Release 6.2
Tellius 6.2 is here! 🔥 Experience a true agentic platform that can build, remember, adapt, and accelerate work across the entire analytics journey.
Kaiya Apps turns completed analyses into interactive, role-specific applications your team can open without touching a query.
Data Architect lets your team go from raw tables to a production-ready Business View without leaving the conversation.
With persistent memory, every definition, correction, and preference Kaiya learns stays learned across every session, for every user who needs it.
One conversation can now be from different Business Views. Kaiya carries your filters, context, and history forward as topics shift.
And this is only the headline story. There’s a lot more in 6.2, so dig in!
🚀 New features
Kaiya Apps: Build data applications from a conversation
Kaiya Apps generates persona-scoped, interactive data applications directly within Tellius.
The distinction from Vizpad is architectural: Vizpad is a general-purpose analytics canvas where users configure charts and filters manually. Kaiya Apps generates a full application (component layout, data bindings, filter logic, navigation structure) through a single conversational exchange, tailored to a specific role and set of recurring questions.
Apps query the underlying Business View at creation time and store a data snapshot. This snapshot powers the app immediately. There is no waiting for a live query on every load. Data can be refreshed on demand by admins and app owners, and the app remains fully usable during refresh using the last-loaded state.
How to create a Kaiya App
There are two entry points:
From a Kaiya Deep Insights conversation: After Kaiya completes an agentic analysis, a “Create App” button appears inline. This converts the analysis into an interactive application. The app inherits the analytical approach that powered the Deep Insight (the analysis logic, metric relationships, and findings). The source conversation and the resulting app are bidirectionally linked via "Go to Source" and "Go to App" navigation.
From the Apps section directly: Navigate to the Apps page, select a Business View, and describe the application in natural language. Kaiya generates a plan (listing goals, key metrics, dimensions, analysis types, and layout) which the user can modify conversationally or approve to trigger generation.
Customizing a Kaiya App
What distinguishes Kaiya Apps from static dashboards is how you modify them. Once an app is generated, you change it the same way you created it: in natural language, directly within the Kaiya conversation.
Tell Kaiya to remove a KPI card, swap a chart type, add a dimension filter, or change the time range. Each instruction triggers recalculation against the Business View and updates the app in place. There is no drag-and-drop editor or configuration panel; you describe what you want, and the app adapts.
Every modification is saved as a version. Changes remain in draft until you explicitly publish, and version history is accessible from within the app so you can roll back if needed.
Data refresh: Admins and app owners can trigger manual refreshes. Refresh is async; the app remains viewable during refresh using the last-loaded data state.
For example, consider a pharma field performance app.
A commercial ops lead navigates to the Apps page, selects the territory performance BV, and types:
"Build an app for regional directors showing TRx and NBRx performance vs target, call frequency trends, and territory coverage gaps".
Kaiya proposes a 3-tab layout:
Summary (KPI cards + trend line)
Territory Comparison (ranked table)
Alerts (territories with coverage_status = Under-covered) with all key findings, insights, and summaries.
The lead approves. The app publishes and is shared with the six RDs. Each Monday, data refreshes automatically. RDs open it on their laptops without touching Kaiya or Vizpad directly.
Data Architect: From raw tables to analysis-ready BV in a single conversation
Setting up a Business View today is a multi-step, multi-module workflow, each step with its own interface. Data Architect collapses this entire workflow (Connect → Prepare → Business View) into a single Kaiya conversation.
Data Architect is a Kaiya agent that creates Business Views conversationally from raw data sources. The agent handles schema analysis, join inference, derived table generation (in the upcoming patch), calculated column creation, metadata enrichment, and BV validation without the user navigating between modules.
When the agent is uncertain (ambiguous join keys, multiple valid aggregation levels, unclear business intent), it asks clarification questions with pre-populated multiple choice options. This keeps the user in control without requiring them to write SQL or understand schema internals.
For ambiguous requests, the agent can generate multiple candidate plans (each representing a different join strategy or aggregation level) and present them side by side for the user to select. This is especially relevant when the same business question could be answered at different granularities (e.g., HCP-level vs territory-level).
The agent walks through five sequential phases: Intent → Tables → Joins → Columns → Validate. Each phase is a distinct checkpoint. You can approve before the agent proceeds to the next phase.
The agent's plan is represented in YAML and displayed in a split-pane editor alongside a visual diagram. The YAML plan is reviewable and editable in Draft state before anything runs. This is a key trust point for data teams handing BV creation to a non-technical user.
Snowflake and Redshift (in Live mode) are the only supported data sources for now. All other data sources require the dataset to be pre-loaded into Tellius before starting.
What can the data architect agent do?
Schema analysis: Detects column types, identifies primary/foreign keys, classifies snapshot vs event tables. No manual profiling.
Join inference: Proposes a join graph with plain-language rationale. M:N relationships are flagged with fix suggestions before anything is built.
Calculated columns: Translates your business questions into formulas, with row-level vs report-level handling. Refineable in natural language.
Metadata import: If you already have metadata documented (a data dictionary, column descriptions, business glossary, or any structured reference file) you can upload it and the agent will use it to configure the BV and populate metadata. Existing definitions are source-of-truth; the LLM fills only what's missing.
Validation: Row count integrity, join explosion detection, formula validation. Issues surface with plain-language explanations before the BV publishes.
Kaiya’s persistent memory: Your definitions, rules, and preferences
Kaiya now maintains a persistent memory layer that survives across conversations. No need to re-establish the same business context. Define it once and it applies automatically in every future session, for the scope you choose.
Memory entries are scoped to the user, group, or organization level, and can be tied globally across the instance or to a specific Business View. Memories created at organization scope apply to all users by default; user-scope memories override org defaults for that individual.
How to create a memory?
Manual entry: Upload a file (PDF, TXT, CSV with glossaries, business rules documents, data dictionaries) or type individual entries directly. Uploaded files are parsed and individual memories are extracted, previewed for review, and saved on confirmation.
Auto-suggested: During conversations, Kaiya identifies clarification moments where saving context would benefit future sessions — metric definition corrections, routing preference signals, entity disambiguation. Kaiya surfaces a prompt: "Should I remember these?" and dsiplays the list of clarifications done. You can choose the required clarifications to remember for the future or decline. No memory is written without explicit user confirmation.
Correction persistence: When you correct a definition mid-conversation or explicitly say "Memorize this:", Kaiya can persist that correction as a memory entry. Future conversations with the same pattern apply the correction automatically.
The scope of memory
Who it applies to: User, Group, or Org. Hard constraints at org scope override individual preferences when they conflict. Admins control critical business logic, users retain flexibility for personal preferences.
What it applies to: Global (all conversations) or scoped to a specific Business View. A definition scoped to the Commercial Performance BV doesn't bleed into unrelated analysis.
When a new memory entry is being saved, Tellius runs a semantic similarity check against existing entries. Entries with most similarity update the existing record rather than creating a duplicate. Entries that are somewhat similar are flagged for review. This prevents memory bloat from repeated corrections phrased slightly differently.
The Learnings feature is deprecated in 6.2. Memories is the replacement going forward, and all Learnings capabilities (saved definitions, correction preferences) are subsumed into the Memories system. Learnings remain accessible during the transition period but will be removed in the future.
Fully dynamic Auto BVs: Cross-topic analysis Without Switching BVs
In real analytical conversations, users naturally move across topics. Sales performance, then inventory, then promo effectiveness. Each may live in a different BV. 6.2 makes BV selection fully dynamic. Kaiya evaluates the optimal BV for each question independently, at query time. The prior conversation context (filters applied, metrics discussed, clarifications made, entity references) carries forward even when the underlying BV changes.
How Kaiya carries the context across different question
When Auto BV switches to a new Business View mid-conversation, it carries forward:
Active dimension filters (e.g., region = Southeast, product = Laundry)
Time period selections
Metric references established in prior turns
Entity clarifications (e.g., 'market' as defined in the conversation)
Kaiya maps these context elements to the schema of the new BV. Where a dimension or metric has a direct equivalent in the new BV, it is applied automatically. Where there is no equivalent, Kaiya surfaces the gap and asks for clarification.
Faster answers to the questions your team asks most
6.2 introduces a semantic caching layer. On the first response, the generated SQL is stored against its natural language query. On subsequent queries, Kaiya checks for semantic matches in the cache. Matching queries route through a faster, lighter model with the cached SQL as a reference, producing the final query more quickly.
How the cache works
When a question comes in, Kaiya checks whether a semantically similar question has been resolved before on that Business View. If the similarity crosses a threshold, the cached SQL becomes a reference point. The final SQL is generated through a faster model that starts from the cached version and adapts it to the current question: different date range, different dimension value, different filter combination. The structure is reused; the specifics are recalculated. This is why filter variations work correctly instead of returning stale results.
The cache is shared across all users on the same Business View.
The cache is per-BV, not per-user. When one user resolves a question, every other user on that BV benefits from it going forward. This is where the value compounds in practice; the questions that slow down individual users tend to be the same questions everyone on the team is asking: Monday morning performance checks, weekly pipeline pulls, recurring exception reports. One resolution, shared benefit.
The cache delivers the largest benefit on high-frequency questions where the intent is the same across sessions even if the filter values change slightly. Examples:
“What were last week's TRx by territory?” — asked every Monday by multiple regional directors
“Show top 10 products by revenue this quarter” — asked across sales and commercial ops
“What is call coverage vs target for this week?” — pulled by field force managers daily
Whenever the underlying Business View changes, the affected cache entries are automatically invalidated. The next question resolves fresh against the updated schema. You will never get a result built on SQL that references columns or relationships that no longer exist.
Storyline-driven PPT and PDF export
Kaiya can now export complex analyses as professionally formatted PowerPoint or PDF presentations, structured around a coherent narrative.
When you trigger an export from a Kaiya conversation, Kaiya generates a storyline that organizes the findings into a logical arc: context, key findings, supporting evidence, implications, and recommendations. The storyline is presented in an editable view where you can reorder sections, modify framing, or provide a textual instruction in natural language (for example, "Add a final slide for the executive summary"). Once the storyline is confirmed, Kaiya generates the final deck.
After a completed Deep Insight, export options appears inline, converting the analysis into a presentation that inherits the full analytical context. Kaiya generates the storyline and builds the deck within the conversation, rendered in a side-by-side Artifact Viewer.
Artifacts reflect your branding
Admins can upload a corporate template (PPTX or PDF) per environment. Kaiya extracts a comprehensive brand style specification from the template, covering color palette, typography, slide layouts, logo placement, chart styling, and visual elements. All subsequent exports are rendered on-brand, matching the uploaded template's visual language. When no template is uploaded, Kaiya default template is used automatically.
Charts from the Kaiya conversation are embedded as high-resolution images to guarantee visual fidelity. Tables are rendered with template-consistent formatting. All numeric values in the exported deck exactly match the source data.
Scheduled Agentic Workflows
Agentic workflows can now be scheduled to run on a recurring basis. You can set up a schedule directly from a completed Kaiya conversation, choosing frequency, timing, and notification preferences. Kaiya executes the workflow automatically on the defined cadence and delivers results to subscribers.
The following buttons will be displayed at the end of every Kaiya Deep Insight response:
Create App
Subscribe to existing schedule
Create new schedule for workflow
Creating a new schedule
The scheduling flow is a two-step wizard.
Step 1 (Basic Information): Select the workflow, give the schedule a name and optional description, and optionally provide any additional context the workflow needs to run.
Step 2 (Schedule and Delivery): Configure the start date, start time, and frequency. You then configure the email delivery: recipient addresses (with Cc and Bcc support), a subject line, and a custom email message body. An "Include report" toggle controls whether the full workflow output is attached to the email as a report.
Subscribing to an existing schedule.
Subscribe to get email alerts and notifications without creating duplicates. When a workflow already has one or more active schedules, clicking "Subscribe to existing schedule" displays the available schedules. Each card shows the schedule name, recurrence pattern (for example, "Runs daily at 7:50 PM GMT+5:30"), the next scheduled run date, the delivery method, and the author who created the schedule.
All scheduled workflows can be viewed and managed under Kaiya → AI Agents → Scheduled Workflows.
📈 Enhancements
Kaiya Embedding: Row-level policy support and BV-based context
Kaiya's embedded mode now supports row-level policy (RLP) enforcement. When Kaiya is embedded in a customer's application, the host application can pass RLP conditions that filter data at query time, ensuring users only see data they are authorized to access.
Additionally, embedded Kaiya now supports Business View pre-selection via parameters. When a customer embeds Kaiya for a specific use case, the embedding URL can enforce a particular Business View without exposing internal BV identifiers to end users. Users in embedded mode are also now prevented from navigating out of the Kaiya interface, keeping them within the intended analytical experience.
Granular Point-in-Time configuration at the column level
Point-in-Time analysis can now be configured at the individual column level rather than only at the Business View level. This gives data teams finer control over which metrics use end-of-period snapshot calculations versus sum or rollup aggregations. For example, headcount and balance-type metrics can be set to Point-in-Time independently from flow metrics like revenue or transactions within the same Business View.
Column-level Point-in-Time configuration is supported only on Snowflake and Redshift Business Views with deduplication enabled.
Admin control: Disable Business Views from Kaiya
Admins can now disable specific Business Views from being available in Kaiya. This provides governance over which data sources are exposed to the conversational AI interface, allowing organizations to restrict Kaiya access to vetted, production-ready Business Views while keeping development or sensitive BVs out of the AI-assisted query path.
TQL monitoring in Resource Tracker
A new TQL monitoring capability has been added to the Resource Tracker. It continuously monitors error rates and performance metrics across the Tellius Query Language layer (which is the backbone for Search, Vizpad, and Insights) and sends alerts when TQL request performance degrades. This gives admins proactive visibility into slowdowns before they cascade into user-facing issues, and is particularly useful for identifying expensive queries, monitoring concurrency, and diagnosing slow-performing objects in production environments.
Pinned Vizpads for user groups
Admins and Super Users can now pin Vizpads to user groups, making critical Vizpads prominently visible to all group members. Pinned Vizpads appear at the top of the group's Vizpad listing, ensuring that key operational dashboards, reports, and monitoring views are immediately accessible without requiring individual users to find and bookmark them.
French language support
Tellius now supports French as a platform language. Users can switch to French in their language preferences, and the interface, labels, navigation, and system messages render in French across all modules.
Native Vertex AI connectivity
Kaiya can now connect natively to Google Vertex AI, expanding the set of supported LLM backends. Organizations running on Google Cloud can route Kaiya's language model calls directly to Vertex AI without requiring an intermediary proxy or API adapter.
Advanced LLM model assignment configuration
Admins now have fine-grained control over which LLM models are assigned to different Kaiya functions. The configuration allows admins to map specific models to specific Kaiya capabilities, thereby assigning the best model for each purpose. For example, a faster model for simple queries and a more capable model for Deep Insights, optimizing for cost, latency, and quality based on the organization's priorities.
Selective chart export as separate CSVs
Users can now export individual charts from a Vizpad as separate CSV files, in addition to exporting the entire Vizpad's data. This is particularly useful when a Vizpad contains multiple charts from different Business Views and the user needs the underlying data for a specific visualization.
Performance Toggle: Disable Stats Calculation
A new toggle in Data → Advanced Configuration allows admins to disable automatic stats calculation during data loading. When enabled, this improves load performance for wide datasets by skipping column-level statistical analysis. Dimension and Measure classification falls back to data-type-based inference. Null percentage and type distribution statistics are not computed. This is recommended for datasets with hundreds of columns where the initial profiling pass is unnecessary.
🛠️ Minor Fixes
Kaiya
Improved memory management during extended Kaiya chat sessions, reducing resource consumption and improving stability over long conversations.
Improved Deep Insights summary generation to ensure all output uses user-facing business language consistently.
Improvements to Kaiya's unstructured data handling, including performance and accuracy refinements for document-based question answering.
Vizpad
Resolved numerous filter-related issues in the redesigned Vizpad, including: pinned filter sync, global filter passthrough to data view and pivot tables, filter value persistence when editing, filter operator switching, column-level filter application for measure columns, date filter editing, and "Select All" / "Deselect All" behavior for dimension column filters.
Fixed chart configuration and interaction issues including: chart type switching, color-by bucket responsiveness, aggregation label display, tooltip rendering, bulk filter application, conditional formatting for tables, and switch metric behavior in view mode.
Resolved export-related issues including: PPT export rendering for large charts, filter display in exported PPTs, section-level export for PPT/CSV/Excel, KPI and cohort chart export options, and location chart JPG download in embed mode.
Fixed pivot table issues including: data mismatch, pinned column behavior in view mode, dynamic match formula calculations for totals and subtotals, and extra spacing with single-measure configurations.
Resolved custom view issues including: chart title retention when switching views, filter and color formatting propagation across tabs, time resolution changes not reflecting in exports, and switch metric configuration retention on cloned tabs.
Fixed map visualization issues including: collapsible range controls in location maps and range update behavior across all map types.
Resolved feed-related issues including: edit feed dialog, three-dot menu not responding, and date change behavior when setting feed schedules.
Fixed miscellaneous Vizpad issues including: text box displacement, image viz upload in view mode, notification "View Details" navigation, chart placeholder image differentiation, unnecessary scroll-up on panel close, and BV configuration panel search and column visibility.
Data and Business Views
Resolved an issue where joining tables could throw an error during data preparation.
Fixed an issue where Insight, Vizpad, and Predict list view objects would not load when scrolling.
Improved navigation stability when returning to the Vizpads landing page after saving changes in Vizpad.
Fixed an issue where dynamic match formulas in pivot tables did not respect reporting granularity for advanced pivot totals and subtotals.
Related fixes ensure that Point-In-Time settings are correctly sent in the payload when adding column groups and publishing, that PIT-enabled metrics return accurate values when date range filters are applied, and that dedup-enabled Business Views correctly validate PIT configurations against primary date columns.
Embedding
Improved embed mode behavior so that tab navigation is correctly scoped and session continuity is maintained when switching between tabs.
Last updated
Was this helpful?