• Anthropic
  • OpenAI
  • Google
  • xAI
  • Ollama

Ask your company. See the evidence.

Private AI for your company's documents and databases. Cited answers, from the model you choose, on servers you control.

Onirix product tour

A conversation on the left, with the document it produced open beside it

Privacy

Your data stays where you put it.

You decide where the data lives, which model reads it, and who can see each document.

Where the data lives

The whole stack runs in Docker on servers you operate. Originals, the search index, and conversations stay in your own databases and object store.

Which model reads it

Chat and embedding calls go only to the provider you connect, under your own credentials. You can change it at any time.

Who sees which document

Visibility is checked inside the search query, before anything reaches the model. An admin who is not on a team does not read that team's documents.

Fully local when it has to be

With Ollama for both chat and embeddings, a deployment answers questions without a single outbound request.

Features

Answers your team can check.

Every answer and every number stays connected to the file it came from.

01 / citations

Open the passage behind a claim

Citation markers open the cited passage beside the answer, with a link to the original file. When two documents disagree, the answer says so.

02 / charts

Chart the numbers in your files

Spreadsheets and reports become bar, line, area, pie, or scatter charts. Each series cites its sheet, and you can open the table or copy the data.

03 / models

Pick the model, change it later

Connect OpenAI, Anthropic, Google, xAI, or your own Ollama server. Enable several at once and set the default.

In the box

Built for the files a company really has.

The formats you already use

PDF, Word, Excel, CSV, Markdown, HTML, JSON, and text, indexed with page and sheet anchors.

Connected sources

Websites, Google Drive, OneDrive, and S3 buckets sync on a schedule, and only what changed is indexed again.

Databases, read-only

Ask questions of a connected database. Each query is a single SELECT in a read-only transaction with a row cap and a timeout.

Says when it does not know

Missing evidence is stated, and sourced facts are kept apart from inference.

Answers stay auditable

Cited passages are saved with the conversation, so last month's answer can still be checked after a reindex.

Roles apart from reading rights

Owner, admin, member, and custom roles govern administration. Reading a document depends on teams and visibility only.

Self-hosting

Your infrastructure. Your model.

One compose file brings up the application, its worker, and the four services they depend on.

docker compose6 services
  • dashboardNext.js application
  • workerBackground indexing
  • postgresMetadata, chats, citations
  • opensearchHybrid vector and keyword index
  • redisIndexing queue
  • minioUploaded originals, or any S3 store
$pnpm run docker:up

Desktop app

Onirix on your desktop.

Your workspace in its own window. Open models download and run on the same computer.

Coming soonAll platforms
  • macOSApple silicon and Intel
  • Windows64-bit installer
  • LinuxAppImage, x64
  • A client for an Onirix server. Free.

FAQ

Questions a security review asks first.

Onirix does not train or fine-tune anything. It retrieves the passages that answer a question and sends them to the provider you connected, under your own API key, so that provider's API terms are the ones that apply. With Ollama, nothing is sent anywhere.

OpenAI, Anthropic, Google, xAI, and self-hosted Ollama. You can enable several at once, choose the default, and use one provider for chat and another for embeddings.

Yes. Self-host the stack and use Ollama for both chat and embeddings, and questions are answered without an outbound request. Cloud sources such as Google Drive still need network access to sync.

Docker. The compose file runs the application and its indexing worker beside PostgreSQL, OpenSearch, Redis, and MinIO or any other S3-compatible store.

You choose: the whole organization, selected teams plus you, or only you. The rule is applied inside the search query, and administrators get no reading rights from their role. Conversations are private to their author.

PDF, Word, Excel, CSV, Markdown, HTML, JSON, and plain text, up to 50 MB per file. It also syncs websites, Google Drive, OneDrive, and S3 buckets. PowerPoint and images are not supported yet.

No, Onirix runs in the browser. The desktop app is a client for an Onirix server, and it can download open models and run them on the same computer.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • Ollama

Private AI workspace

Ask your company. See the evidence.

Create a workspace, connect a provider, upload a folder.