A small semantic-layer release makes distributed cancellation safer and malformed success responses explicit—both important when software, not a person, is driving the query loop.
Version 1.14 measures how much schema selection can shrink an NL2SQL prompt—and how often the necessary tables disappear with it.
Version 1.13 lets the model inspect a few real values before committing to SQL. The gains are large on BIRD and uneven elsewhere, making the release useful as both a feature and an evaluation lesson.
Microsoft’s multi-step query generation moves into Fabric’s standard runtime Nov. 2–10, alongside GPT-5.6 Terra; the vendor’s own numbers are strong, but they are internal.
The release adds OpenMetadata project import and Apache Ossie context exchange, while fixing several SQL and metadata correctness edges.
Private Agent Factory 26.10 adds topic-level context, clarification, SQL review and pre-publish evaluation instead of letting user feedback rewrite shared knowledge automatically.
The failure did not throw an error: it could produce a plausible but incorrect ranking when ORDER BY and LIMIT landed in different query levels.
The new MIT-licensed MCP server turns PostgreSQL, SQLite and spreadsheet queries into inspectable boards, while its cost and speed claims remain narrowly author-tested.
The new self-hosted MCP pipeline catches syntax and execution failures, yet production teams still need an independent semantic check and licensed q infrastructure.
The semantic-core patch landed first; the Python binding followed 46 minutes later. Pin and test the package your application actually loads.
The new AGPL project is unusually complete for a day-one release; its default credentials and unproven SQL guard still make a sandbox the right starting point.
Release 1.10 reports a jump from 35.0% to 64.5% first-try accuracy on 1,534 UniQL Oracle questions with Gemini 2.5 Flash. The result is promising, but still project-reported.
The GA service offers roughly 30-second live views, transactional apply and a seven-day change feed. Choosing the wrong query surface can still give an agent a stale or unexpectedly broad view.
Apache Ossie support and a rewritten native YAML model widen interoperability. Teams should diff both manifests before making the new parser or metric schema part of production.
Version 15.0 packages OAuth, optional Vault-backed source credentials and SQL guidance into a containerized MCP server. Teams should verify what the Community Edition actually prevents before connecting an agent.
Semantic core 0.5.0 lets Rust applications enforce custom row and column rules while SQL is planned, with fail-closed errors and explicit limits around masking and language bindings.
PDF, Word, PowerPoint and HTML generation is generally available, yet every export depends on a code sandbox with its own rights, persistence and approval boundaries.
The new output_schema argument makes semantic aggregation easier to feed into SQL pipelines, while error thresholds, multimodal limits and token caps still define the production boundary.
The synchronous APIs turn incoming files into structured agent inputs; they do not make every downstream decision production-ready.
The new Apache-2.0 workbench puts retrieval, context and candidate selection behind a bounded experiment loop, while explicitly refusing to turn local search into a universal accuracy claim.
Native Apps now redact orchestrator-authored content in consumer logs and require explicit approval before providers receive content-level telemetry.
The October release gives existing deployments a planning agent, a Teams interface and a source-side WordPress connector. The operational gain is real; so is the need to review every live system the agent can reach.
The ai_decide beta returns probabilities, named choices and rubric scores over governed data. Its speed is useful; its probabilistic output still needs explicit escalation and review rules.
OpenSharing and system-table triggers remove polling from one part of the loop; batching, unsupported dependencies and filtered views still define the freshness boundary.
DCM Projects now manages metrics, relationships, AI instructions and verified queries declaratively—but its governance envelope still has edges.
The new PostgreSQL text-to-SQL stack separates its safety controls from its prompt—and labels every planned BIRD number as unmeasured.
The launch notice and model card give opposite answers; production teams should treat caching as unverified until AWS reconciles them.
Microsoft’s MCP Tasks support turns long analytics questions into resumable jobs—but clients must preserve task IDs and implement cancellation correctly.
Google’s managed endpoint exposes creation, update, run and deletion controls, so the safety boundary must cover ingestion operations—not only generated SQL.
A new MIT repository separates probabilistic question analysis from ontology-backed planning and deterministic SQL compilation—but launches without a README or benchmark.
The release exposes context, reasoning effort, routing geography and service tier as separate deployment choices, while structured output and invocation logs remain the practical guardrails.
Version 0.17.0 exposes selected database connections to AI clients, but puts enforcement, auditability and hard limits below the model layer.
The flaw reached both ordinary queries and the agent path, showing why read-only execution and connection isolation must be tested separately.
The 5.3.7 beta now writes a uniquely tagged plan, reads it back through DBMS_XPLAN and restores the client transaction—turning a success message into an inspectable result.
The release adds orderBy to the agent-facing cube tool and fixes query-rewrite edge cases that could change or break generated SQL.
The stable release adds persisted cron jobs, ownership-aware authorization and metrics—the controls that matter after an analytics agent leaves the demo.
The experimental feature turns live workspace data into branded PDF and PowerPoint deliverables, but explicitly stops short of autonomous reporting.
A new beta separates permission to create a governed metric from permission to read, share or edit it—and blocks calculations when any dependency is hidden.
Local metric views let teams prototype semantic logic inside one AI/BI dashboard, then promote the definitions when they are ready for wider reuse.
The MCP server now closes the loop from schema selection to read-only query execution, while explicitly warning that principal and role claims are not authenticated.
The release reports a three-question retrieval gain across 98 tests, while preserving the hashed model for base installs and adding an opt-out for persisted deployments.
Microsoft’s Build Agent with AI preview turns schema and conversation context into instructions, source guidance and examples. Teams should treat that output as proposed configuration, not production policy.
Personal and company-scoped procedures can now steer AI Analyst, while the public marketplace still documents Claude and n8n as today’s execution surfaces.
The new Playground panel turns natural-language requests into runnable Visual Embed SDK code, while leaving execution and conflict resolution with the developer.
Adaptive Instructed-Retriever decides when one pass is enough and when a harder request merits more search, while keeping a fixed step ceiling.
The alpha Python package now derives dimension-like tables from the join graph, while prompt and validator fixes unblock CTEs, window functions and multi-table joins.
The release adds CSV mapping and MySQL transaction inspection, but operators should prioritize its SSH-log redaction and connection-close API change.
The September release adds natural-language filter expressions, generated column descriptions and consumer-facing Auto Insights—but keeps review and model choice with administrators and authors.
The new Apache-2.0 package filters schema objects before they reach a model. That closes one information-leak path while leaving query execution controls firmly in place.
Select AI’s new ADB-D capabilities collapse NL2SQL, feedback, memory and tool orchestration into database objects. The operational burden moves, not disappears.