cade v0.2.1

cade runs beyond Linux: it builds from source on FreeBSD (CPU, or the GPU through Vulkan), and this release adds experimental archives for macOS on Apple Silicon and Windows x86-64. Nobody on the project has a Mac or a Windows machine: those two are built and tested only on GitHub's runners, so please report how they behave on a real one.

Português: o cade roda além do Linux: compila do código-fonte no FreeBSD (CPU, ou a GPU pelo Vulkan), e esta versão traz arquivos experimentais para macOS em Apple Silicon e Windows x86-64. Ninguém no projeto tem Mac nem Windows: esses dois são compilados e testados só nas máquinas do GitHub, então conte como eles se comportam numa de verdade. As notas abaixo estão também no CHANGELOG.pt-BR.md.

Install

• Linux x86-64 (CPU, any CPU with AVX2): cade-v0.2.1-linux-amd64-cpu.tar.gz below, checked with its .sha256, then the README. The models are the same as v0.2.0's.
• macOS, Apple Silicon (experimental): cade-v0.2.1-darwin-arm64-metal.tar.gz. It is not signed: run xattr -d com.apple.quarantine cade once. The models run on the CPU by default; "gpu_layers": -1 tries Metal, which produced garbage on GitHub's virtual Macs. The config is ~/Library/Application Support/cade/config.json.
• Windows x86-64 (experimental, CPU): cade-v0.2.1-windows-amd64-cpu.zip, with cade.exe linked statically. It is not signed, so SmartScreen may warn. The config is %APPDATA%\cade\config.json, the database and models live in %LOCALAPPDATA%\cade, and git must be on the PATH for the git source.
• FreeBSD and NVIDIA GPUs: from source, as DEVELOPMENT explains (go tool mage cpu, go tool mage vulkan or go tool mage build).

Before upgrading from v0.2.0

• On Linux nothing changes: no migration, no reindex, same models, config and ask --json.
• macOS and Windows are new installs: run cade init, then cade doctor.

What is in it

• FreeBSD (#101): builds from source, CPU or Vulkan; tested on FreeBSD 15.1, where the retrieval suite matches Linux on the CPU.
• macOS (#53): Metal build, browsers under Application Support, background ingestion with throttled disk I/O.
• Windows (#54): paths stored with / like on Linux, data in AppData, browsers and the new Teams desktop app found by cade init, background ingestion with pause, stop and --gentle without signals, and the progress bar in the console.
• Everywhere: go tool mage vulkan for GPUs without CUDA, and two doctor/loading messages that were wrong.

See the CHANGELOG for every change and measurement, and PRIVACY for where the files live on each system.

cade v0.2.0

Third version of cade: it now understands who and what appears in your history. People, files and images become entities, with the relationships between them, and the search uses them ("imagens do João que o Pedro me mandou", "quem mais mexeu no arquivo que o Rui alterou?"). Other apps' messages are read from the browser's storages through schemas, the database drops to less than half its size (on the reference machine, 1.8 GB → 0.77 GB with int8 vectors and one set of vectors per text), and you can export everything, or forget everything about one person.

Português: terceira versão do cade: ele passa a entender quem e o que aparece no seu histórico. Pessoas, arquivos e imagens viram entidades, com as relações entre elas, e a busca usa isso ("imagens do João que o Pedro me mandou", "quem mais mexeu no arquivo que o Rui alterou?"). Mensagens de outros apps são lidas dos armazenamentos do navegador por schemas, o banco cai para menos da metade do tamanho (na máquina de referência, 1,8 GB → 0,77 GB, com vetores int8 e um conjunto de vetores por texto), e dá para exportar tudo ou esquecer tudo sobre uma pessoa. As notas abaixo estão também no CHANGELOG.pt-BR.md.

Install

• Binary (Linux x86-64, CPU, any CPU with AVX2): download cade-v0.2.0-linux-amd64-cpu.tar.gz below, check it with the .sha256 file, and follow the README. The models are downloaded separately; they are the same as v0.1.0's.
• From source (and for NVIDIA GPUs): the README's steps. go tool mage build compiles the NVIDIA binary and needs the CUDA Toolkit; without it, use go tool mage cpu.
• Then cade init, cade doctor, cade ingest all, cade ask.

Before upgrading from v0.1.0

• The first command copies the database, then migrates it. Migrations 11 to 19 run on the first command that opens a v0.1.0 database. Before the first one that rewrites data, cade writes cade.db.before-v11-<date> (mode 600) next to the database: it needs as much free disk as the database, plus the vectors while they are rewritten, and on a 270 k-event database the migrations add up to about two minutes. cade doctor tells, without changing anything, what is pending and whether it will copy.
• The copy keeps everything, including what you forget later. Once the new version works, delete it with cade forget --backups; cade forget and cade doctor now list such copies.
• No reindex, no new embedding, no new model.
• ask --json: the plan may carry modality, entities and related; they are left out when empty, so other questions give the same JSON.
• Do not re-import instead: the Teams cache expires and Chrome keeps ~90 days of history, so forget + ingest loses data.

New, and off until you turn it on

• ingest.names_from_email: keeps on each commit the name its author's e-mail spells ("F. Rocha <fernanda.rocha@…>" is found as Fernanda Rocha). It is a new kind of personal data; turning it off deletes the names on the next cade ingest git.
• Teams pictures (with sources.images on): the pictures in Teams messages are read from the browser's HTTP cache and described like image files; only the description is kept.
• Other apps are read only from the origins and URLs you configure (sources.storage_origins, sources.request_cache_urls) and the schemas you save.

See PRIVACY for what each source keeps, and the new comparison with the LGPD and the GDPR.

What is in it

See the CHANGELOG for every change, measurement and migration.

cade v0.1.0

Second version of cade: secrets kept out of the database, deletion of single events, a generation model under Apache-2.0 (Qwen3.5-2B), search by image description and ingestion in the background.

Português: segunda versão do cade: segredos fora do banco, exclusão de eventos isolados, um modelo de geração sob Apache-2.0 (Qwen3.5-2B), busca pela descrição de imagens e ingestão em segundo plano. As notas abaixo estão também no CHANGELOG.pt-BR.md.

Install

• Binary (Linux x86-64, CPU, any CPU with AVX2): download cade-v0.1.0-linux-amd64-cpu.tar.gz below, check it with the .sha256 file, and follow the README. The models are downloaded separately.
• From source (and for NVIDIA GPUs): the README's steps, now with go tool mage instead of make (#16). go tool mage build compiles the NVIDIA binary and needs the CUDA Toolkit; without it, use go tool mage cpu.
• Then cade init, cade doctor, cade ingest all, cade ask.

Before upgrading from v0.0.0

• Migration 8 copies the database first. It masks known secrets in stored events, deletes credential files (.env*, *.pem, id_rsa*…) and compacts the database. Before it runs, cade writes cade.db.before-v8-<date> (mode 600) next to the database: it takes about a minute and as much free disk as the database. cade doctor tells, without changing anything, whether a migration is pending and whether it will copy. Migrations 9 and 10 only add a table and an index.
• Run cade reindex once after migration 8. Events whose text had a secret masked have no vectors until it finishes.
• ask --json changed: the task status "concluida" is now "pr_aberto". It means local history saw the PR creation page, not that the PR was merged. Scripts that read the old code must change.
• New generation model: run go tool mage models again. The default is now Qwen3.5-2B, with its vision projector (mmproj-Qwen3.5-2B-F16.gguf, 0.67 GB). Configurations written by cade init with the old default switch to it by themselves, so qwen2.5-3b-instruct-q4_k_m.gguf can be deleted. An explicit generation.model_path keeps the model it names.
• Images are described only with sources.images on. It is off by default; cade init asks and shows the cost (about 1.7 s per image on an RTX 3060, 22 s on a 6-core CPU).
• Do not re-import instead: the Teams cache expires and Chrome keeps ~90 days of history, so forget + ingest loses data.

Models and licenses

• cade is now under the GPLv3 or later; v0.0.0 stays under the GPLv2.
• The default generation model, Qwen3.5-2B, and its mmproj are under Apache-2.0. The non-commercial restriction of v0.0.0's Qwen2.5-3B no longer applies to the defaults.
• The binary's own dependencies are listed with their licenses in THIRD_PARTY_NOTICES.md, included in the archive.

What is in it

See the CHANGELOG for every change, measurement and migration.

cade v0.0.0

First version of cade: a local, offline history of your git commits, browser history, files and Teams messages, queried by date or in natural language.

Português: primeira versão do cade: histórico local e offline dos seus commits, histórico do navegador, arquivos e mensagens do Teams, consultado por data ou em linguagem natural. As notas abaixo estão também no CHANGELOG.pt-BR.md.

Install

• Binary (Linux x86-64, CPU, any CPU with AVX2): download cade-v0.0.0-linux-amd64-cpu.tar.gz below, check it with the .sha256 file, and follow "Or from a release binary" in the README. The models are downloaded separately.
• From source (and for NVIDIA GPUs, make cuda): the README's steps from a clean clone to the first cade ask.
• Then cade init, cade doctor, cade ingest all, cade ask.

Before upgrading a database from a development build

• Migrations run by themselves the first time a command opens the database. Before the first one that rewrites data, cade writes a copy next to it, cade.db.before-vN-<date> (mode 600): it takes about a minute and as much free disk as the database (516 MB for a real history of 108 thousand events). cade doctor tells, without changing anything, whether a migration is pending and whether it will copy.
• Run cade reindex once after migration 5 (vectors per chunk). Until it finishes, long texts have no vectors and ask warns about it; on a real history it took ~80 s on an RTX 3060.
• Do not re-import instead: the Teams cache expires and Chrome keeps ~90 days of history, so forget + ingest loses data.

Models and licenses

• The default generation model, Qwen2.5-3B-Instruct, is under the Qwen Research License: non-commercial ("research or evaluation purposes only"). For commercial use, get a license from Alibaba Cloud or switch to Qwen2.5-1.5B-Instruct (Apache-2.0), which reads questions worse (see the README's Models section).
• The binary's own dependencies (llama.cpp, SQLite, go-sqlite3, sqlite-vec, klauspost/compress, Go) are listed with their licenses in THIRD_PARTY_NOTICES.md, included in the archive.

What is in it

See the CHANGELOG for every change, measurement and migration.