macOS with Apple Silicon only

The Ferrari of RAG.
On your Mac.

m9 is a native desktop app that turns your Mac into an enterprise-grade retrieval machine. Hybrid pipeline, parallel embedding, 4-phase encryption at rest. Your data never leaves your computer — and even on disk, it's cryptographically sealed.

Request information
100%
Local
3x
Faster
0
Cloud data
4
Encryption layers

Every detail engineered for performance

m9 is not a ChatGPT wrapper. It's a native RAG engine built from scratch to maximize Apple Silicon hardware.

Fused Hybrid Retrieval

BM25 (keyword) + Dense (semantic) pipeline combined with Reciprocal Rank Fusion. It doesn't rely on vectors alone: it also finds exact matches for specific terms, names, acronyms and codes.

Parallel Embedding

Leverages Apple Silicon Performance + Efficiency cores with batch-concurrent embedding. Hundreds of chunks per second, directly on the integrated GPU via Ollama.

Total Privacy + Encryption

No data leaves your Mac. LLM and embedding run locally with Ollama. 4-phase encryption at rest protects documents, metadata, and vectors. Not even a database administrator can read your data.

Multi-Query Expansion

Before searching, m9 reformulates your question into different variants to maximize recall. Synonyms, paraphrases and automatic reformulations expand search coverage.

All Formats

PDF, DOCX, TXT, Markdown, web pages. Ingestion from files, folders or integrated web crawler with configurable depth and automatic download of PDFs found online.

Precise Citations

Every answer includes citations with reference to the source file and exact chunk. Maximum transparency: you can always verify where every piece of information comes from.

RAG pipeline in 6 stages

From ingestion to response, every stage is optimized for quality and speed.

1

Ingest

Parallel extraction from PDF, DOCX, TXT, Web

2

Chunk

Intelligent sentence-aware segmentation with overlap

3

Embed

Batch parallel embedding on Apple Silicon GPU

4

Index

Dual index: LanceDB (vectors) + MiniSearch (BM25)

5

Fuse

Reciprocal Rank Fusion with optimized BM25+Dense weights

6

Answer

Local LLM generates response with verifiable citations

Zero compromises on privacy

In an era when data is the new oil, m9 chooses a different path: everything stays on your Mac.

Your data never leaves. Ever.

m9 has no servers, no accounts, no telemetry. LLM and embedding run entirely locally via Ollama. The vector database (LanceDB) and metadata database (SQLite) are files on your disk. You can disconnect from the internet and m9 works exactly as before. And with 4-phase encryption at rest, even the raw database files are cryptographically unintelligible without your password.

No Cloud No Account No Telemetry Offline Ready GDPR Compliant AES-256-GCM Differential Privacy

4-phase cryptographic shield

Every layer of your data — vectors, metadata, documents, and search index — is protected by a dedicated cryptographic phase. No single attack vector can expose your information.

0

Orthogonal Rotation

Signed Perm + Givens · Differential Privacy

Embedding vectors are rotated through a secret orthogonal matrix and injected with calibrated Gaussian noise. Cosine similarity is preserved for search, but the raw vectors are unintelligible.

Embedding vectors Cosine similarity
1

Format-Preserving

FF3-1 FPE · AES-256 round function

Filenames, paths, URLs, titles, and collection names are encrypted while preserving length and format. The database schema never changes — SQL queries still work.

Filenames Paths URLs Titles Collection names
2

Authenticated Encryption

AES-256-GCM · Random nonce · AAD binding

Document text and chat messages are encrypted with unique random nonces. The authentication tag detects any tampering. AAD binds each ciphertext to its document ID.

Document text Chunk content Chat messages
3

Encrypted Search

HMAC-SHA256 · Searchable Symmetric Encryption

BM25 keyword search works over opaque HMAC tokens instead of plaintext. The server finds documents without ever seeing the actual search terms.

Search index Keyword tokens
Password → PBKDF2 (600k rounds) → Master Key → HKDF-SHA256 → AES Key · SSE Key · FPE Key · ORT Key
Zero-knowledge: master key in memory only DB owner cannot read data No Rust, no C — pure Node.js crypto

Built with cutting-edge technologies

Modern stack, native macOS, no dependencies on external services.

Platform Native macOS (Apple Silicon M1/M2/M3/M4)
Framework Electron + React + TypeScript
Vector Database LanceDB (Rust, embedded)
Metadata Database SQLite (better-sqlite3)
LLM Provider Ollama (local, any model)
Embedding nomic-embed-text / bge-m3 / mxbai
Retrieval Hybrid BM25 + Dense + RRF
Encryption at Rest 4-phase: ORT+DP / FF3-1 FPE / AES-256-GCM / SSE
Key Derivation PBKDF2 + HKDF-SHA256 (4 sub-keys)
Supported Formats PDF, DOCX, TXT, MD, HTML (Web Crawler)
Web Crawler Integrated, with PDF download and domain filter
License Proprietary — Marinuzzi & Associati

Want m9 for your organization?

Fill out the form and we'll get back to you for a custom demo and quote.