Knowledge on your own machines. Every answer cited.
Metis turns the documents on your drives — and the files, notes and facts you add — into knowledge bases you can ask questions of. Each question is matched by meaning and by its exact words, reranked, and answered with the passages that answer it best, each with where it comes from: the document, its page or headings, the line. The documents and the index stay on your machines.
A knowledge base is sources, an index, and one search.
It runs as ordinary work on one of your workspace's machines — a CPU core and a little memory by default — with its index on a drive of its own, so it survives restarts. Its models are Eos deployments of yours.
Serve two small models
An embedding model and a reranker from the Eos catalog — BGE-M3 and BGE Reranker v2 M3 are small enough to run on CPUs, next to the documents.
Point it at your drives
Any drive, or a folder of one: a folder on a machine, a shared filesystem, a bucket or a Hugging Face repository as a data source. Drives are read, never written.
Ask it
Searchable while it indexes. When files change, only the changed ones are read again; a data source's new revision is indexed by itself.
Measure it
An evaluation set — questions, and the documents that answer them — gives recall@k and MRR. Change a model or the chunking, and run it again to see whether it helped.
Found by meaning, and by the exact words.
Meaning finds “how many holidays do I get” in a passage about vacation days. Words find the names, codes and numbers meaning blurs — Xid 79, nvidia-smi. Metis does both, then lets a reranker read the question with the best candidates.
- CitationsEvery result says where it comes from — the drive, the document, its title, the page of a PDF or the headings above it, the line — with a snippet of what matched.
- Context keptDocuments are cut along their sections, and each passage is read with its headings in front, so a passage under “Installation” is found by a question about installing.
- Why it came upEach result carries its score at every step — by meaning, by words, fused, reranked — so you can see which retriever found it.
- Still answersIf the embedding model does not answer, the search falls back to words; without the reranker, to the fused order — and says so.
- MeasuredAn evaluation set gives recall@k and MRR, each way — hybrid, meaning only, words only, with or without the reranker — so “it finds it” is a number you can compare.
# A knowledge base of a drive's folder
$ astra metis create handbook \
--drive docs:handbook \
--embedding bge-m3 --reranker bge-reranker
# Ask it: passages, each with its citation
$ astra metis search handbook \
"How many vacation days do I have?"
# Measure it: recall@k, MRR, what was missed
$ astra metis evaluate handbook \
--file docs:.metis/evals.jsonlYour drives, and whatever else you want it to know.
Besides the drives it reads, every base keeps notes of its own: files you upload, notes saved from conversations and a memory's facts — indexed and cited like any document. A base may be only that, with no drive at all.
Your drives
Up to 16 sources per base: drives or folders of them, of any kind. Markdown, text, HTML, PDF and Word files are read; a scanned PDF without text is listed with why.
Files from anywhere
Drop files or folders on the console, tap Upload on your phone, or astra metis add. Written on the base's own machine, searchable seconds later.
Save from a conversation
In the app's chat or the Playground, save a message — or a summary the model writes, which you read and correct — as a dated note with where it came from.
Keep what you attach
A PDF or document attached to a chat message can be kept in a base too, so later searches find it.
A model proposes, you save
Ask a model to save something and it proposes a note. You see the title and the full text, then Save or Discard. A draft nobody decides on is dropped after a day.
Evaluation sets
Keep the questions in a hidden folder of the drive, or upload them: read by evaluations, never indexed.
Why a model never saves by itself: a base is what later answers rely on. A note saved from a misunderstanding, or from instructions hidden in a document the model read, would be cited as truth — so saving takes a person every time, and what is saved is exactly what was shown.
Facts that know when they were true.
A memory keeps dated facts instead of documents: who or what, what about it, the fact, its source and when it was said. When a newer fact says otherwise, it replaces the older one, which stays in the history — so an assistant that asks gets what holds now, and “what was it in September?” still has an answer.
- Remember this conversationIn the app's chat or the Playground. A small model on a CPU reads it and proposes the facts, each with what it would replace; you untick, correct and keep.
- Yours to seeEvery fact by subject, with its date and source, and its history. Edit one and its history is kept; delete one, or everything kept about it.
- Only what you chooseA memory reads the conversations you send it, when you send them, and keeps only the facts you keep. It does not read your chats by itself.
- No secretsWhat looks like a password, a key or a card number is left out of what is proposed, and refused if kept by hand.
$ astra metis remember team --from standup.json
reading the conversation for facts…
new Ana — role: Ana works in research.
replaces: Ana leads the data team.
new Launch — date: The launch moved
to Friday 16 October.
Keep 2 fact(s) in team? [y/N] y
kept 2; 1 replaced an older factFrom Claude Code in one command. From everything else too.
The console (and the app on your phone), the astra command line, the API, Claude Code and any AI assistant over MCP — each as you, with your role in the workspace.
$ astra connect claude-code
installed the Astralyx knowledge skill:
~/.claude/skills/astralyx-knowledge/SKILL.md
# Then, in Claude Code:
# "What does our handbook say about leave?"
# → astra metis search handbook "leave" …
# → an answer citing handbook/leave.md- Claude CodeOne command installs a skill: Claude searches your bases with
astra metis search, cites the documents it used, says so when nothing answers, and saves a note only after your yes. - MCPClaude, ChatGPT, Cursor, VS Code and other assistants list and search your bases, and propose notes that wait for your approval in the console.
- APICreate, search, add files, keep facts and evaluate, over REST with a workspace token — also where astra cannot be installed.
- PermissionsViewers search; editors and admins create, change and add to bases. With custom roles, each action is a permission of its own.
In the app's chat and the Playground: save a message or a summary, keep an attached file, remember the conversation. Give a deployment the save_to_knowledge tool, and it proposes notes for you to save.
Give an agent the Metis API as a tool, with a workspace token kept as a credential: each search is decided by the agent's policy and recorded, like any other call.
A notebook, like any run of the workspace, searches a base at its own address on the workspace's network — the same body as the search API.
What stays on your machines.
The same rule as everything Astralyx builds: the control plane decides, and never holds what you would not want it to.
- Your documentsRead where they are, on your drives. Source drives are read-only: nothing a base does changes them.
- The indexOn a drive of the base's own, on the machine it runs on.
- What you addUploads, saved notes and a memory's facts are written on the base's machine, in its own folder, and deleted with it.
- ConversationsThe app keeps its chats on your phone and the Playground in your browser. Only what you save, or a conversation you ask a memory to read, is sent — to the base's machine.
- What Astralyx holdsA base's specification and totals — documents, chunks, progress — never a document's name or text.
What we are building next.
None of this is available yet.
- A knowledge graphEntities and relations across your documents, with dated facts, for questions that span several of them — who decided what, and when.
- Memory that offers itselfFacts proposed from a conversation when it ends, for you to review — off unless you switch it on.
- ConnectorsObsidian vaults, exports from Notion and Evernote, Google Drive, Gmail and Calendar, Outlook and OneDrive — each with your consent, and a switch that disconnects it and deletes what came from it.
Questions we are asked.
Is it RAG on my own machines?
It is the retrieval half, done carefully: the passages that answer a question, each with its citation, from documents that never leave your machines. Your model, agent or assistant writes the answer from them — Claude Code with the Astralyx skill does exactly that.
Which files does it read?
Markdown, text (including CSV, JSON and logs), HTML, PDF and Word (.docx), up to 64 MB each. A PDF's text layer is read; a scanned PDF without one is listed as such — there is no OCR yet.
Does it need a GPU?
No. The embedding model and the reranker the docs suggest run on CPUs, served with Eos, and a base takes one CPU core and 2 GiB of memory by default.
Can a model change my knowledge base by itself?
No. A model — a deployment of yours, or an assistant over MCP — can only propose a note. You see what it would keep and save or discard it; an assistant cannot approve its own proposal, and neither can an API token.
Does memory read all my conversations?
No: only a conversation you choose, when you choose — Remember this conversation in the app or the Playground, or astra metis remember. Nothing is kept but the facts you keep.
Can I use it from Claude Code?
Yes: astra connect claude-code installs the skill, and Claude searches as you, in the workspace you chose with astra use. Add --mcp to connect Astralyx's MCP server as well (how).
How is it priced?
Metis is in early access, as the other products are. A knowledge base runs as ordinary work on your machines and counts like any other; tell us what you want to search and we will work out terms with you.
Ask your documents. Check every answer.
Point a knowledge base at a drive, ask it from where you already work, and measure what it finds.