Release date: September 2026. This note also covers 9.02 (July 2026) and 9.03 (August 2026); the version in which an item first shipped is given in brackets.

Datafor 9.02–9.04 add two new visualizations (Decomposition Tree and Histogram), the Metrics Library with governed metric bindings, a rebuilt Data Security editor, a major Model Designer update, and a new generation of the AI Agent query engine with stage-level LLM configuration, usage statistics and quotas.

Highlights

1. Decomposition Tree (9.02)

A new chart for driver analysis. The root node shows the total of one measure; every level splits the selected node by a dimension. Users choose the next dimension manually, or let Smart split pick the highest or lowest contributor (absolute, or relative to the dimension average).

• Up to 50 Explain by dimensions, each queried independently instead of one large cross join; up to 10 members per level with data bars.
• The selected path is published to linked components and to drill-through and jump actions, so other charts follow the analysis.
• The path saved with the report is the default state; Drill Reset restores it.

Decomposition Tree

2. Histogram (9.03)

A new chart that shows the distribution of a numeric field. Binning is computed in the database: Auto bins, a fixed number of bins (2–100) or a fixed bin width; the Y axis shows Count or Percentage; empty bins are kept; the tooltip always shows both count and share.

Clicking a bin filters linked components by its numeric range (several bins from the same histogram are combined with OR). Report filters, component filters and data permissions are applied. 9.04 fixes linked-filter write-back, axis labels, re-rendering after style changes and validation.

Histogram

3. Metrics Library and metric bindings (9.02 → 9.04)

Data → Metrics Library stores the governed definition of each enterprise metric: a stable Metric ID, business name, synonyms, definition, calculation relationship (Base metric, Ratio, Difference, Attainment rate or Custom formula), unit, direction, owner, and a Draft / Certified status. Libraries can be imported and exported as JSON or CSV.

In the Model Designer, a measure or calculated measure is bound to an enterprise metric under Business semantics → Metric governance, optionally with an effective grain. The Metric bindings panel lists every binding with its status. Compare definition (or Compare all) asks the AI to compare the model implementation with the business definition and records the verdict: Matches definition, Possible drift, Needs comparison or Not enough evidence. Duplicate, missing and stale bindings are reported by Model diagnostics.

The AI Assistant resolves a metric name or synonym in a question to the certified definition and warns when a definition is draft, unbound, ambiguous or drifting.

Metrics Library

4. Data Security rebuilt (9.02)

The Data security editor of a database datasource was rebuilt.

• Row access policies use a condition builder: field, operator and value per condition, AND/OR groups, an Effective condition preview and Validate expression; Return all rows creates an explicit exception.
• Table & column access policies hide selected tables or columns from selected subjects, or make them visible only to selected subjects.
• Policies target users, roles or user types through one mixed search box. Save draft keeps a policy inactive, Save & enable activates it; the policy list has an enable switch, search and filters.
• Test access simulates a user, role or user type and shows the resolved roles, the effective row condition, field visibility and the policies in effect, with a data preview of up to 50 rows.
• Uploading a model can overwrite its existing permissions; row permissions of system roles are applied correctly.

Data Security

5. Model Designer update (9.03 → 9.04)

• A unified property panel with the same four sections for every object; format and unit are combo boxes that list the values already used in the model.
• A new Insights panel at the bottom of the designer with Model diagnostics (errors, warnings and hints for structure, references, semantics and metric governance; errors are reviewed when saving) and Metric bindings.
• Tidy model into a star schema proposes removing dimension or measure-group roles that conflict with the inferred star-schema role, with a preview before anything is changed.
• The model-level Default time dimension and the measure-level Default time field are separate settings; time semantic roles are assigned per attribute.
• Business semantics for AI: captions, descriptions, aliases, sample values, semantic roles, unit, direction and recommended dimensions on dimensions and measures.
• Four additional statistical aggregators for measures; distinct counts on levels.
• Datasource and schema are selected together; tables land where they are dropped; dimensions and measure groups can be deleted and re-added from the canvas table menu; context menus fixed; connector routing on the canvas redone (9.04).
• SQL Views with check-and-preview; the SQL of a modeling query can be displayed.
• File datasets: table structure and field type maintenance; failed imports no longer stay "in progress"; wide-table imports on PostgreSQL fixed; decimal precision follows the database dialect.
• Units and display scale (9.04): measure units are stored as language-neutral codes (ISO 4217 currencies, CLDR units, %, counts) and shown in each user's language; a new Display scale setting (K / M / B, or 万 / 亿 for Chinese readers) becomes the measure's default number format in reports. Metrics Library uses the same unit picker and exports both the code and the display label. Units and Display Scale

Analysis Model Overview · Model Diagnostics · Tidy Model into a Star Schema · Business Semantics for AI · Time Semantics and Default Time Settings · Creating SQL Views

6. AI Agent: new query engine (9.02 → 9.04)

The Agent's query pipeline was rebuilt (9.03) around a verified semantic draft: the LLM proposes a structured query intent, Datafor proves it against the model metadata and executes it under the user's permissions. The AI never writes SQL or MDX.

What users notice:

• Faster, more direct answers and fewer unnecessary clarifications. The date basis is assumed and disclosed instead of asked; when a clarification is needed, the options can be answered in your own words.
• More calculation types: month-by-month year-over-year, period comparisons labelled by the actual windows, distinct counts ("how many customers"), ratios and averages, rankings along a time axis, filters on calculated measures, and unions of conditions.
• Evidence and boundaries with every answer: the time window used, data coverage, adopted member mappings and any limitation; citations link each conclusion to the exact cells.
• Bounded multi-step analysis for questions such as "where did the decline come from": a plan with a few steps, each with its own result; an incomplete answer is marked partial instead of failing.
• Follow-ups reuse the previous result (the table you were shown is the referent); the chart type can be switched without a new query; follow-up chips state the question they will ask.
• Dashboard insights stream as they are generated, explain only the data on the page and say when the page context is truncated; Generate business brief produces a management summary; the three insight entry points share one implementation (9.04).
• Add to page inserts the exact query the answer ran, including multi-step answers (9.04); charts and tables can be exported as image or Excel.
• History restores the model that was used; cancelled questions are labelled; failure explanations and insight text follow the language of the question (9.04).
• Unsupported operations are refused with an explanation rather than a silent failure; relationship questions (scatter and correlation) are supported.
• MCP entry point: ask Datafor from Claude Desktop, Claude Code and other MCP clients under the same permissions and rules.

AI Assistant · AI Agent Overview and Roadmap

7. AI Agent administration (9.02 → 9.04)

• LLM: model profiles with provider presets (OpenAI, Qwen, DeepSeek, Gemini, Anthropic, custom), Verify and enable (a real structured-output verification before a model can be assigned), a model assignment for each of the 14 agent stages plus the embedding role, and schemes that can be saved, exported, imported and switched, with built-in templates for global and China-region providers. API keys are never exported. (Stage-level assignment 9.02; verification and schemes 9.03; panel clean-up 9.04.)

• Ops → Usage shows questions, success rate, response time and p95, and tokens per period. Ops → Quota sets a daily question quota per user with overrides by role and user type and an administrator exemption; submissions beyond the quota are refused. (9.03)

• Vector indexes show the embedding model used, status and progress; the chat shows a notice when the selected model has no index (9.04). Changing the embedding model requires a rebuild.
• Operations: /ai/health endpoint, background retention clean-up, multi-worker deployment on Linux, and a packaged runbook.

LLM Configuration · AI Operations and Quotas · Preparing Data for AI

Query engine

• Filters on calculated measures are planned generically and pushed down safely, including calculated measures across several fact tables, partitioned filters and large sets (9.02–9.03).
• Query-level SQL levels and calculated levels are pushed down natively (this powers histogram binning and range filters); two-segment level queries; distinct-count auto measures on levels; native null filters on dimension SQL (9.03).
• Dimension and level annotations are loaded and passed through metadata, including shared dimensions (9.02).
• Multi-level member queries run in parallel and return normalized errors; empty-result pagination counts and <= boundary comparisons fixed (9.03).
• PostgreSQL regex literal escaping fixed; SQL member queries and MDX member enumeration no longer leak statements or connections (9.03).

Console and platform

• Connection management follows the user's permissions (9.03); Portal login credentials are no longer persisted in plain text (9.03).
• Uploaded datasets: field management and append with field mapping (9.02).
• The vector index list labels the model as Embedding model (9.04); Metrics Library CSV import accepts headers in any supported interface language (9.04); export scope is the selected rows or the whole library (9.04).
• Report SDK resources carry a cache key so browsers load the new bundle after an upgrade (9.04).
• Front-end distribution: dependency clean-up, npm as the only package manager, third-party license notices included in the package (9.03).
• Toolbar size inputs are limited to 1–20000 and toolbar buttons have accessible two-state styling; invalid canvas sizes are rejected (9.04).

Important bug fixes

• Metadata without annotations no longer renders a blank table; missing level information is handled (9.02).
• Collapsing the property panel no longer closes the AI Agent panel (9.02).
• Histogram: a linked filter replaces the previous range instead of accumulating; axis labels; re-render after panel changes; height of copied components; advanced-filter validation (9.04).
• Inserting an AI result into the canvas keeps the hierarchy level; panel configuration write-back crash and dialog style scoping fixed (9.04).
• Percentage functions with hidden named sets, legacy model metadata null pointer, model metadata resource inheritance and refresh API error messages (9.03).
• Simplified query API: silent aggregation downgrade, duplicate aggregation suffixes and parameter validation (9.03).
• Modeler: property panel horizontal scrollbar; renaming a measure migrates its metric binding; duplicate bindings reported as diagnostics; no preview for incomplete file imports; datasource page blank screen (9.04).