Skip to content

Release Notes -- 2026-03-09

Databricks Forge v0.15.0 → v0.18.0


v0.15.0 -- Design System Overhaul

New Features

Design System Overhaul

Complete visual redesign moving from scaffold defaults to a branded, intentional design language. The new "Precision Intelligence" aesthetic transforms every surface of the application from a developer workbench into a polished business tool.

Typography

Replaced Geist (the Next.js default) with Plus Jakarta Sans for body/headings and JetBrains Mono for code, creating a distinctive typographic identity.

Motion System

Installed motion (Framer Motion v11+) and created shared animation variants. The sidebar now uses spring-based collapse/expand, page transitions use fade+slide via template.tsx, and dashboard cards enter with staggered animations.

Tile Dashboard Layout

Redesigned the home page with an atmospheric hero banner featuring a Databricks diamond texture pattern, an asymmetric tile layout with a hero use-case metric, quick-action cards, secondary KPI tiles, side-by-side recent runs and activity feed, and a refreshed empty state with branded onboarding.

Improvements

Replaced flat bg-muted/30 with gradient background, stronger active states with left-border accent in Databricks Red, refined section labels, and animated brand text on collapse/expand.

Header Refinement

Increased height from 48px to 56px, grouped right-side actions with a visual divider, strengthened page title weight, and added subtle bottom shadow for depth.

Consistent Page Layouts

All 23 pages now use max-w-[1400px] content constraint, space-y-8 section spacing, and standardized PageHeader component where applicable. New PageShell and PageHeader shared components enforce consistency.

Component Refinements

  • Card: Added hover transition support and tighter title tracking
  • Badge: New brand variant using Databricks Red with subtle border
  • Button: Primary buttons now have shadow, hover shadow, and active scale
  • Skeleton: Brand-tinted shimmer instead of generic grey pulse

Error & Status Pages

Refreshed 404 page with branded icon and atmospheric large-number background, refined error boundary with icon treatment and action buttons, and loading skeletons that mirror the new tile layout.

Ask Forge

Updated layout margins for the new header height, refined persona toggle and conversation history backgrounds.


v0.16.0 -- LLM Skills System + Genie Detail Page

New Features

LLM Skills System

Introduced a structured knowledge injection framework (lib/skills/) that enriches all AI surfaces with domain-specific expertise. Skills are registered in a central registry and resolved at prompt-build time based on context signals (industry, data patterns, intent). Ships with five built-in skill packs: Databricks SQL Patterns, Genie Space Design, Industry Enrichment, Metric View Patterns, and System Tables. Skills are embedded into the vector store for RAG retrieval and injected directly into LLM prompts, improving generation quality across the Genie Engine, Ask Forge, and pipeline steps.

Genie Space Detail Page Uplift

Completely redesigned the Genie Space detail page (app/genie/[spaceId]/page.tsx) with a tab-based layout featuring a new hero section (space-detail-hero.tsx), Overview tab (space-overview-tab.tsx) with table/instruction/expression summaries, Health tab (space-health-tab.tsx) with inline health checks and fix actions, Benchmarks tab (space-benchmarks-tab.tsx), and a refreshed Config Viewer. The page now matches the app's Precision Intelligence design language.

Genie Space Improvement Workflow

Added a background improvement job system (lib/genie/improve-jobs.ts) with new API routes (/api/genie-spaces/[spaceId]/improve and /api/genie-spaces/improve-status). Users can trigger space improvements from the detail page and poll for progress, with status indicators and result surfacing in the UI.

SQL Extraction in Space Fixer

Enhanced the space fixer (lib/genie/space-fixer.ts) with SQL extraction and reference handling capabilities, enabling more precise automated fixes when resolving health check failures.

Improvements

Design Restoration -- Outcomes & Ask Forge

Restored the design-overhaul UX to the Outcomes page, Use-Case Table, and Ask Forge chat interface that had been lost during branch merges.

Dynamic Application Naming

Updated lib/lakebase/provision.ts to support a FORGE_APP_NAME environment variable for PROJECT_ID_BASE and display name.

Layout & Spacing Consistency

Updated layout and spacing across multiple pages (Fabric, Environment, Metadata Genie, Run Detail) for consistent margins, padding, and content constraints matching the design system.

Trusted Assets Pass Enhancements

Expanded the trusted assets pass (lib/genie/passes/trusted-assets.ts) with richer generation logic and better handling of SQL patterns.

Benchmark Generation Improvements

Enhanced the benchmark generation pass with more robust question generation and better context handling for improved Genie Space benchmark coverage.


v0.17.0 -- LLM Skills Overhaul (8 Modules, 12 Surfaces)

New Features

LLM Skills Overhaul -- 8 Modules, 12 Blind Surfaces Wired

Comprehensive overhaul of the in-app LLM skills system. Created four new skill modules (Databricks Data Modeling, AI Functions, SQL Scripting, Dashboard SQL) and substantially rewrote four existing modules (SQL Patterns, Genie Design, Metric View Patterns, System Tables). Extended the type system with new GeniePass values (exampleQueries, joinInference) and PipelineStep value (dashboard-design). Wired the skill resolver into 12 previously blind LLM surfaces so every AI-powered generation path now benefits from structured domain knowledge.

Improvements

SQL Generation Craft Skills

Rewrote databricks-sql-patterns to focus purely on SQL generation craft: window function recipes, CTE composition patterns, date idioms, lambda/higher-order function patterns, and a comprehensive anti-patterns catalogue.

Databricks Data Modeling Skill

New dedicated skill covering star schema design, keys and constraints, Liquid Clustering best practices, SCD Type 1/2 patterns, and data modeling anti-patterns. Wired into table-filtering and SQL generation pipeline steps.

Databricks AI Functions Skill

New skill covering ai_query rules and structured output, the cost-optimisation funnel, and all task-specific AI functions (ai_classify, ai_extract, ai_forecast, vector_search, http_request, remote_query, read_files).

Databricks SQL Scripting Skill

New skill for procedural SQL: BEGIN...END blocks, variable declaration, control flow, exception handling, stored procedures with IN/OUT/INOUT parameters, and recursive CTE patterns.

Dashboard SQL Skill

New skill for Lakeview dashboard dataset SQL rules, widget field matching constraints, 6-column grid layout rules, and metric view SQL compliance.

Enhanced Genie Design Skill

Added five new chunks: snippet expression complexity rules, alias-based join inference patterns, auto-generated date filters, question style guidelines, and instruction anti-patterns.

Enhanced Metric View Patterns Skill

Added star/snowflake join YAML examples, common gotchas, and window measure examples (YTD, rolling, period-over-period).

Enhanced System Tables Skill

Added deeper coverage: job/pipeline monitoring, storage/compute metrics, and data product observability patterns.

Skill Resolver Wired to 12 LLM Surfaces

Previously blind surfaces now receive skill context: SQL Reviewer, Pipeline SQL Generation, Dashboard Engine, Genie Example Query Generation, Genie Join Inference, Genie Adhoc Engine, Genie Optimise, Genie Space Fixer, Health Check Synthesis, Fabric Gold Proposer, Fabric DAX-to-SQL, and Pipeline Table Filtering.


v0.17.1 -- SQL Hallucination Fix

Bug Fixes

  • False-positive hallucination flags on AI use case SQL -- The column validator was rejecting valid SQL that used from_json() struct field access (e.g. parsed.confidence). The validator now recognises column-alias prefixes as struct field access and skips them during validation.

  • String literal content matched as column references -- Prose inside SQL string literals was being matched by the column reference regex. Added stripStringLiterals() to the extraction pipeline.

  • Reviewer fixes introducing new hallucinated columns -- The SQL reviewer fix prompt now includes an explicit column-grounding constraint.

Improvements

LATERAL VIEW EXPLODE Deprecation

Added rules across all prompt surfaces warning against LATERAL VIEW EXPLODE (deprecated Hive syntax that causes parse errors when combined with JOINs).

Canonical from_json() Alias Naming

Enforced ai_result and parsed_result as the mandatory alias names for ai_query() and from_json() outputs respectively.

Targeted Fix Guidance for from_json Violations

When the hallucination fixer detects parsed.* prefix patterns in flagged columns, the error message now includes specific instructions about canonical from_json() alias naming.


v0.18.0 -- Pipeline & Engine Performance (17 Areas)

Improvements

Parallelized pipeline steps for faster execution

Domain-clustering (Step 5) and scoring/dedup (Step 6) now process domains in parallel with bounded concurrency, and LLM batches within each step run concurrently via Promise.all. Metadata extraction (Step 2) runs all per-scope queries in parallel. Expected 3-5x speedup on steps 2, 5, and 6.

Genie Engine pass reordering and concurrency

Metric view proposals now run in parallel with Phase 3 passes instead of sequentially before them. Genie LLM concurrency default raised from 3 to 6. Metric view SQL reviews now run in parallel.

Run-level sample data cache

SQL generation (Step 7) now caches sample data per table FQN across use cases within a run, eliminating 50-80% of redundant warehouse queries.

Parallelized embedding batches

Embedding generation now runs up to 3 batches concurrently with increased batch size (32, up from 16), reducing embedding phase latency by 2-3x.

SQL reviewer LLM cache for Genie surfaces

SQL review calls for genie-* surfaces are now routed through the Genie LLM cache, avoiding redundant model serving calls.

Buffered prompt log inserts

Prompt log entries are now buffered in memory and flushed via createMany every 20 entries or 2 seconds, reducing DB round-trips.

Progress update throttling

updateRunMessage calls are now debounced to at most one write per 500ms.

Reduced SQL generation wave delay

Default SQL_GEN_WAVE_DELAY_MS reduced from 2000ms to 500ms.

Frontend polling backoff

All status polling now uses progressive backoff (starting at 2s, scaling to 8-10s) instead of fixed intervals.

Cache-Control headers on read-heavy API routes

Added s-maxage + stale-while-revalidate headers to environment/table, table-coverage, outcome-maps, genie-spaces, and aggregate ERD endpoints.

Lazy-loaded CodeMirror editor

The SQL editor component is now dynamically imported with ssr: false, removing ~200KB from the initial Ask Forge bundle.

Batched N+1 update in enrich-pbi

Individual prisma.update calls replaced with a single prisma.$transaction batch.

Deduplicated aggregate estate view queries

Added a 30-second in-memory TTL cache for getAggregateEstateView().


All Commits

Hash Summary
0591d98 Design system overhaul: replace scaffold defaults with branded, intentional UI
c51529c Add industry context handling in assistant context builder
8bd4f19 Enhance evaluation framework and persona handling in assistant
1608d53 Add resource prefix support for Unity Catalog resources and enable metric view generation
167dbd5 Restore design-overhaul UX for outcomes page and use-case table
3552a07 Restore Ask Forge design overhaul from stashed design-system-overhaul branch
18f3c85 Update PROJECT_ID_BASE and DISPLAY_NAME logic in provision.ts for dynamic app naming
1e299e5 Uplift Genie Space detail tabs to match app design quality
9e012eb Add LLM skills system for structured knowledge injection across all AI surfaces
d95c7b9 Add SQL extraction and reference handling in space fixer
1884492 Implement improvement polling and status handling in Genie pages
5b179cb Update layout and spacing for consistency across multiple pages
29dcd04 Refactor OutcomesPage and IngestOutcomeMapPage for improved state management and UI consistency
4aa9c57 Overhaul LLM skills: 8 skill modules, 12 blind surfaces wired
246a27d Fix false-positive hallucination flags and tighten SQL generation quality
307032e Improve pipeline and engine performance across 17 areas