AI Assistance (Preview)

BimlFlex offers AI assistance across the app: suggestions for names, descriptions and column mappings, drafts of integration keys, extension point fixes and rule sets, and plain explanations of settings, lineage and build runs. Every AI surface is a Preview capability. The assistance only proposes. You review each suggestion or draft and accept or reject it, and nothing the AI proposes reaches your metadata until you accept it. Where an accepted suggestion changes metadata, it lands as an unsaved change that you save the way you save any other edit.

Enabling AI Suggestions

Before any AI request, while the AI Enable Suggestions setting is off, the Enable AI Suggestions? dialog opens. It states that enabling the feature sends your metadata to a third-party LLM provider for processing, and that by enabling AI suggestions you consent to this data transmission under the End User License Agreement (EULA). It also tells you that you can revoke the permission at any time by turning off the AI Enable Suggestions setting in the Core category of Settings. The setting's key is AiEnableSuggestions.

Cancel closes the dialog and stops the request with nothing changed. Enable AI Suggestions turns the setting on, creating it when your workspace lacks it, and the request goes ahead. Every AI surface on this page passes through this same check.

The standard installer ships with the AI features on, and the separate no-AI installer for customers who need the legacy connectors ships with them off and shows none of the AI surfaces.

When a press cannot be answered, for example because your account has no AI access or the AI service cannot be reached, the surface shows one plain sentence beside the control you pressed. The control stays visible and enabled so you can try once more, and your next press clears the sentence. A refusal from the AI service shows the service's own sentence. Any other failure, such as a server fault or a lost connection, shows the fixed sentence The AI service is not available right now.

In a hosted deployment, an AI request to a paused workspace brings up the pause overlay, a modal dialog in front of the app, and the request ends with the plain failure sentence instead of being sent a second time.

Ask Flexi

Ask Flexi opens the AI assistant panel in the right drawer. The top bar carries an Ask Flexi button after the notifications button. The toolbars of the connection, column and object editors, in their form and grid views, and the Model Lineage page carry one too. Every opener runs the consent check first.

The panel, titled Ask Flexi, lists the actions the current page offers, each with a short description. On a connection, for example, it offers Suggest Description, and on the connections grid it offers the same suggestion for the selected rows, or for every row when none is selected. On the Build page, it offers Diagnose this run, which opens the Diagnostics tab and explains the selected run.

Below the actions, the panel shows guidance that needs no AI request. This item lists the failing validation findings of the item you are editing, each with its remediation text. Workspace lists the most severe validation findings across the workspace, each with an Open link to its item, such as Open Connection. When there is nothing to show, the panel reads Nothing needs attention in this workspace.

Naming and Description Suggestions

BimlFlex suggests values for two fields, Description and Business Name. Ask for a suggestion from the Ask Flexi panel. The Suggest Description dialog, or Suggest Descriptions for several items, takes optional Additional Info and a Thinking Mode choice, and notes that using the feature sends data to a third-party large language model. Suggest closes the dialog, and the suggestions arrive in the background while you keep working. Cancel closes it with nothing sent.

A field that holds a waiting suggestion shows the suggested text with two buttons, Accept and Reject. A form fills only an empty field with the suggested text. Accept writes the value into the field as an unsaved change of the form, which you save with the form. Reject puts the field back to its value and records the rejection right away, with no form save needed.

In a grid, a cell whose suggestion differs from its value is tinted, and a filled cell keeps its own value as its display. When you start editing an empty cell that has a suggestion, the cell editor opens with the suggested value. The column, project, macro and object grids also carry suggestion columns, hidden by default, whose cells show the suggestion with Accept and Reject on hover.

On the Import Metadata page, a checkbox beside the connection and project selectors, off by default, queues naming and description suggestions for the objects you import and their columns. The suggestions land on the imported items for you to accept or reject, and no field value is written until you accept.

Triage on the same page classifies the candidate tables of the loaded list by their names alone. The Triage candidate tables dialog lists each table with its classification, a Keep or Skip recommendation, a confidence and a rationale. Apply to selection updates only the checkboxes of the table list: keep becomes checked, skip becomes unchecked, and unanswered tables stay as they are. Nothing is imported, so review the selection before you import.

Mapping Suggestions

Suggest Column Mappings on the Columns grid toolbar of a saved object opens the Suggest Column Mappings dialog with that object as the target. The object menu of the Data Lineage designer opens the same dialog with the object you clicked as the source.

In the dialog, choose the Source Objects and the Target Object, read the count of source and target columns, and add Additional Instructions if you want to guide the AI. Thinking Mode asks for deeper reasoning. Suggest sends the request, and Cancel closes the dialog with nothing sent.

The result dialog lists each suggested mapping from a source column to a target column with its confidence and a Rationale you can expand. A mapping that needs a conversion you author carries a transform badge, and a column the AI advises against mapping reads Do not map. Filter the list with Min Confidence, use Select All or Deselect All or pick rows one by one, and choose Apply. Applying writes each selected mapping as an unsaved change of its source column and skips a column that already has a mapping. Do not map rows and unanswered rows never write. Apply Results counts the applied and skipped mappings.

Integration Key Drafts

With the AI features on, the Generate Integration Key dialog carries Draft key with AI under its column list. When it finds a usable key, it proposes the key columns of the table and checks them in the proposed order. You still confirm with the dialog's own action, and cancelling the dialog applies nothing.

Settings Help

In the dialog that Explain this setting opens, Explain with AI narrates the explanation of the setting. The narration shows above the details with an AI chip and the caption Narrated from the details below. Those details stay the record. For a sensitive setting, its values are not sent.

A toolbar action of a settings group form opens the Review values for a target platform dialog. Pick the Target platform and choose Review values. Only settings that apply to that platform and differ from their shipped default are reviewed, and sensitive settings are never sent. Each row shows a proposed value with Apply, or reads Reviewed, no change or Unreviewed. Applied values become unsaved changes on the settings form, which you save with its Save button.

A second toolbar action opens the Find a setting dialog. Describe what you want under What do you want to change? and choose Search. Each result names a setting, with a short reason where the AI ranked it, and selecting a result opens that setting in the editor.

Lineage

On the Model Lineage page, the AI narrates what the lineage computes, and the counts, the trace and the diagram stay the record.

In the impact panel, Write impact brief narrates the grouped counts the panel shows. The brief carries an AI chip and the caption Narrated from the grouped counts. The numbers of record are in the table above.

In the trace view, Orient me narrates the statistics of the trace you are looking at. When the lineage loads a snapshot that differs from the one you viewed in the same session, a Snapshot changes panel shows the counts of added and removed items, and Narrate changes narrates them.

The Ask about this lineage panel, marked Preview, answers questions such as where the values of a column come from. Type the question and choose Ask, or Cancel while it runs. The AI answers from the lineage already loaded in the view: it can look up objects and columns by name, trace a column and list what a column reaches. Names in the answer are links that select the table or column in the view, and a name that is not part of the loaded lineage reads (unverified). The panel notes that the answer is narrated from the loaded lineage, that the trace and the diagram are the record, and that nothing is saved.

Review heuristic mappings on the lineage graph toolbar opens the Review Heuristic Mappings dialog over the lineage edges that rest on a column name match alone, with no authored mapping behind them. The AI confirms or challenges each one. Accept on a confirmed row stages the mapping as an unsaved change on the source column, which you save as usual. A challenged row shows what the AI proposes under Proposed instead:, and challenged and unanswered rows are advisory only.

Build Diagnostics

On the Diagnostics tab of the Build page, Explain this run explains a finished run that has diagnostics. The explanation is advisory and changes nothing. An entity it names becomes a link only when the name resolves to an entity of your metadata. To start the same explanation from the Ask Flexi panel, use Diagnose this run.

Recommendations and Auto-fix with Flexi

A dashboard recommendation about a missing description, a missing business name or another single text field carries Auto-fix with Flexi. Its tooltip reads Flexi suggests a value for this field. Nothing changes until you accept it. The press asks for one suggestion, shows Flexi is writing a suggestion. and then the suggested value with Accept and Dismiss suggestion. Accept lands the value as an unsaved change of the entity's field, and the recommendation leaves the list. An unsaved changes line with Save and Discard sits on the dashboard. The detail panel of the recommendations workbench offers the same button.

On the recommendations workbench, a selection of rows or a rule group with a fixable rule offers Auto-fix with Flexi over every open row at once. A confirm bar names the rows and requests of the run before it starts, and Stop sits beside the progress of the run. Each suggestion shows beside its row for you to accept or dismiss one at a time, or all together with Accept all and Dismiss all. Undo reverts Accept all, and Retry asks for the rows that got no suggestion. When a fixable rule has many open rows, the dashboard points you to the workbench.

Drafts

After you validate an extension point, each finding carries Draft a fix with a Preview chip. The Diagnostics tab of the Build page shows the same button beside Open extension point. The draft shows in the Drafted fix panel, marked AI and Preview, as an explanation and a difference of the current text against the drafted text. Apply draft puts the drafted text into the editor as an unsaved edit, which you save and validate as usual. Dismiss drops the draft. When the AI finds nothing to change, the panel says so and the text stays as it is. A draft does not apply when the extension point text changed after the draft started.

Proposing and Critiquing a Rule Set

A rule set models a table's business rules: its inputs, joins, filters and grouping, and one rule for each output column, from which BimlFlex generates the view SQL. Two verbs work on rule sets, each with a Preview chip. Propose a rule set proposes a rule set from imported metadata, and Critique this rule set critiques an existing one. Both answer with reviewable proposals that you accept one rule at a time. Both verbs sit in the Draft with AI group on the Rules tab toolbar of the modeled table's editor and in the Draft with AI menu of the Dimensional Accelerator page, and Propose a rule set also appears on the import result screen.

The proposal shows under Proposed rule set as one table shape row with the column rules grouped under it, each with Accept and Skip, and Accept all reviewed. A column rule can be accepted once its table row is accepted, and a skipped rule stays out until you add it by hand. Accepted rows become unsaved changes. When you save, BimlFlex generates the view SQL from the accepted rules, and a licensing refusal shows as a warning with nothing lost.

The Rules Editor guide describes the Rules tab of a modeled view, where you edit the rules one at a time. The Dimensional Accelerator guide describes the accelerator's own drafts, Suggest roles and Critique this star, beside the two rule set verbs.

Where Suggestions Are Kept

BimlFlex keeps each suggestion together with its outcome, accepted, rejected or applied, so a resolved suggestion never shows as a waiting one.

In a database workspace, the suggestions are kept with each entity in the metadata database.

In a git repository workspace, the suggestion text the AI writes is kept in one central file, ai-suggestions.json, at the root of the workspace's store in the repository. Saving a suggestion or its outcome changes that one file in your source control and never an entity file. A value you accept into a field saves with the editor, as any edit does. AI requests in a repository workspace read the same metadata every user of the branch sees.

The suggestions save carries the version of the store you last read. When the suggestions file changed outside the app since then, the save is refused and the Save Conflict prompt opens. Keep mine saves your version over the disk version. Keep the disk version reloads the store and ends the save without applying it, with no error alert.

What AI Assistance Does Not Do

AI assistance never applies a suggestion or a draft that you have not accepted. Explanations and narrations change nothing, and the details, counts, traces and diagnostics they narrate stay the record.

AI assistance takes no part in generating your warehouse. Generation stays deterministic: the same metadata produces the same output every build, with no language model in the generation path. The SQL of a modeled view comes from the BimlFlex generator, which writes it from the rule model.