← Back to the journal

DataFocus · Journal

Build your first OT Flash ontology: three tables, two documents, one answer you can audit

From first login to an agent answer you can audit: bring in three tables, create an ontology, bind tables and definition documents, then check View processing details against Data Explore before inviting colleagues.

Getting started with DataFocus OT Flash means building one small ontology and proving it answers correctly before anyone else depends on it. The r57 user manual's path has five steps: sign in and change the default password; get two or three tables into Flash from a file or a read-only database connection; create an ontology, bind those tables and upload the one or two documents that define the business terms; ask the ontology agent a question and open View processing details to see which tables and tools it used; then walk a short checklist with the business owner before inviting colleagues. The fourth step is where first ontologies fail quietly: an answer that looks right but was built on the wrong table, or without the definition document, is not one you can hand to a colleague.

Every screen, field and label below comes from the OT Flash r57 user manual. This is the documented procedure, not a report from a running instance.

What you are building, and what must exist first

OT Flash is a data analysis and knowledge platform for business users: type keywords and the system draws a chart, ask FocusGPT in natural language, or talk to an ontology agent that combines business data with documents. In Flash an ontology is a knowledge space organized around one business topic: related tables, business documents, indicators and pinboards, plus an agent that answers from them and can call executions to run an analysis or write results back. Each ontology has 11 modules (Agent, Ontology Graph, Document Center, Data Tables, Data Explore, Answers, Pinboards, Data Apps, Indicators, Execution and Control Panel); a first ontology uses five.

Have these in hand first:

You needWhy it mattersWho provides it
The Flash address, for example https://<server address>:31600Flash runs in a browser; the manual recommends the latest ChromeAdministrator
An account whose roles include Tables and OntologyMenus depend on your roles and permissionsAdministrator
A configured large language model serviceOntology agents work only after one is configuredAdministrator
One business question with a named owner and written acceptance criteriaKeeps the ontology small enough to check by handYou and the business owner
Sample or synthetic rows for that question's tablesThe getting-started guide says to start with synthetic dataYou

If there is no instance yet, OT Flash is distributed free; installers and release notes, with system requirements and supported platforms, are on the download page and in the GitHub repository.

The fourth row decides how the rest goes. DataFocus's getting-started guide asks you to define the task, its business owner, its acceptance criteria and the permitted scope before building anything. The product brief narrows the choice: pick a question that already has data, policy documents and a clear owner, such as order delivery or purchase approvals; build one ontology for it; get business users working on it; only then authorize an AI agent or move on. A practical test: if you cannot name the document that defines the question's key term, you have not picked the question yet.

Step 1: Sign in and read the menu

  1. Open the address in Chrome and sign in with the account you were given. The first administrator login forces a change of the default password; do not reuse that default anywhere.
  2. Ask now whether a mail server has been configured. Self-service password recovery works only when one has.
  3. Read the left menu, which runs from Home (the Ontology Workbench) through Tables, Ontology and Logs to Admin and Recycle. If Tables or Ontology is missing, your role does not include it. Ask for the role rather than borrowing a login; Step 5 explains why.

Step 2: Get the tables into Flash

An ontology binds tables that already exist in the system, so the data comes first. The r57 manual lists five routes in: a local CSV, Excel or JSON file, an external data source, an API, or an HTTP source. The decision rule for a first ontology:

SituationRouteReason
Two or three tables, synthetic or exported rows, no refresh neededFile import under TablesNothing to configure or secure, easy to delete
The same tables, but they must stay currentExternal data source with a scheduled importThe import wizard can set a schedule
Twenty tables "to be safe"Neither, yetYou cannot verify twenty tables by hand

File import is the default for a first attempt: export the tables your question needs, say orders, shipments and customers, and import them under Tables.

Connecting a database means filling in Source Name, Server, Port, Database, an optional Schema and JDBC Param, a username and password, then pressing Connect to test. The manual recommends a dedicated read-only database account for OT Flash, limited to the databases and tables you plan to analyze; ask the DBA for one rather than using your own credentials. Import then runs as a three-step wizard: select the data source, select the tables, configure how to load them, with the import type and an optional schedule.

Either way, open each imported table and check the column names and types; the agent can cite only what is there.

Step 3: Create the ontology and bind tables and documents

Create it

Go to Ontology. The page shows how many ontologies exist and how many more you can create; the manual's example reads "2 items / 98 more can be created". Creating one takes a Name and an icon. Name it after the question, "Order delivery", not after a department or a source system; "Finance tables" tells a colleague nothing about what it answers.

Bind tables and define the joins

Open the ontology and go to Data Tables. Import data table adds tables from the system to this ontology; the Relationships tab sets up joins between them, as in the Tables chapter. If orders and shipments share an order number, define that join here. Without it the agent and Data Explore see two unrelated tables.

Then use Data Explore, which offers the same search features as Data Search but only over this ontology's tables. Run the two or three filters your acceptance criteria depend on, such as orders shipped after their promised date last month, and write down the counts. Those are the reference numbers for Step 4.

Upload the documents that define the terms

Go to Document Center and upload the documents that explain the data: metric definitions, business rules, operating procedures. Each document shows File Name, Type, Size, Related Tables, Updated time and Status. The agent can search and cite a document only once its Status reads Completed. Ask before that and you get an answer built on the tables alone, easy to mistake for an answer built on the rule.

The documents earn their place. In a 2023 benchmark by data.world researchers (Sequeda, Allemang and Jacob, arXiv), GPT-4 answered 16% of enterprise questions correctly when prompted directly against an insurance SQL schema, and 54% when the same questions were posed over a knowledge graph built from an ontology and mappings. One schema and one model generation, so treat the figures as a direction, not a law. The direction is clear: answer quality is bounded by the context you bind, and a one-page definition is the cheapest context you will ever add.

A worked example

Consider a hypothetical "Order delivery" ontology, not taken from the manual:

Bound objectContentWhat it gives the agent
orders tableOrder number, customer ID, promised delivery dateThe population and the promise
shipments tableOrder number, ship date, carrier, delivered dateWhat happened
customers tableCustomer ID, name, regionThe grouping colleagues ask for
Relationshipsorders to shipments on order number; orders to customers on customer IDThe ability to combine them
Document 1One page defining "on-time delivery"The rule that decides which orders count as late
Document 2The escalation procedure for late ordersThe basis for recommended actions

Read the graph before you ask anything

Ontology Graph draws the knowledge nodes (tables, documents, fields) and their relations as a network graph. Its Graph quality panel lists nodes, relations, isolated nodes, connected components, unmapped mentions and failed sources. After adding data or documents, use Rebuild or Refresh, then ask whether the structure matches the table above. An isolated table node means a join is missing; a failed source means a document did not come through. Fix both before the agent has a chance to route around them. In the English UI the graph toolbar labels are in Chinese, so keep the manual's screenshots open.

Step 4: Ask the agent, then audit its work

You can ask in two places. On Home, the Ontology Workbench, Talking With Ontology sends a question to the selected ontology's agent; the "@ ontology name" tag shows which. Inside the ontology, the Agent module shows the agent reading the tables and documents and working through the question step by step; Stop task interrupts it. For test questions use the Agent module and press Clear Context before each one, so the agent starts a new topic and the earlier conversation does not shape the answer you are about to check.

Expect to wait. The manual says an answer usually takes tens of seconds, longer for complex questions. It returns a conclusion, detailed records, risk notes and recommended actions, states which tables the data came from, and Run Summary gives the number of steps.

Now audit it. Cited data comes from the tables and documents bound to the ontology, and View processing details shows which data the agent queried and which tools it called. For the first few questions, read it every time against this table.

What processing details showLikely causeWhere to fix it
A table you did not expect was queriedOver-binding, or a join that leads somewhere surprisingData Tables: remove the table or correct the relationship
The definition document was not cited on a question that depends on itDocument not yet Completed, or the question did not use the document's termsDocument Center; then ask once more using the defined term
Detailed records differ from the Data Explore count for the same filterMissing relationship, or a filter interpreted differentlyRelationships tab; compare the filter the agent applied with the one you ran

Rewording the question appears once in that table, and only after the document is Completed. When the agent uses the wrong table or skips the document, the cause is nearly always in Step 3; prompting around a structural gap tells you nothing about whether the next colleague's question will fail the same way.

Step 5: What to confirm before inviting colleagues

Walk this list with the business owner you named at the start.

  • [ ] The ontology covers one business question, and its name says which.
  • [ ] Every bound table is needed. Tables bound "just in case" are gone.
  • [ ] Every document shows Completed in Document Center and is the current version.
  • [ ] Graph quality shows no failed sources, and you can explain any isolated node.
  • [ ] Three to five typical questions have been asked, and their processing details reviewed against Data Explore.
  • [ ] The answers meet the written acceptance criteria. The getting-started guide warns against treating production completion, dispatch and customer acceptance as the same event without confirming them; in a delivery question, that is the distinction your definition document should settle.
  • [ ] Database connections use a read-only account.
  • [ ] Access is granted through roles, not shared logins.

The last item is why Step 1 said to ask for a role rather than borrow a login. r57 ships preset system roles that cannot be modified, among them SYSTEM_MANAGER, SOURCE_MANAGER, DEPARTMENT_MANAGER, DEVELOPER and LOG_MANAGER, and administrators can add custom roles, granting each permissions per resource under All Resources, Ontology and Data Entry. Operate Logs records operations such as logins and table imports with the operating user on each entry; viewing it needs LOG_MANAGER or SYSTEM_MANAGER. When a colleague reports that yesterday's answer has changed, that log tells you whether a table was re-imported, and by whom. A shared login turns it into noise.

Limits to know before you start

  • The agent depends on the model your administrator configured. The manual names no vendor or model, and this post claims nothing about answer quality or speed. Flash is self-hosted; if the configured model is external, confirm what is sent to it under your data policy.
  • The ontology is a workspace, not a formal modelling language. In r57 you build it from tables, relationships, documents, indicators and the graph. If you need explicit object types and rule definitions, the Flash manuals do not describe that.
  • Small is the point, not a compromise. Three tables and two documents can be verified by one person. Bind everything the department owns and nobody can check the answers, a different problem from the answers being wrong, and harder to notice.

Next step

Keep the r57 user manual open at Chapter 10 (Ontology) and Chapter 6 (data sources) while you work. Write down the task, owner and acceptance criteria first, as the getting-started guide asks; the criteria are what tell you whether the ontology worked.