ROADMAP

What ships when. The v0.1 scope is locked. Everything else is "interesting later."

v0.1 — first public release (target: BSides Melbourne, May 16–17, 2026)

Goal: one person can install AllArkive on a laptop in an afternoon, ask the local AI a question about the local archive, and get an answer with citations they can verify.

In scope

docker-compose install of Kiwix + Ollama + Open WebUI on one machine.
Default knowledge bundle (balanced) with English Wikipedia text-only, a medical reference wiki, iFixit, Project Gutenberg, Stack Exchange.
Two additional bundles: minimal (Pi-friendly) and comprehensive (full Wikipedia with images for users with disk).
RAG pipeline with citations. No-source = no-answer behaviour.
Local landing page (search, chat, manage).
Install guides for laptop, Linux server, macOS, Windows (via WSL2).
Deployment patterns for Pi text-only, Pi archive-only, and laptop+Pi split.
Threat model and explicit disclaimers.
AGPL-3.0 licensed glue code; bundle content tracked with its own licenses.
GitHub primary, Codeberg mirror.
Signed releases, signed git tags.
Demo GIF in README.

Out of scope (explicitly deferred)

Clustering, replication, federation between nodes.
Mesh networking, serverless transport, IPFS-style content addressing.
Phone apps (Android, iOS).
Hardened or airgap-only build target.
Specialised bundles for medical, agricultural, or legal use.
Bundle deltas (incremental archive updates).
Multi-language UI.
Multi-user accounts / roles / quotas.
Cloud sync of any kind.
Telemetry of any kind.
Public-internet-facing default deployment.

v0.2 — quality of life

Goal: the install is smoother, the AI is better, the docs cover the long tail.

Shipped (post-BSides 2026-05-18)

Improved RAG: better embeddings, hybrid search (vector + keyword). Indexer rewritten with batched async embeddings (10–30× faster on CPU), int8 vector quantization (4× smaller vectors), offset-only chunk storage (~60% smaller index), and an opt-in Xapian/BM25 hybrid mode that skips dense indexing on multi-100-GB ZIMs entirely. Bundled under RAG_PROFILE=pi|laptop|workstation presets. See docs/rag-optimization.md and the [Unreleased] CHANGELOG entry.
Custom bundles. scripts/fetch-bundle.sh custom --add <url|handle> lets users compose their own bundle from any ZIM on download.kiwix.org or a direct URL. Manifest is generated incrementally and gitignored. See docs/bundles/README.md.

Candidates (not commitments)

Bundle deltas — update an archive without re-downloading 50 GB.
Pre-built index distribution — ship a vector index alongside the ZIM bundle so a Pi can skip indexing entirely. Feasibility validated in the v0.2 storage work; not yet wired into the release pipeline.
Multi-language retrieval.
A first-run wizard on the landing page (pick a bundle, pick a model).
Pre-flight check script: "does your machine actually have what this needs."
Backup and restore scripts for the archive and the user's chat history.
Documented opt-in LAN access with reverse proxy + auth examples.

v0.3+ — ideas

In rough order of how much they interest us, not how likely they are:

Specialised bundles: a curated medical bundle (with extra disclaimers), an agricultural bundle (FAO docs, soil/crop guides), a software-development bundle (curated Stack Exchange + selected open-source docs).
A second landing-page experience for non-technical users — bigger fonts, fewer options, more guidance.
Mesh / LAN federation — multiple AllArkive nodes on the same LAN exchange archive availability so a query can hit any of them.
Phone-as-a-client app — connect to a home AllArkive node from a phone on the same network. Not a full local stack on the phone.
Hardened build — minimal base images, SELinux profiles, signed kernel modules, etc. For users who want this as part of a security posture.
Reproducible builds end-to-end, including the model weights provenance chain.
A formal security audit once the project has any users worth auditing for.

What we will probably never do

Become a hosted SaaS. The whole point is that it runs on your machine.
Add tracking, telemetry, or "anonymous usage statistics."
Take VC money for the core project.
Promise this is a survival tool, a medical tool, or a legal tool. It isn't.
Ship anything we can't reproduce from source.
Bundle proprietary content.

How to propose something for the roadmap

Read this doc and check it isn't already here under a different name.
Open an issue with the roadmap label.
Describe the user need, not the technical solution.
Note which version it might fit (v0.2, v0.3, "someday").
We'll discuss in the issue and either move it into a milestone or close with a reason.

Scope creep is the most likely way this project dies. We'd rather say no twenty times than ship a sprawl.