Release Notes -- 2026-03-04¶

Databricks Forge v0.6.0

New Features¶

Ask Forge SQL Column Grounding¶

Ask Forge no longer fabricates column names when generating SQL. A four-layer defense-in-depth approach ensures the LLM always sees exact column metadata: - Layer 1 -- Schema snapshot: A compact schemaSnapshotJson field is now persisted on ForgeRun at the end of metadata extraction, storing all table/column metadata with an isBusinessTable flag stamped after table filtering. This covers pipeline-only users who never ran an estate scan. - Layer 2 -- Enrichment columns: fetchTableEnrichments() now parses ForgeTableDetail.columnsJson and includes per-table column schemas with backtick-quoted names in the context injected into the LLM. - Layer 3 -- Backtick quoting rules: DATABRICKS_SQL_RULES now includes explicit identifier quoting rules. The assistant system prompt has been hardened with column name fidelity instructions that prohibit normalising, snake_casing, or transforming column names. - Layer 4 -- EXPLAIN fix cycle: After SQL extraction, each SQL block is validated via EXPLAIN. If a column resolution error is detected, a single LLM fix attempt is made with the exact error message and correct column schema. The fix is only accepted if the corrected SQL passes EXPLAIN.

Ad-hoc Dashboard Engine¶

New dashboard deployment flow from Ask Forge conversations. Users can deploy dashboards directly from chat by clicking "Deploy as Dashboard" when the assistant proposes relevant SQL and tables.

Pipeline Cancellation¶

Users can now cancel in-progress pipeline runs directly from the UI. A new cancelled status has been added to the run lifecycle, and the pipeline engine gracefully halts step execution when cancellation is requested. Cancelled runs can be resumed later.

Introduced a dedicated Genie Builder modal component (genie-builder-modal.tsx) extracted from the inline page, providing a cleaner UX for creating Genie Spaces from Ask Forge. Users can now configure question complexity (simple, medium, complex) when generating example queries, allowing tailored output for different audience levels.

Custom App Name for Deployments¶

The deploy.sh script now accepts an --app-name option for multi-instance Databricks App deployments, enabling isolated naming across environments. Updated QUICKSTART.md with usage documentation.

Superannuation Industry Outcomes¶

Added a new superannuation-specific industry outcomes module with domain-relevant KPIs and use case templates.

Improvements¶

Standardised LLM JSON Parsing¶

Replaced ad-hoc parseJSONResponse calls across all pipeline steps with the robust parseLLMJson utility. This provides consistent multi-strategy parsing (direct parse, fence extraction, bracket extraction, repair, and truncation recovery) with structured error logging throughout the codebase.

`parseLLMJson` Caller Diagnostics (in progress)¶

Added an optional caller parameter to parseLLMJson and tryRepairAndParse so that log entries and error messages identify which pipeline step or module triggered the failure, significantly improving debuggability.

SQL Generation: Hallucinated Column Detection & Fix Cycle (in progress)¶

The SQL validation step now detects hallucinated columns (columns referenced in generated SQL that don't exist in the schema). When found, a targeted fix prompt is sent to the LLM with an explicit list of valid columns per table. If the fix still contains invalid columns, the SQL is rejected rather than returned with errors.

SQL Fix Prompt Now Includes Sample Data (in progress)¶

The SQL error-fix template and attemptSqlFix function now pass sample data context to the LLM, giving it concrete examples of actual column values when repairing broken queries.

maxTokens Tuning¶

Increased maxTokens across 17 components including environment intelligence, Genie passes (benchmarks, queries, instructions, joins, titles, trusted assets), pipeline steps (domain clustering, scoring, table filtering, use case generation), assistant engine, and export summaries to accommodate larger and more complex LLM responses.

AI Query Result Metadata¶

Added finishReason to the AIQueryResult interface, capturing whether model output was complete or truncated -- enabling downstream logic to detect and handle incomplete generations.

Environment Page Component Extraction¶

Refactored the monolithic environment page (1,400+ lines) into focused components: aggregate-summary, data-maturity-card, executive-summary, governance-quality-view, scan-progress-card, scan-trends-panel, single-scan-summary, stat-card, and table-coverage-view.

API Error Standardisation¶

Introduced lib/api-error.ts and lib/fetch-json.ts utilities for consistent error response formatting across all API routes.

Genie Space Refactoring¶

Cleaned up mergeSpaces by removing unused trackedBySpaceId mapping. Switched NewGenieSpacePage to use a ref for jobId instead of state to prevent unnecessary re-renders. Removed the buildGenieUrl helper in favour of inline logic.

Bug Fixes¶

Backtick-quoting for column names with spaces -- Column names containing spaces (e.g. Origination Quarter, Loan Duration (Months)) are now backtick-quoted throughout the Genie Engine pipeline. Previously, the time-period generator produced invalid SQL like table.Origination Quarter which the identifier validator flagged as unknown, silently dropping all auto-generated time filters and dimensions for affected columns. Now produces valid table.\Origination Quarter`references. The schema-allowlist validator, assembler join rewriter, and metric view FQN stripper all handle backtick-quoted identifiers. Schema context blocks (buildCompactColumnsBlock,buildSchemaContextBlock`) now present space-containing columns as backtick-quoted to the LLM.
Metric view: measure name shadowing detection -- Added static validation detecting metric view measures whose names are identical to source column names. In Databricks metric views, measure names take priority over column names in the shared namespace, causing the column reference inside the expr to resolve to the measure itself -- producing a recursive NESTED_AGGREGATE_FUNCTION error. The validator now catches this before dry-run and the LLM prompt explicitly prohibits it.
Metric view: backtick-quoted column validation -- The validateColumnReferences function now parses backtick-quoted column references (e.g. lending.\Defaulted Loans``). Previously, hallucinated columns in backtick-quoted form slipped past static validation and only failed during dry-run SQL execution. They now trigger the LLM repair loop correctly.
Metric view: expanded repair loop triggers -- Added NESTED_AGGREGATE, FIELD_NOT_FOUND, and measure shadowing patterns to the error patterns that trigger the LLM repair loop, so dry-run failures for these error types get a chance at automated repair.
SQL column validation threshold lowered -- Previously, unknown column warnings only fired when more than one hallucinated column was detected. Now triggers on any unknown column, catching single-column hallucinations that were previously silently passed through.
Removed stale middleware.ts -- Eliminated the unused authentication/rate-limiting middleware that was interfering with Next.js routing. The file has been renamed to proxy.ts to preserve any reusable logic.

Other Changes¶

Added iaisweb.org to the public host allowlist for benchmark source fetching.
Added Prettier configuration (.prettierrc, .prettierignore) for consistent code formatting.
Added rate-limiting utility (lib/rate-limit.ts) for API route protection.
Genie trash preview component added for space deletion confirmation UX.
Pipeline run detail page now includes config field, coverage gap card, and summary card sub-components.

Commits (8)¶

Hash	Summary
`89ba443`	Enhance deployment script and API functionality
`5427ba0`	Add iaisweb.org to public host allowlist
`5051052`	Refactor Genie space handling and update API routes
`77dd219`	Enhance Genie query generation and question complexity
`bb7447f`	Remove middleware.ts, rename to proxy.ts
`bf103b6`	Update AI query handling and response parsing
`8a51124`	Implement pipeline cancellation feature
`d811362`	Update maxTokens parameters across components

Uncommitted changes: Ask Forge SQL column grounding (context-builder, engine, sql-proposer, prompts, sql-rules), schema snapshot persistence (metadata-extraction, table-filtering, runs, Prisma schema), ad-hoc dashboard engine, deploy-dashboard-dialog enhancements, and prior uncommitted work (SQL hallucination fix cycle, sample data in fix prompts, parseLLMJson caller diagnostics, backtick-quoting for spaced column names, measure shadowing detection, expanded metric view validation and repair loop).