github yjid/kensaku v1.0.18
Kensaku 1.0.18 — Pre-release

pre-release43 minutes ago

Kensaku 1.0.18 — Pre-release

[!IMPORTANT]
This is a pre-release build for testing, not the final stable release.

Windows: Binaries are signed using SignPath's test certificate. Windows SmartScreen may show "Unknown publisher". You may need to select More info → Run anyway.

The production signing certificate is still pending. Windows binaries will be signed with the production certificate before the stable release.

macOS: Builds remain unsigned and are not notarized (#73).

Please report regressions or unexpected behaviour before the stable release.

Kensaku 1.0.18 is a larger update than usual. It introduces per-folder indexing modes, improves Markdown and Outlook .msg extraction, reworks OCR thread scheduling, and includes several indexing, search and performance fixes.

Existing settings and indexes are compatible. No re-index is required to upgrade.

Downloads

Platform | Installer | Portable / CLI -- | -- | -- Windows x64 | kensaku_1.0.18_x64-setup.exe, kensaku_1.0.18_x64_en-US.msi | kensaku-portable-1.0.18-x64.zip, kensaku-cli-windows-x86_64.exe Windows ARM64 | kensaku_1.0.18_arm64-setup.exe, kensaku_1.0.18_arm64_en-US.msi | kensaku-portable-1.0.18-arm64.zip, kensaku-cli-windows-aarch64.exe macOS (universal) | kensaku_1.0.18_universal.dmg | kensaku-cli-macos-universal, kensaku-cli-macos-aarch64 Linux x86_64 | kensaku_1.0.18_amd64.deb, kensaku-1.0.18-1.x86_64.rpm, kensaku_1.0.18_amd64.AppImage | kensaku-cli-linux-x86_64

SHA256SUMS covers every asset.

New

  • Per-folder "names only" indexing (#145, #140). Any folder can be indexed by name and path only: files below it stay searchable and appear in results, but they are never read, OCR'd or embedded. This keeps media libraries, archives and network shares cheap to index. The list is empty by default, so existing setups are unchanged; changing it reconciles the index on the next scan. Configure it with the Names only switch in Settings → Folders or in first-run setup, or through the name_only_dirs setting.

  • Cancel the first-run scan and change folders (#188, #164). The setup scan screen now offers Cancel and change folders, which stops the crawl, releases its worker slots and returns to folder selection. A cancelled crawl does not mark the first scan complete, so the wizard resumes properly.

  • Outlook messages whose body is stored only as HTML or RTF are indexed (#167, follow-up to #63). Some .msg files carry no plain-text body at all; those messages were previously indexed by subject and attachment names only. Extraction now falls back to the HTML body, then to the HTML embedded in the compressed RTF body.

  • Cleaner Markdown search text and previews ([#139], #44). .md and .markdown files keep the existing plain-text reader, but link and image destinations, image paths with empty alt text, and top-level code-fence markers are stripped from what is indexed, from result snippets and from previews. Link labels, alt text, frontmatter and code inside fences stay searchable. This is clean-up, not a Markdown renderer: previews still show the document as text. .markdown files also get the document category, the category filter and the language facet, including uppercase .MARKDOWN.

Fixed

  • Renaming a folder no longer removes its indexed files on Windows (#187). A rename arrives as a single event for the directory itself; the watcher now re-indexes the children it gained instead of treating the folder as new and empty.

  • Folders and files starting with a dot are indexed again when "Skip hidden files" is off (#179, #178). Both traversal filters hard-coded the hidden check, so turning the setting off had no effect on rebuilds. Configured exclusions, sensitive-name rules, ignore files and cloud-placeholder handling are unchanged.

  • Search no longer returns unrelated "letter-soup" results from the semantic tier (#177). Every token the static embedding vocabulary could not represent mapped to the same [UNK] row, so random strings, unsupported scripts and identifier-heavy documents all collapsed onto one vector. Measured on 461 real chunks, random letters previously scored 0.56 cosine against genuine content.

  • The headless Linux daemon no longer stalls at 0/N during extraction (#180). The extraction pipeline could deadlock when the CPU quota was one; it also now tries every candidate control-socket path before giving up.

  • Custom CA bundles and the system trust store are honoured (#171, #169). Downloads previously used only bundled roots, so HTTPS inspection proxies broke model, OCR and update downloads with UnknownIssuer. SSL_CERT_FILE and SSL_CERT_DIR are now respected as well.

  • No false "crash log" warning after a WebView2 restart (#176). When the previous instance's msedgewebview2.exe children still held the user-data folder, window creation failed with ERROR_BUSY and was recorded as a crash. Creation is retried three times (300 ms, 900 ms, 2 s), and if it still fails the message names the real cause and the recovery step.

  • Reopening the index no longer fails when Tantivy's writer lock is briefly busy (#191). An index rebuild drops the old writer and reopens the directory immediately; on Windows the lock can stay held for a few milliseconds, which discarded the rebuild. The lock is now retried for up to two seconds.

  • A stale or truncated Indra OCR pack is detected, and the self-test result is no longer cached forever (#174). Pack presence was checked with is_file(), so an interrupted download counted as installed and the engine then read nothing. A pack is now only "present" at its expected size, the self-test cache is cleared after a successful download, and Indra reports how many lines the detector found versus how many the recogniser read — so a failure identifies which stage broke.

Performance

  • The first scan no longer oversubscribes every core (#172, #170). Each detector and recogniser session opened a core-sized ONNX thread pool, so extraction concurrency multiplied. Threads are now pinned to one, recogniser sessions are capped at eight, regions are processed serially while documents are parallelised, and two CPUs are reserved where available. Background priority is applied to worker threads only, instead of to the whole process on Windows and macOS.

Upgrade notes

  • Configuration and the existing index are reused as-is. Nothing needs to be re-indexed to upgrade.

  • New extraction only applies to newly indexed or modified files. Files already in the index keep the text extracted by 1.0.17. To apply the Markdown and .msg improvements to an existing corpus, run Save & re-index from Settings, or the /reindex command.

  • "Names only" drops content for that folder by design. Switching a folder to names-only removes its content documents and semantic units; switching it back queues its files for extraction again.

  • A cancelled first scan is not remembered as complete, so the setup wizard will offer to scan again rather than pretending the corpus is indexed.

Under the hood

  • Dependency majors: mail-parser 0.11.9 → 1.0.1, cfb 0.14 → 0.16, parking_lot 0.12 → 0.13. Also notify → 8.2.2 and interprocess → 2.4.8.

  • Windows release binaries are Authenticode-signed through SignPath Foundation for the first time. The app executable is signed first, the installers are rebuilt from that signed binary (--no-binary-patching), then signed themselves, so the setup, MSI, portable executable and CLI all carry a signature. This pre-release uses the test certificate.

  • Windows test binaries now link the Common Controls v6 manifest (#192), so the test harness matches the shipped executable. This affects CI only.

  • The real-model OCR workflow now also runs on master pushes, not only on pull requests.

  • The release workflow fails a tagged build if any Windows binary is unsigned, or if a release-signing build does not verify as Valid.

Verifying your download

Every platform:

sha256sum -c SHA256SUMS

Windows:

Get-AuthenticodeSignature .\kensaku_1.0.18_x64-setup.exe |
  Format-List Status, SignerCertificate, TimeStamperCertificate

The test certificate may report UnknownError because its root is not trusted. This does not necessarily indicate a missing signature.

macOS: The DMG is unsigned and not notarized. After copying Kensaku to Applications:

xattr -dr com.apple.quarantine /Applications/Kensaku.app

Or use System Settings → Privacy & Security → Open Anyway.

The DMG is a universal binary, so Apple silicon runs natively.

Known issues

Not addressed in this release:

  • #73 — macOS repeats the disk-access request after indexing completes.

  • #163 — jumping from a search result to the matching line.

  • #131 — Android version (not planned).

  • #67 — Flow Launcher integration.

  • #138 — Liquid Glass and Frost themes.

  • #71 — evaluation of the OXR document model.

Reporting problems

Please report any regressions or unexpected behaviour.

Include:

  • Your OS and architecture.

  • Kensaku version (1.0.18).

  • Installation method (installer, portable folder, AppImage, etc.).

  • Whether the same behaviour occurs on 1.0.17.

For OCR problems, attach the [Kensaku OCR] lines from a terminal launch; they report detected versus recognised line counts.


What's Changed

New Contributors

Full Changelog: v1.0.17...v1.0.18

# Kensaku 1.0.18 — Pre-release

Important

This is a pre-release build for testing, not the final stable release.

Windows: Binaries are signed using SignPath's test certificate. Windows SmartScreen may show "Unknown publisher". You may need to select More info → Run anyway.

The production signing certificate is still pending. Windows binaries will be signed with the production certificate before the stable release.

macOS: Builds remain unsigned and are not notarized ([#73]).

Please report regressions or unexpected behaviour before the stable release.

Kensaku 1.0.18 is a larger update than usual. It introduces per-folder indexing modes, improves Markdown and Outlook .msg extraction, reworks OCR thread scheduling, and includes several indexing, search and performance fixes.

Existing settings and indexes are compatible. No re-index is required to upgrade.

Downloads

Platform Installer Portable / CLI
Windows x64 kensaku_1.0.18_x64-setup.exe, kensaku_1.0.18_x64_en-US.msi kensaku-portable-1.0.18-x64.zip, kensaku-cli-windows-x86_64.exe
Windows ARM64 kensaku_1.0.18_arm64-setup.exe, kensaku_1.0.18_arm64_en-US.msi kensaku-portable-1.0.18-arm64.zip, kensaku-cli-windows-aarch64.exe
macOS (universal) kensaku_1.0.18_universal.dmg kensaku-cli-macos-universal, kensaku-cli-macos-aarch64
Linux x86_64 kensaku_1.0.18_amd64.deb, kensaku-1.0.18-1.x86_64.rpm, kensaku_1.0.18_amd64.AppImage kensaku-cli-linux-x86_64

SHA256SUMS covers every asset.

New

  • Per-folder "names only" indexing ([#145], [#140]). Any folder can be indexed by name and path only: files below it stay searchable and appear in results, but they are never read, OCR'd or embedded. This keeps media libraries, archives and network shares cheap to index. The list is empty by default, so existing setups are unchanged; changing it reconciles the index on the next scan. Configure it with the Names only switch in Settings → Folders or in first-run setup, or through the name_only_dirs setting.
  • Cancel the first-run scan and change folders ([#188], [#164]). The setup scan screen now offers Cancel and change folders, which stops the crawl, releases its worker slots and returns to folder selection. A cancelled crawl does not mark the first scan complete, so the wizard resumes properly.
  • Outlook messages whose body is stored only as HTML or RTF are indexed ([#167], follow-up to [#63]). Some .msg files carry no plain-text body at all; those messages were previously indexed by subject and attachment names only. Extraction now falls back to the HTML body, then to the HTML embedded in the compressed RTF body.
  • Cleaner Markdown search text and previews ([#139], [#44]). .md and .markdown files keep the existing plain-text reader, but link and image destinations, image paths with empty alt text, and top-level code-fence markers are stripped from what is indexed, from result snippets and from previews. Link labels, alt text, frontmatter and code inside fences stay searchable. This is clean-up, not a Markdown renderer: previews still show the document as text. .markdown files also get the document category, the category filter and the language facet, including uppercase .MARKDOWN.

Fixed

  • Renaming a folder no longer removes its indexed files on Windows ([#187]). A rename arrives as a single event for the directory itself; the watcher now re-indexes the children it gained instead of treating the folder as new and empty.
  • Folders and files starting with a dot are indexed again when "Skip hidden files" is off ([#179], [#178]). Both traversal filters hard-coded the hidden check, so turning the setting off had no effect on rebuilds. Configured exclusions, sensitive-name rules, ignore files and cloud-placeholder handling are unchanged.
  • Search no longer returns unrelated "letter-soup" results from the semantic tier ([#177]). Every token the static embedding vocabulary could not represent mapped to the same [UNK] row, so random strings, unsupported scripts and identifier-heavy documents all collapsed onto one vector. Measured on 461 real chunks, random letters previously scored 0.56 cosine against genuine content.
  • The headless Linux daemon no longer stalls at 0/N during extraction ([#180]). The extraction pipeline could deadlock when the CPU quota was one; it also now tries every candidate control-socket path before giving up.
  • Custom CA bundles and the system trust store are honoured ([#171], [#169]). Downloads previously used only bundled roots, so HTTPS inspection proxies broke model, OCR and update downloads with UnknownIssuer. SSL_CERT_FILE and SSL_CERT_DIR are now respected as well.
  • No false "crash log" warning after a WebView2 restart ([#176]). When the previous instance's msedgewebview2.exe children still held the user-data folder, window creation failed with ERROR_BUSY and was recorded as a crash. Creation is retried three times (300 ms, 900 ms, 2 s), and if it still fails the message names the real cause and the recovery step.
  • Reopening the index no longer fails when Tantivy's writer lock is briefly busy ([#191]). An index rebuild drops the old writer and reopens the directory immediately; on Windows the lock can stay held for a few milliseconds, which discarded the rebuild. The lock is now retried for up to two seconds.
  • A stale or truncated Indra OCR pack is detected, and the self-test result is no longer cached forever ([#174]). Pack presence was checked with is_file(), so an interrupted download counted as installed and the engine then read nothing. A pack is now only "present" at its expected size, the self-test cache is cleared after a successful download, and Indra reports how many lines the detector found versus how many the recogniser read — so a failure identifies which stage broke.

Performance

  • The first scan no longer oversubscribes every core ([#172], [#170]). Each detector and recogniser session opened a core-sized ONNX thread pool, so extraction concurrency multiplied. Threads are now pinned to one, recogniser sessions are capped at eight, regions are processed serially while documents are parallelised, and two CPUs are reserved where available. Background priority is applied to worker threads only, instead of to the whole process on Windows and macOS.

Upgrade notes

  • Configuration and the existing index are reused as-is. Nothing needs to be re-indexed to upgrade.
  • New extraction only applies to newly indexed or modified files. Files already in the index keep the text extracted by 1.0.17. To apply the Markdown and .msg improvements to an existing corpus, run Save & re-index from Settings, or the /reindex command.
  • "Names only" drops content for that folder by design. Switching a folder to names-only removes its content documents and semantic units; switching it back queues its files for extraction again.
  • A cancelled first scan is not remembered as complete, so the setup wizard will offer to scan again rather than pretending the corpus is indexed.

Under the hood

  • Dependency majors: mail-parser 0.11.9 → 1.0.1, cfb 0.14 → 0.16, parking_lot 0.12 → 0.13. Also notify → 8.2.2 and interprocess → 2.4.8.
  • Windows release binaries are Authenticode-signed through [SignPath Foundation] for the first time. The app executable is signed first, the installers are rebuilt from that signed binary (--no-binary-patching), then signed themselves, so the setup, MSI, portable executable and CLI all carry a signature. This pre-release uses the test certificate.
  • Windows test binaries now link the Common Controls v6 manifest ([#192]), so the test harness matches the shipped executable. This affects CI only.
  • The real-model OCR workflow now also runs on master pushes, not only on pull requests.
  • The release workflow fails a tagged build if any Windows binary is unsigned, or if a release-signing build does not verify as Valid.

Verifying your download

Every platform:

sha256sum -c SHA256SUMS

Windows:

Get-AuthenticodeSignature .\kensaku_1.0.18_x64-setup.exe |
  Format-List Status, SignerCertificate, TimeStamperCertificate

The test certificate may report UnknownError because its root is not trusted. This does not necessarily indicate a missing signature.

macOS: The DMG is unsigned and not notarized. After copying Kensaku to Applications:

xattr -dr com.apple.quarantine /Applications/Kensaku.app

Or use System Settings → Privacy & Security → Open Anyway.

The DMG is a universal binary, so Apple silicon runs natively.

Known issues

Not addressed in this release:

  • [#73] — macOS repeats the disk-access request after indexing completes.
  • [#163] — jumping from a search result to the matching line.
  • [#131] — Android version (not planned).
  • [#67] — Flow Launcher integration.
  • [#138] — Liquid Glass and Frost themes.
  • [#71] — evaluation of the OXR document model.

Reporting problems

Please report any regressions or unexpected behaviour.

Include:

  • Your OS and architecture.
  • Kensaku version (1.0.18).
  • Installation method (installer, portable folder, AppImage, etc.).
  • Whether the same behaviour occurs on 1.0.17.

For OCR problems, attach the [Kensaku OCR] lines from a terminal launch; they report detected versus recognised line counts.


What's Changed

  • fix(indexer): honor Skip Hidden Files during directory scans by @codewithpom in #179
  • fix(indexer): index .msg bodies stored only as HTML or RTF (follow-up to #63) by @mtjia6 in #167
  • fix(tls): respect system and custom CA bundles by @codewithpom in #171
  • Clean up Markdown search text and previews by @surkhayka in #139
  • fix(ocr): bound inference threads and preserve UI priority by @codewithpom in #172
  • Per-directory indexing mode: names only or names and content by @rhurta in #145
  • fix(build): give Windows test binaries the Common Controls v6 manifest by @yjid in #192
  • fix(indexer): wait out a busy Tantivy writer lock on reopen by @yjid in #191
  • chore(deps): bump mail-parser from 0.11.9 to 1.0.1 by @yjid in #190
  • chore(deps): bump interprocess from 2.4.7 to 2.4.8 in /cli by @dependabot[bot] in #181
  • chore(deps): bump interprocess from 2.4.7 to 2.4.8 in /src-tauri by @dependabot[bot] in #183
  • chore(deps): bump cfb from 0.14.0 to 0.16.0 in /src-tauri by @yjid in #186
  • chore(deps): bump notify from 8.2.1 to 8.2.2 in /src-tauri by @yjid in #184
  • chore(deps): bump parking_lot from 0.12.5 to 0.13.0 by @yjid in #182

New Contributors

Full Changelog: v1.0.17...v1.0.18

Don't miss a new kensaku release

NewReleases is sending notifications on new releases.