notebookMLX

Drop a document in, ask it questions, and get answers whose citations open the passage they name.

Embedding happens on the machine, through MLXEmbedders - the documents never leave it. Generation is answered by a dAI gateway, which is itself a fleet of local machines, so the same holds there.

Three repositories

This one is the app. It is a client of a dAI fleet, and shares that project's fork of Apple's MLX examples.

flowchart TB
  fork["<b>mlx-swift-examples</b> · MIT<br/>fork of ml-explore's<br/>adds pipeline parallelism for dense Qwen"]
  dai["<b>dAI</b> · Apache-2.0<br/>control plane + node agent<br/>harvests idle Macs, serves models"]
  nb["<b>notebookMLX</b> · Apache-2.0 · you are here<br/>drop a document in, ask it questions"]
  fork -->|"MLXLLM, MLXLMCommon<br/>pinned revision"| dai
  fork -->|"MLXEmbedders<br/>the same pinned revision"| nb
  nb -->|"HTTP: asks the gateway<br/>to generate an answer"| dai

Loading

	What it is
notebookMLX	This app. Ingests documents, embeds and indexes them locally, answers with citations that resolve to a page.
dAI	The fleet behind the answers: idle Apple Silicon machines, centrally scheduled, serving models over an OpenAI-compatible gateway. Nothing leaves the building.
mlx-swift-examples	The fork both depend on. MLXEmbedders here, MLXLLM there, one pinned revision for both.

Without a gateway configured the app still ingests, embeds and searches - that part is entirely local. It is generation that needs the fleet.

Pointing it somewhere other than dAI

It is an ordinary OpenAI client: POST /v1/chat/completions with a bearer token, reading choices[0].message.content. So any OpenAI-compatible endpoint works - a dAI fleet, a private vLLM or LM Studio, OpenAI itself.

Destinations are saved by name. Settings holds a list: add one, give it a URL, a model and a key, and switch between them from the picker. Each keeps its own key in the Keychain. Deleting a destination removes its key too.

Two things to know before you do.

• Clear the CA certificate path. With one set the app trusts only that authority (SecTrustSetAnchorCertificatesOnly), which is right for a fleet with its own CA and fails every handshake against a public host. Empty means the system roots.
• Name a model. dAI serves whatever the group is pinned to, so the field is optional there; other servers return a 400 without one. The app refuses before sending rather than passing that back as a refusal that looks like a bad key.

Embedding stays local either way: there is no /v1/embeddings call in this app and the index never leaves the machine. What travels is the question and the passages retrieved for it. Settings says so when the endpoint is not a fleet.

What it does

• Ingests PDF, plain text and xlsx. A spreadsheet is a zip of XML, which is why CoreXLSX is a dependency rather than something written here.
• Chunks, embeds and indexes locally, with pooling chosen by model family because the conversions do not ship a 1_Pooling directory to read it from.
• Answers with citations that resolve to a page, and an evidence column beside the answer rather than footnotes underneath it.
• Says when an answer was cut off, and what cut it.

Building

swift build

The app bundle is assembled by packaging/make-app.sh, which wraps the built executable - a menu-bar-free document app needs an Info.plist, and SwiftPM produces a bare binary.

Command-line tools alongside the app: notebook-ingest (one document, with memory reported as it goes), notebook-import (adopt an existing index), notebook-reembed (re-embed keeping chunks and citations), notebook-verify (check this embedder against the agent's fixture), and notebook-demo.

The pinned fork

Package.swift pins djsincla/mlx-swift-examples to an exact revision, and dAI pins the same one. Both use MLXEmbedders; two copies at different revisions could pool or normalise differently, and an index built by one would be silently incomparable with a query from the other. That is not a failure anything reports, so Tests asserts the two agree.

Security

It parses documents it did not write and holds a gateway key in the Keychain. SECURITY.md covers reporting and what those two facts mean.

Licence

Apache License 2.0, © 2026 Dwayne Sinclair - see LICENSE and NOTICE. The pinned fork is MIT and is not covered by that notice.

Model weights are not in this repository and carry their own licences.