Repository navigation
Antfly version v0.3.1
Documentation (2026-11-17)
The documentation is available here:
What's Changed
- fix(inference): isolate Whisper workspace and enable Arm64 OpenBLAS by @ajroetker in #1659
- Fix v0.3.0 Homebrew and embedded publication by @ajroetker in #1661
- Honor graph work limits and batch minimum-hop shortest-path reads by @ajroetker in #1657
- Make Dictation onboarding and output settings clearer by @ajroetker in #1656
- Prepare v0.3.0 SDK releases and independent Rust publishing by @ajroetker in #1663
- fix(inference): enforce Linux CPU budgets on Arm by @ajroetker in #1665
- Prune exact dense search with certified bounded scoring waves by @ajroetker in #1666
- fix(inference): honor configured native media chunk ceilings by @ajroetker in #1667
- Improve Dictation navigation, speech languages, and local cleanup latency by @ajroetker in #1664
- fix(storage): bound mapped vector descriptors through process admission by @ajroetker in #1669
- Fix Swift release-runner compilation of dictation cancellation fixture by @ajroetker in #1674
- Add opt-in Apple dictation prefix prewarming and cleanup benchmarks by @ajroetker in #1672
- fix(generating): initialize Apple backend state before early return by @ajroetker in #1676
- fix(inference): complete CPU media admission and bounded preparation by @ajroetker in #1671
- Fix main CI regressions and reduce repeated test setup by @ajroetker in #1660
- test: qualify descriptor pressure ingestion and real media execution by @ajroetker in #1675
- Accelerate GLiNER2.5-Decide CUDA and Metal inference for FP32 and Q8_0 by @timkaye11 in #1673
- fix(auth): AND row filters across all matching scopes by @ajroetker in #1683
- feat(inference): serve Fastino GLiDE-0.4B, 0.8B and 2B decision models natively on Metal and CUDA by @timkaye11 in #1649
- fix: restore native privacy text extraction and durable document routing by @ajroetker in #1686
- fix(inference): make GLiDE cleanup and CUDA measurements reliable by @ajroetker in #1688
- fix(inference): complete reader preparation admission and nested cancellation by @ajroetker in #1677
- test(inference): synchronize GLiDE measurement lifecycle phases by @ajroetker in #1691
- fix(licenses): preserve source comments after short SPDX notices by @ajroetker in #1692
- fix(auth): bound row-filter materialization before rejection by @ajroetker in #1690
- fix: preserve native text batching in mixed document cohorts by @ajroetker in #1689
- fix(inference): complete reader admission and move owned OCR text by @ajroetker in #1693
- fix: preserve per-item retries alongside publication waits by @ajroetker in #1694
- fix(inference): make reader transfers atomic and share CPU executors by @ajroetker in #1696
- fix(inference): separate retained executors from load-request I/O by @ajroetker in #1697
- Fix graph transport, catalog admission races, and CI unit resource use by @ajroetker in #1684
- feat: gate direct enrichments across storage, Raft, transactions and serverless by @ajroetker in #1685
- feat: add durable index aliases for atomic model cutovers by @ajroetker in #1687
- fix: pin dense inference and narrow serverless admission conflicts by @ajroetker in #1701
- Fix GLiNER CUDA batching limits and invalid-model errors by @timkaye11 in #1695
- Fix compiled metadata adapter for index alias validation by @ajroetker in #1703
- Fix privacy policy generations, protected index migration, and producer recovery by @ajroetker in #1707
- Bound missing-input replay retries under continuing writes by @ajroetker in #1708
- docs: draft v0.3 series changelog and upgrade notes by @ajroetker in #1711
- Expose bounded administrator inspection of protected dense vectors by @ajroetker in #1712
- Bound standalone readiness without redundant vector dependency scopes by @ajroetker in #1714
- feat(grepstein): explore PDFs and timestamped audio collections by @ajroetker in #1658
- test(inference): cover mixed Decide batches and malformed configs by @ajroetker in #1702
- perf(inference): accelerate Gemma 4 E4B CUDA inference on L4 by @timkaye11 in #1709
- fix: qualify index aliases across native deployment boundaries by @ajroetker in #1713
- docs: make EmbeddingGemma 2 the multimodal embedder, mark SQL and lakes shipped by @drewlanenga in #1715
- Fix protected index availability during unrelated model backfill by @ajroetker in #1717
- perf(inference): add opt-in host placement for Gemma4 PLE by @ajroetker in #1718
- docs: add SQL, foreign sources, Postgres extension, auth, PII, MCP, and deployment guides by @drewlanenga in #1720
- fix(enrichment): batch materialized backfills and defer vector snapshot repair by @ajroetker in #1721
- fix: batch protected migrations with independent publication and bounded maintenance by @ajroetker in #1719
- fix(antfarm): distinguish authentication and connection failures by @bpopadiuk in #1731
- fix: stabilize protected backfill ownership and vector publication by @ajroetker in #1733
Full Changelog: v0.3.0...v0.3.1
Released by Antfly release scripts.