Vodou v0.6.23 Alpha

Documents become memory. Vodou could always remember your conversations; it could not read the contract, the plan, or the spec those conversations were about. Now it can — and it can tell you which document answers a question, quote the passage, and hand it to any model on request.

The Document Library

• Ingest what you actually keep context in — pdf, docx, pptx, xlsx, csv, epub, md, txt — from a file, a whole folder, a URL, or the page open in your browser.
• Browse it at /library — search, filter by state (broken / un-carded / watched), read the extracted text, open the original, and add documents with a paste-a-path box that shows live progress on a folder import.
• Attach a document to any chat with `@doc:<name>` — the Vodou console, Slack, Telegram, anywhere. A typo suggests the right document instead of failing silently; a document too large to attach whole says so and points at the section you want.
• Vodou knows which document answers a question. Each document gets a routing card — what it is, what it answers, and what it is not about — so "what is our liability cap?" reaches the contract rather than a guess. Documents now surface in ordinary memory search for the first time.
• The panel tells you when a page relates to something you saved, and stays quiet when it doesn't — deliberately, because a chip that lights up on everything gets ignored.
• "Add to Vodou Library" in the browser — right-click any page, or use the keyboard shortcut for pages that own their right-click menu (Google Docs, Notion). Requires Vodou Bridge 0.5.97.73+.

Also in this release

• Documents are chunked as documents. A bullet in a personal note is a standalone fact; a bullet in a plan is prose. Treating them alike split one plan into 181 fragments — it now produces 50 coherent passages with nothing lost.
• Document matching is fast — ~10s to under a second, by asking the already-running memory service instead of starting a new one per question.
• One reranker per install. The background service and the command line were quietly using different relevance models, so the same question could score differently depending on which answered it.
• Contradiction detection reads across documents — a governing-law clause is compared against other governing-law clauses, not the whole corpus.

Fixed

• Two silent text-loss bugs in document import, both found by reading a stored document rather than trusting a success message: a paragraph-splitting bug dropped 940 bytes of a 9,199-byte passage mid-word, and an oversized first paragraph was never split at all.
• Pasting a bare document reference no longer confuses the router into running unrelated tools.
• Browser capture of Notion-style pages no longer welds every block into one run-on line, nor eats words inside links and buttons.

Install

curl -fsSL https://raw.githubusercontent.com/VodouAI/OS/main/install-vodou.sh | bash

Engine: VodouAI/vodou-core v0.6.23 · full history in CHANGELOG.md

Known issues

• Windows is unsigned — SmartScreen will warn. Treat it as a preview.
• Model lists for six providers (groq, deepseek, xai, mistral, together, kimi) are ~3 weeks old in this build.

Alpha.