live wire
▸Gemma-4-E2B text-to-SQL adapter ships weights but no SQL evaluation, dataset disclosure or task-specific exampleHugging Face model card▸Qwen2.5-3B GRPO adapter posts 62.67% Spider dev execution accuracy, 1.8 points above base; base leads execution successHugging Face model card▸NextGen Healthcare opens N.I.A. analytics agent to early adopters; all Insights clients get it in early NovemberNextGen Healthcare▸Overseas Tax Agent posts a 200-case synthetic eval; tax math is deterministic, but SQL and RAG tools remain stubsGitHub repository▸Databricks says Genie rose from 32% to 90%+ on an internal benchmark; task count and artifacts are undisclosedDatabricks▸Tableau says agent traces and human corrections continually reshape its runtime knowledge graphTableau▸ThoughtSpot urges separate read/write permissions and explicit approval for production-impacting agent actionsThoughtSpot▸SchemaGate v1.11.1 drops sentence-like notes after complete SQL; 28,785-reply replay changed one trailing commentSchemaGate▸Teradata adds agent-aware ABAC, identity reduction and immutable backups for enterprise AI dataTeradata▸Huel says it moved staff to Spotter-first analytics after a direct Snowflake-to-Claude trial lacked explainabilityThoughtSpot customer essay▸KG mediation reaches 89.47%/90.29% F1 on 158 enterprise prompts; only 30 receive preliminary end-to-end trialsECIS 2026 paper▸MotherDuck says hiding query_rw is insufficient for MCP read-only mode; other write/delete tools need separate reviewMotherDuck▸Oracle AIDP adds progressively disclosed Agent Skills with optional declared Python entrypoints for high-code agentsOracle▸ThoughtSpot says Spotter 3 may retry forecasts with a different model after evaluating confidence intervalsThoughtSpot▸Gemma-4-E2B text-to-SQL adapter ships weights but no SQL evaluation, dataset disclosure or task-specific exampleHugging Face model card▸Qwen2.5-3B GRPO adapter posts 62.67% Spider dev execution accuracy, 1.8 points above base; base leads execution successHugging Face model card▸NextGen Healthcare opens N.I.A. analytics agent to early adopters; all Insights clients get it in early NovemberNextGen Healthcare▸Overseas Tax Agent posts a 200-case synthetic eval; tax math is deterministic, but SQL and RAG tools remain stubsGitHub repository▸Databricks says Genie rose from 32% to 90%+ on an internal benchmark; task count and artifacts are undisclosedDatabricks▸Tableau says agent traces and human corrections continually reshape its runtime knowledge graphTableau▸ThoughtSpot urges separate read/write permissions and explicit approval for production-impacting agent actionsThoughtSpot▸SchemaGate v1.11.1 drops sentence-like notes after complete SQL; 28,785-reply replay changed one trailing commentSchemaGate▸Teradata adds agent-aware ABAC, identity reduction and immutable backups for enterprise AI dataTeradata▸Huel says it moved staff to Spotter-first analytics after a direct Snowflake-to-Claude trial lacked explainabilityThoughtSpot customer essay▸KG mediation reaches 89.47%/90.29% F1 on 158 enterprise prompts; only 30 receive preliminary end-to-end trialsECIS 2026 paper▸MotherDuck says hiding query_rw is insufficient for MCP read-only mode; other write/delete tools need separate reviewMotherDuck▸Oracle AIDP adds progressively disclosed Agent Skills with optional declared Python entrypoints for high-code agentsOracle▸ThoughtSpot says Spotter 3 may retry forecasts with a different model after evaluating confidence intervalsThoughtSpot
nl2sql.ai
guideDEPLOYMENT GUIDE

Tableau’s knowledge graph learns from agent traces—but human corrections still anchor it

The useful deployment pattern is not a static semantic layer. It is a runtime map whose failed paths reveal gaps and whose approved definitions prevent those gaps becoming policy.

By The News Desk· Oct 25, 2026the quick take — two AI hosts go live when you do

Tableau’s latest architecture essay makes a useful distinction for teams building analytics agents: a semantic layer gives business terms meaning, an ontology constrains valid relationships, and a knowledge graph gives the agent a structure it can traverse while planning a task. Tableau argues that these are not three disconnected systems, but different views of the same enterprise knowledge used at runtime. Tableau

What changes at runtime

In Tableau’s description, an analytics agent does more than retrieve a stored answer. It maps an objective to business concepts, expands through relationships, executes against data and tools, and adjusts when results fail. Semantic models define terms such as “pipeline” or “active customer”; ontology captures entities, valid joins and constraints; the graph connects those concepts to data assets and operational processes. Tableau

That separation matters operationally. A model can know a metric’s definition yet still choose the wrong relationship or process. Conversely, a graph can connect the right assets while leaving an ambiguous business term unresolved. Tableau’s proposed knowledge engine combines meaning, structure, context and relationships so the agent has both definitions and possible routes. Tableau

Traces are diagnostics, not automatic truth

The strongest practical point is Tableau’s feedback loop. Every completed task produces traces: which paths the agent selected, which queries succeeded or failed, and where it retried or changed course. Tableau says those traces can expose missing relationships and ambiguous mappings. Human input supplies a different signal through curated semantic models, refined metrics, verified queries, corrected outputs and maintained documentation. Tableau

Teams should preserve that distinction in deployment. A frequently selected path is evidence of usage, not proof that the path is correct. Failed queries can identify a gap, but they do not decide the approved fix. Tableau’s own architecture assigns stabilization and correction to explicit human input while usage reveals incompleteness. Tableau

A rollout test

Before treating a knowledge graph as self-improving, operators should test three loops separately: whether task traces are captured, whether failed or retried paths create reviewable diagnostics, and whether approved human corrections update the knowledge used by later tasks. They should also retain the semantic definition, relationship change and validating query behind each correction. That test follows directly from Tableau’s split between usage signals and curated knowledge. Tableau

Tableau does not publish a benchmark in this essay showing how much the feedback loop improves accuracy or how quickly corrections propagate. The defensible takeaway is architectural: runtime traces can show where an agent’s business map is weak, but governed human input remains the mechanism that turns those observations into trusted knowledge. Tableau

sources

Filed by The News Desk. Corrections: desk@nl2sql.ai · Our standards →

comments · 0

    Comments are moderated before they appear. Your email is used once to confirm it is you — never shown, never sold. Corrections and questions get an answer from the desk when we have one.

    more from the desk