse it, shows the parsed statement and offers a Revert option. For a figure that will go into a board deck, run it through Data Search too.
Keyword-style search has its own tuning, separate from the knowledge base. The Search tab under Admin lists custom locations, custom dates such as a fiscal year, synonyms and custom keywords. If your problem is "FY" or a region nickname inside a keyword search rather than a full question to FocusGPT, look there first.
Knowledge base or ontology documents?
OT Flash has two places where business text lives, and they do different jobs.
- My Knowledge Base is where you tell FocusGPT about definitions and terminology. It is fast to edit, structured by keyword, type and category, and can be shared system-wide by administrators.
- Document Center in an ontology manages business documents bound to that ontology, for example metric definitions, business rules and operating procedures. Each document has a file name, type, size, related tables, updated time and status. Once the status is Completed, the ontology agent can search and cite its content.
A practical way to choose: if you want a short rule that should shape how questions are interpreted in FocusGPT, write a knowledge entry. If you have a whole procedure, policy or metric handbook that people and the ontology agent should be able to quote from, upload it to the Document Center of the relevant ontology and link it to the tables it governs. They are not exclusive. A metric can have a one-line knowledge entry for FocusGPT and a full definition document in the ontology.
When this is not enough
A knowledge entry guides interpretation; it does not change the data. If "active customer" needs a calculation that does not exist in any table or formula, the entry can describe it but cannot produce it. In that case, create the column or formula first, then write the entry that points to it.
Entries also need an owner. A glossary nobody maintains drifts from what the business means, and FocusGPT will faithfully repeat the old definition. Put the review date in your team calendar, not just in the entry.
Next steps
OT Flash is free to download from the editions page, and the installers and release notes are on GitHub. The r57 user manual covers FocusGPT in Chapter 3 and the Admin > Knowledge tabs in Chapter 13. Start with ten terms, check each one against the parse line, and expand from there.