zeekay
de5e797fca
A version written into a manifest is written into every manifest that wants it, which is how eslint came to exist in a hundred and four forms across this estate and how a package came to resolve twice at run time. The catalog moves the number out of the manifest: a manifest names the package, the workspace names the version. Four are held below their latest because the newer release cannot be used, and the reason is recorded rather than the version alone: @noble/hashes 2 renamed every subpath the web3 stack imports, @splinetool/runtime 2 depends on a package that was never published, @dynamic-labs 5.4.1 the same, and react-resizable-panels 3 points its types condition at the bundle.
3.1 KiB
Hanzo Index
Overview
A search index in Rust: documents in, ranked results out. Full-text (BM25, typo tolerant), vector, and hybrid retrieval, with faceted filtering, geo search, and chat completions grounded in your own indexes.
One instance holds one tenant's indexes behind one master key, so a deployment is per tenant rather than one shared keyspace.
Upstream: forked from Meilisearch (Meili SAS, MIT + a commercial EE). LICENSE, LICENSE-MIT and LICENSE-EE keep their copyright and NOTICE records the fork — that is what a fork owes, and it stays.
Tech Stack
- Language: Rust (toolchain pinned by
rust-toolchain.toml)
Build & Run
cargo build --release -p search -p searchtool
./target/release/search --master-key="…"
Binaries are named after their crates: the server is search, the CLI companion is
searchtool. The container's WORKDIR is /data, which is where a volume belongs.
Naming
Four different things, deliberately not merged:
| thing | what it is |
|---|---|
this repo (hanzoai/index) |
the index engine — where documents go |
search.hanzo.ai |
Hanzo's own instance of it, serving the docs corpus |
cloud clients/websearch |
queries the outside world |
hanzoai/crawl |
fetches pages |
Environment
Every option is a const in crates/search/src/option.rs, all prefixed INDEX_
(INDEX_MASTER_KEY, INDEX_DB_PATH, INDEX_ENV, …). With INDEX_ENV=production and
no master key the process refuses to start, so a misconfigured name is a crash-loop and
never an unauthenticated index.
mini-dashboard is NOT a default feature — it vendors a prebuilt UI from an upstream
release. Without it / answers {"status":"Hanzo Index is running"} in every
environment.
Upgrading a live database
The database records the version that wrote it, and a newer binary will not open an
older one. INDEX_EXPERIMENTAL_DUMPLESS_UPGRADE=true migrates in place on boot
(supported from 1.12.0 up). Take a dump first — POST /dumps — and check its document
counts against /stats before rolling: that dump restores into the newer binary with
--import-dump, and that is the rollback.
CI
No GitHub Actions. Every push reaches the platform through the Hanzo GitHub App and the
platform builds on our own runner fabric; pipelines live in .hanzo/workflows.
Key Files
crates/search/-- HTTP server, routes, auth, search queuecrates/milli/-- the engine: indexing, ranking, filtering, vectorscrates/index-scheduler/-- async task scheduling, batching, upgrade stepsDockerfile-- container builddocs/index.mdx-- the reference
Version pin (kept here, not in customer docs)
Current workspace version 1.45.2, 23 workspace members. Naming as shipped:
server binary search, operator CLI searchtool, auth crate search-auth,
core index engine milli (unchanged), env var prefix INDEX_*.
docs.hanzo.ai/docs/skills/hanzo-search no longer pins the upstream release in its introduction — it had drifted (it said v1.37.0) and it is an engineering fact, not a customer-facing one. Update the version above when the base moves.