Seyk v0.2.0

What's Changed

• Pin Rust for CI; check the newest stable monthly by @Naoray in #4
• Keyword index stores words without their order by @Naoray in #6
• Fix PR #5 review blockers: tie order, weight validation, file names by @Naoray in #7
• Hybrid search: meaning, keywords and file names in one query by @Naoray in #5
• docs: add project north star by @Naoray in #8
• Phase 2: URI resource ids and a resource provider by @Naoray in #9
• Phase 3: extract readable text from search hits by @Naoray in #10
• Reusable embedding crates: seyk-embed and seyk-embeddinggemma by @Naoray in #11
• docs(readme): new Seyk banner, badges, drop title heading by @Naoray in #13
• [agent] Add verified text PDF search and extraction (4a) by @Naoray in #12
• refactor(core): centralize text embedding job construction by @Naoray in #14
• [agent] Validate embedding output before caching by @Naoray in #16
• [agent] Add scanned PDF page retrieval by @Naoray in #17
• perf: scope PDF eligibility lookup to lexical search by @Naoray in #18
• Fix MCP inline sync ownership across overlapping watchers by @Naoray in #19
• [agent] Bind model identity and downloads to resolved snapshots by @Naoray in #20
• fix: serialize indexing through commits and orphan pruning by @Naoray in #21
• [agent] Fix daemon session admission and idle lifetime by @Naoray in #22
• fix: recover indexing after lost watcher events by @Naoray in #23
• [agent] Report verified semantic overlap in eval corpus by @Naoray in #24
• [agent] Bound UTF-8 segment splitting to the current block by @Naoray in #25
• [agent] Hydrate complete resources once per search by @Naoray in #26
• [agent] Reject corrupt vector bytes and zero-width indexes by @Naoray in #27
• [agent] Cache budget-rejected PDF preview pages by @Naoray in #28
• [agent] Measure cross-profile BM25 score and top-k contamination by @Naoray in #29
• [agent] Apply audit C1 documentation and PDF allocation hygiene by @Naoray in #30
• [agent] Type hydration test readback scenarios by @Naoray in #31
• feat: add opt-in audio retrieval with time evidence by @Naoray in #32
• fix(cache): isolate BM25 profiles and repair deletion statistics by @Naoray in #33
• feat: add CPU session controls and paired benchmarks by @Naoray in #34
• refactor(core): share PDF and audio resource commit boundary by @Naoray in #35
• feat: add public retrieval benchmark harness by @Naoray in #36
• feat(audio): stream verified recordings with configurable duration limits by @Naoray in #37
• Add offline on-demand speech-to-text extraction by @Naoray in #38
• refactor(audio): share descriptor identity between sync and search by @Naoray in #39
• chore: resolve unmaintained dependency advisories by @Naoray in #40
• perf: defer audio encoder hashing until first audio use by @Naoray in #44
• Guard cache directories against backup and sync exposure by @Naoray in #43
• fix: skip empty IRPAPERS transcriptions with provenance by @Naoray in #45
• test: cover pinned HF downloads and prove indexing lease contention by @Naoray in #41
• security: erase deleted cache content and WAL remnants by @Naoray in #42
• test: gate MCP deadlines on CLI startup readiness by @Naoray in #46
• fix: safely publish model cache files under concurrency by @Naoray in #48
• fix: preserve valid PDF pages after CMap panics by @Naoray in #47
• docs: describe keyword and hybrid search in README by @Naoray in #52
• fix: guard filesystem content reads against placeholders by @Naoray in #49
• feat: add a confined local-content mode for automatic collections by @Naoray in #50
• Avoid recursive discovery for cached local-content queries by @Naoray in #51
• feat: expose per-hit match ranks and query terms by @Naoray in #53
• fix: deny secret files across filesystem search entry points by @Naoray in #54
• fix(embeddinggemma): serialise ORT session init on Apple (dirname race) by @Naoray in #56
• Document fresh retrieval benchmarks and sourced comparisons by @Naoray in #58
• feat: add result kind counts and per-root indexing progress by @Naoray in #57
• fix: treat Gemma special-token strings as plain text by @Naoray in #59
• feat: serve lexical results while semantic search is unavailable by @Naoray in #55
• fix(core): pin discovery walks and content reads to the walked root by @Naoray in #60
• refactor(core): pair logical and resolved policy roots by @Naoray in #61
• fix(embeddinggemma): clarify text and template token diagnostics by @Naoray in #62
• refactor(core): harden PolicyRoot validation boundaries by @Naoray in #63
• docs: plain-English pass over README and docs (open-source prep) by @Naoray in #65
• feat(embed): add pinned Photo Anchor Q4 codec by @Naoray in #64
• Measure Photo Anchor P2 quality and record NO-GO result by @Naoray in #67
• Benchmark offline QMD, DigUp and LEANN retrieval by @Naoray in #68
• Improve hybrid retrieval with RRF k=5 by @Naoray in #69
• test: cover audio match reasons and stabilize exact-score goldens by @Naoray in #70
• Bind audio sessions to the bytes used for cache identity by @Naoray in #71
• Add measured competitor bars and polish the README benchmark chart by @Naoray in #72
• fix(core): pin opened directories across Linux automounts by @Naoray in #74
• chore: prepare repository for open source by @Naoray in #66
• Validate PolicyRoot capture and normalize MCP confined paths by @Naoray in #73
• fix: account for PDF benchmark omissions and parser failures by @Naoray in #76
• test: restore Windows discovery and mtime cache tests by @Naoray in #75
• Add verified cache import for re-staged paths by @Naoray in #77
• feat: opt-in Photo Anchor q4 image scan (P3) by @Naoray in #78
• Expose live JSONL progress for standalone indexing by @Naoray in #79
• Identify non-release builds in CLI version output by @Naoray in #80
• fix: isolate daemons by build and cache schema version by @Naoray in #81
• Add model profile inspection without loading weights by @Naoray in #82
• Fix daemon socket paths under long HOME and cache directories by @Naoray in #83
• Turn the README into a short decision page by @Naoray in #84
• Add opt-in typed extract error codes by @Naoray in #85
• chore: bundle release licences and prepare public source notices by @Naoray in #86
• Point installs at the source build until binaries return by @Naoray in #87
• Include audio in CLI builds and add seyk setup audio by @Naoray in #88
• fix: make Windows CI and retrieval harness portable by @Naoray in #89
• docs: use a neutral garden note in the README demo by @Naoray in #90
• docs: one Keep a Changelog file at the repository root by @Naoray in #92
• fix: keep semantic text search when vision files are missing by @Naoray in #91
• Install public binaries without gh and verify release archives by @Naoray in #93
• Prepare Photo Anchor calibration data and offline quality comparisons by @Naoray in #96
• test: make MCP background indexing deterministic by @Naoray in #94
• Prepare all workspace crates for crates.io by @Naoray in #97
• Attest release artifacts and pin GitHub Actions by @Naoray in #98
• docs: fix agent MCP setup and add Codex guidance by @Naoray in #104
• docs: remove Gaze from the north star by @Naoray in #105
• docs: sharpen north star scope and success signals by @Naoray in #107
• docs: position Seyk as local retrieval for AI agents by @Naoray in #106
• test: prevent PDF ownership checks racing parallel subprocesses by @Naoray in #108
• Fix MCP progress double-counting for shared watchers by @Naoray in #109
• fix: retain daemon endpoint ownership through runtime shutdown by @Naoray in #110
• Simplify PDF diagnostics ownership and progress completion by @Naoray in #111
• ci: give crate preparation one package list by @Naoray in #112
• test: validate typed audio golden references by @Naoray in #115
• fix(fixtures): separate mel provenance from decoder capture by @Naoray in #116
• Add Embedding::from_preserved for exact vector restoration by @Naoray in #117
• refactor(stt): join graphs to assets by file name by @Naoray in #114
• ci: run the full Rust check sequence in the stable canary by @Naoray in #113
• docs: clarify filesystem freshness hints and stale ranking by @Naoray in #118
• chore(deps): bump astral-sh/setup-uv from 10.1.0 to 10.3.2 by @dependabot[bot] in #103
• chore(deps): bump actions/setup-python from 5.6.0 to 7.1.0 by @dependabot[bot] in #100
• chore(deps): bump actions/checkout from 4.4.0 to 7.0.1 by @dependabot[bot] in #99
• chore(deps): bump actions/upload-artifact from 4.6.2 to 7.0.2 by @dependabot[bot] in #102
• chore(deps): bump actions/download-artifact from 4.3.0 to 8.0.2 by @dependabot[bot] in #101
• Bound embedding batches by padded work budgets by @Naoray in #95
• Declare MCP tool annotations and document unattended Codex approvals by @Naoray in #119
• Make MCP search results compact by default (protocol 11) by @Naoray in #120
• Fix MCP relative paths and accept a single query alias by @Naoray in #121
• Publish the agent evaluation (grep only vs grep + Seyk) by @Naoray in #123
• docs(north-star): record the paths-and-headings cache exception by @Naoray in #122
• Preserve MCP warning counts and freeze reference edge cases by @Naoray in #124
• Clarify when agents should use search or grep by @Naoray in #125
• Release v0.2.0 (version bump and changelog) by @Naoray in #127

New Contributors

• @dependabot[bot] made their first contribution in #103

Full Changelog: v0.1.0...v0.2.0