June 2026¶

The following enhancements, fixes, and completed work shipped in Aunsight Data Services during June 2026. Changes are summarized below by feature area.

Vector Search on Accelerated Warehouse¶

A major new platform capability: semantic similarity (vector) search as a first-class feature of the accelerated warehouse (AWHS). No separate embedding pipelines or bespoke vector database management required.

New — Vector search is now a managed feature of the accelerated warehouse. Solution authors designate string fields for embedding in the ERM table spec (up to three vector fields per table); ADS handles embedding, indexing, and lifecycle automatically.
New — Compute-environment configuration for an external embedding provider (via a LiteLLM proxy) and a platform-managed Milvus cluster — solution authors never interact with Milvus directly.
New — Table lifecycle migration actions added: enable-vector-search, alter-vector-search, and disable-vector-search, with Milvus collection management handled automatically and guards preventing the removal of fields in use by vector search.
New — Automatic embedding on load: dense and sparse embeddings are computed for incoming data during the LWHS→AWHS transfer, loaded into a new versioned Milvus collection, and made live via an atomic alias swap for zero-downtime refresh.
New — Embedding cache: previously computed embeddings are cached and reused for unchanged rows, so only added or changed rows are re-embedded — keeping load times manageable as data volumes grow.
New — Hybrid search (dense + sparse) via a weighted reranker. Queries accept a plain text string, an optional scalar filter, a dense/sparse weight, and a result limit; ADS computes the query embedding internally and returns matching rows with vector-score and matched-field metadata.
New — Vector search is exposed through the Consumption API for consumption-app-enabled AWHS tables, with the same installation-level scoping and authorization applied to warehouse queries generally.
New — Sparse embedding model (bge-m3) exposed through the LiteLLM AI gateway so all model traffic routes through a single managed gateway.
New — S3 bulk import path added for large vector-search loads, streaming pre-formatted embedding data via S3 rather than row-by-row inserts for significantly higher throughput on full loads and large reloads.
New — Milvus client performance tuning: after benchmarking the HTTP client against the gRPC client across deployment modes, vector search settled on the faster gRPC client path, with reduced insert batch sizes and streamed bulk-import preparation to avoid oversized payloads and memory pressure at scale.
New — Production Milvus and LiteLLM deployments provisioned for production compute environments as infrastructure prerequisites, isolated from non-production instances.
New — Vector-search usage documented for solution authors and operators, covering compute-environment setup, ERM spec configuration, load behavior, the query interface, and consumption-API usage.

Local Bucket (LBKT) — Unstructured File Storage¶

A new compute-environment-managed unstructured file storage container, analogous to the Local Warehouse but for files and folders. Local Bucket is the primary unstructured-data storage mechanism in ADS, bringing remote documents into the platform for downstream extraction, chunking, and delivery to the Accelerated Warehouse.

New — Compute Environment and Compute Environment Tenant DTOs extended with a bucket_storage configuration alongside warehouse_storage, plus an optional path field on the Task Artifact DTO to track where each transferred object originated relative to its source root.
New — Bucket CRUD and Bucket-Location CRUD commands reviewed and shored up for LBKT — create, list, view, update, list-history, and list-activity all verified against Local Bucket storage.
New — ADS-managed manifest files (.ads-lbkt-manifest) maintained per prefix to track per-object metadata that S3 does not natively support: checksums, logical created/modified timestamps, object size, and task provenance. Manifests are reconciled and kept in sync on every write.
New — User Bucket → Local Bucket transfer task (bucket:transfer-to-bucket) implemented, moving files from a user bucket into a Local Bucket using a two-phase commit so content is only persisted once fully assembled without error.
New — Sub-path support added to the transfer task, so transfers can target or source a specific path within a bucket rather than only the bucket root.
New — Full set of configurable merge settings implemented for transfers — pre/post state assertions and granular handling for source-only, destination-only, conflict, no-conflict, and type-conflict cases — giving precise control over how differing content is reconciled.
New — Bucket-Location support added to the transfer task, letting scripts reference a pre-declared bucket location (path plus transfer directives) by ID instead of embedding configuration inline.
New — bucket:list-objects task implemented for LBKT (renamed from bucket:list-paths), returning per-object metadata — type, timestamps, size, and checksum — from the manifest, with recursion, max-depth, include, and exclude options.
New — bucket:remove-objects task implemented for LBKT, removing objects at or under specified paths, producing a list of what was removed, and updating affected manifest files.
New — Remote Bucket → Local Bucket transfer task implemented — the primary acquisition path for unstructured content, pulling raw files from an external source (via a Sync Agent plugin) into the Local Bucket.
New — Local Bucket → Local Bucket transfer task implemented, enabling intra-environment file movement for staging, intermediate processing, and pipeline handoffs within a single compute environment.

Solution Builder — Actions¶

Core capabilities for Actions within the Aunsight 3 Solution Manager.

New — Manage Solution Actions: define, configure, edit, and execute solution-specific actions (e.g. recovery, recalculation) from the Solution Manager UI. Actions are available in Draft and Execution modes, can be triggered manually or linked to schedules, and surface execution status and logs.
New — Actions Tab: a dedicated tab in the Solution Manager that lists all configured actions with their key metadata — action name, ID, and description — and a Run Action button to execute an action directly from the list.

Workgraph Submission Recovery & Schedule Phase¶

The Schedule-Oriented phase epic was completed in June. The schedule-metrics feature work (solution framework changes, task metadata, Solution Manager monitoring, email notifications) and the initial Submission Trace UI and recover-submission option were delivered and reported in the May 2026 Release Summary. The items below are the net-new enhancements completed since then; the epic completion is also recorded here.

New — Workgraph Submission Recovery now supports partial recovery: operators can apply optional node and path filters to recover a specific portion of a stalled workgraph submission rather than re-running the entire workflow.
New — Schedule Phase Development epic completed, closing out the schedule-oriented execution work for the solution framework.

Bug fixes¶

Defects resolved during June 2026.

Fix — Arithmetic operator precedence in dataflow expressions now evaluates as written — addition combined with multiplication (e.g. (1900 + field) * 10000) produces the expected result without reordering operations.
Fix — Division operations on float fields now return the expected value without requiring the dividend to be pre-rounded to a fixed number of decimal places.
Fix — Bucket transfers now respect user-level permissions — transfers performed as a non-admin user no longer fail where they previously only worked as an admin.