Skip to content

Router that virtually distributes inference across connected devices in the home.

Go

Update onboarding screenshots and overview documentation

Terveed72f8cf18ce
diff --git a/docs/assets/onboarding/engine-lifecycle/01-engine-installing.png b/docs/assets/onboarding/engine-lifecycle/01-engine-installing.png
index 1a371a9..d3b8dc1 100644
Binary files a/docs/assets/onboarding/engine-lifecycle/01-engine-installing.png and b/docs/assets/onboarding/engine-lifecycle/01-engine-installing.png differ
diff --git a/docs/assets/onboarding/engine-lifecycle/04-model-actions.png b/docs/assets/onboarding/engine-lifecycle/04-model-actions.png
index 80708fb..6d1eb30 100644
Binary files a/docs/assets/onboarding/engine-lifecycle/04-model-actions.png and b/docs/assets/onboarding/engine-lifecycle/04-model-actions.png differ
diff --git a/docs/assets/onboarding/getting-started/01-first-run-setup.png b/docs/assets/onboarding/getting-started/01-first-run-setup.png
index df83f0f..2de9123 100644
Binary files a/docs/assets/onboarding/getting-started/01-first-run-setup.png and b/docs/assets/onboarding/getting-started/01-first-run-setup.png differ
diff --git a/docs/assets/onboarding/getting-started/02-overview-single-node.png b/docs/assets/onboarding/getting-started/02-overview-single-node.png
index 781fb64..082116d 100644
Binary files a/docs/assets/onboarding/getting-started/02-overview-single-node.png and b/docs/assets/onboarding/getting-started/02-overview-single-node.png differ
diff --git a/docs/assets/onboarding/getting-started/03-add-node.png b/docs/assets/onboarding/getting-started/03-add-node.png
index 3a29413..d12177e 100644
Binary files a/docs/assets/onboarding/getting-started/03-add-node.png and b/docs/assets/onboarding/getting-started/03-add-node.png differ
diff --git a/docs/assets/onboarding/getting-started/04-pairing-pin.png b/docs/assets/onboarding/getting-started/04-pairing-pin.png
index a700461..5154492 100644
Binary files a/docs/assets/onboarding/getting-started/04-pairing-pin.png and b/docs/assets/onboarding/getting-started/04-pairing-pin.png differ
diff --git a/docs/assets/onboarding/getting-started/05-connected-nodes.png b/docs/assets/onboarding/getting-started/05-connected-nodes.png
index b0a0e8e..0c61700 100644
Binary files a/docs/assets/onboarding/getting-started/05-connected-nodes.png and b/docs/assets/onboarding/getting-started/05-connected-nodes.png differ
diff --git a/docs/assets/onboarding/getting-started/06-engine-settings.png b/docs/assets/onboarding/getting-started/06-engine-settings.png
index 78ad973..d4fe338 100644
Binary files a/docs/assets/onboarding/getting-started/06-engine-settings.png and b/docs/assets/onboarding/getting-started/06-engine-settings.png differ
diff --git a/docs/assets/onboarding/getting-started/07-add-model.png b/docs/assets/onboarding/getting-started/07-add-model.png
index e485e55..065b00c 100644
Binary files a/docs/assets/onboarding/getting-started/07-add-model.png and b/docs/assets/onboarding/getting-started/07-add-model.png differ
diff --git a/docs/assets/onboarding/getting-started/08-api-endpoints.png b/docs/assets/onboarding/getting-started/08-api-endpoints.png
index 1682f95..56cb836 100644
Binary files a/docs/assets/onboarding/getting-started/08-api-endpoints.png and b/docs/assets/onboarding/getting-started/08-api-endpoints.png differ
diff --git a/docs/assets/onboarding/getting-started/09-job-served-by-node.png b/docs/assets/onboarding/getting-started/09-job-served-by-node.png
index d46b5ba..3bd0724 100644
Binary files a/docs/assets/onboarding/getting-started/09-job-served-by-node.png and b/docs/assets/onboarding/getting-started/09-job-served-by-node.png differ
diff --git a/docs/assets/onboarding/getting-started/10-engine-ports.png b/docs/assets/onboarding/getting-started/10-engine-ports.png
index cb79770..0279dac 100644
Binary files a/docs/assets/onboarding/getting-started/10-engine-ports.png and b/docs/assets/onboarding/getting-started/10-engine-ports.png differ
diff --git a/docs/assets/onboarding/getting-started/11-service-health.png b/docs/assets/onboarding/getting-started/11-service-health.png
index 8d2456b..704379c 100644
Binary files a/docs/assets/onboarding/getting-started/11-service-health.png and b/docs/assets/onboarding/getting-started/11-service-health.png differ
diff --git a/docs/assets/onboarding/getting-started/12-demo-traffic.png b/docs/assets/onboarding/getting-started/12-demo-traffic.png
index 8e5bf31..841a444 100644
Binary files a/docs/assets/onboarding/getting-started/12-demo-traffic.png and b/docs/assets/onboarding/getting-started/12-demo-traffic.png differ
diff --git a/docs/assets/onboarding/overview/01-two-node-cluster.png b/docs/assets/onboarding/overview/01-two-node-cluster.png
index e9370c1..136707b 100644
Binary files a/docs/assets/onboarding/overview/01-two-node-cluster.png and b/docs/assets/onboarding/overview/01-two-node-cluster.png differ
diff --git a/docs/assets/onboarding/terminal-interface/01-tui-overview.png b/docs/assets/onboarding/terminal-interface/01-tui-overview.png
index bb87aeb..1a40bbb 100644
Binary files a/docs/assets/onboarding/terminal-interface/01-tui-overview.png and b/docs/assets/onboarding/terminal-interface/01-tui-overview.png differ
diff --git a/docs/assets/onboarding/terminal-interface/02-tui-pairing.png b/docs/assets/onboarding/terminal-interface/02-tui-pairing.png
index c513459..db5c852 100644
Binary files a/docs/assets/onboarding/terminal-interface/02-tui-pairing.png and b/docs/assets/onboarding/terminal-interface/02-tui-pairing.png differ
diff --git a/docs/assets/onboarding/terminal-interface/03-tui-workloads.png b/docs/assets/onboarding/terminal-interface/03-tui-workloads.png
index 5770d29..0c1bcaa 100644
Binary files a/docs/assets/onboarding/terminal-interface/03-tui-workloads.png and b/docs/assets/onboarding/terminal-interface/03-tui-workloads.png differ
diff --git a/docs/overview.mdx b/docs/overview.mdx
index d7a9d4b..fb519d5 100644
--- a/docs/overview.mdx
+++ b/docs/overview.mdx
@@ -3,7 +3,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All
SPDX-License-Identifier: Apache-2.0
*/}
-# NVIDIA Personal AI Router overview
+# NVIDIA Personal AI Router Overview
NVIDIA Personal AI Router (PAIR) turns several machines on your local network
into one place to send inference requests. You point an application at a local
@@ -15,11 +15,14 @@ the inference engine those applications already speak to, so existing tools work
unchanged while gaining the ability to use another machine's GPU.
You install PAIR on each machine you want to contribute compute and pair those
-machines together. On every one of them PAIR runs two things: an **application**
-you interact with — a desktop window, or a terminal interface for machines with
-no desktop — and a set of **background services** that do the real work. The
-services find the other machines on your network, keep track of which engine and
-which models each one has, and decide where a request should go.
+machines together. On every one of them, PAIR runs two things:
+
+- An **application** you interact with: a desktop window or a terminal
+ interface for machines with no desktop
+- A set of **background services** that do the real work
+
+The services find the other machines on your network, keep track of which engine
+and which models each one has, and decide where a request should go.
This document establishes the vocabulary the rest of the documentation uses. To
install PAIR and send a first request, go to
@@ -53,7 +56,7 @@ These terms have specific meanings in PAIR:
and exposes the JSON-RPC API that both the desktop application and the terminal
interface use.
-## What Happens When You Run PAIR
+## What Happens when You Run PAIR
Running PAIR follows five steps:
@@ -75,8 +78,8 @@ you form a cluster, only the last step repeats.
An application sends an ordinary Ollama-compatible or OpenAI-compatible HTTP
request to a local proxy. The proxy selects one eligible node and forwards the
-whole request. That node's engine performs the inference and the response streams
-back through the proxy.
+whole request. That node's engine performs the inference, and the response
+streams back through the proxy.
```mermaid
flowchart LR
@@ -85,18 +88,28 @@ flowchart LR
Node --> Engine["Inference engine"]
```
-Both legs between machines are encrypted — the request out and the reply in — so
-the whole round trip stays protected rather than just the outbound half. Your
+Between machines, PAIR encrypts both the request out and the reply in, so the
+whole round trip stays protected rather than the outbound half alone. Your
application sees an ordinary reply and never learns which machine served it,
though you can see that yourself.
-A node is eligible when it is reachable, is running an engine that can serve the
-request, and that engine's current inventory advertises the requested model.
-Unknown and non-matching inventories are excluded; if no owner is available, the
-proxy returns a local `502`. Among eligible owners, the proxy honors manual
-selection first, then the scheduler's priority order, then a deterministic
-default. The scheduler ranks nodes by total pending work across both engines, and
-each proxy also counts requests it has just dispatched, so a burst of concurrent
+A node is eligible when all three of these hold:
+
+- It is reachable.
+- It is running an engine that can serve the request.
+- That engine's current inventory advertises the requested model.
+
+The proxy excludes unknown and non-matching inventories. If no owner is
+available, it returns a local `502`.
+
+Among eligible owners, the proxy applies this precedence:
+
+1. Manual selection.
+2. The scheduler's priority order.
+3. A deterministic default.
+
+The scheduler ranks nodes by total pending work across both engines, and each
+proxy also counts requests it has recently dispatched, so a burst of concurrent
requests spreads out instead of waiting for workload reports to catch up.
Because routing is a decision rather than a guarantee, use the **Jobs** view to
@@ -119,14 +132,16 @@ For the trust boundaries in detail, refer to
## What PAIR Provides
+PAIR provides these capabilities:
+
- A local endpoint for compatible AI applications and development tools.
- LAN discovery plus manually configured nodes.
- Ollama-compatible and LM Studio/OpenAI-compatible routing proxies.
- Pairing and cluster membership managed by the background services.
- Model-aware, workload-informed routing of independent requests.
- Encrypted routing between machines: a request sent to another node travels over
- mutual TLS restricted to the nodes you have paired, and a machine that is not a
- cluster member is refused. Local applications reach the proxy over loopback.
+ mutual TLS restricted to the nodes you have paired, and the cluster refuses a
+ machine that is not a member. Local applications reach the proxy over loopback.
- A desktop application, plus a terminal interface for headless machines, driving
the same services.
- Visibility into nodes, engines, models, workloads, and service errors.
diff --git a/scripts/rewrite-docs-links.mjs b/scripts/rewrite-docs-links.mjs
new file mode 100644
index 0000000..1f1d368
--- /dev/null
+++ b/scripts/rewrite-docs-links.mjs
@@ -0,0 +1,151 @@
+#!/usr/bin/env node
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Point the docs pages' repository links at GitHub, for the published site only.
+ *
+ * node scripts/rewrite-docs-links.mjs report what would change
+ * node scripts/rewrite-docs-links.mjs --write rewrite in place
+ *
+ * `docs/*.mdx` is read in two places. On GitHub it is browsed inside the
+ * repository, where `../SECURITY.md` resolves from `docs/` up to the root and
+ * lands on the file in whatever branch or fork the reader is looking at. On the
+ * documentation site those files are not pages at all, and the relative path
+ * climbs above the site root, so every one of them is a dead link.
+ *
+ * Keeping the source relative is what preserves the GitHub behavior, so the
+ * publishing step runs this with `--write` against a throwaway checkout just
+ * before it builds the site. Nothing is committed.
+ *
+ * Never run `--write` against a working tree you intend to commit.
+ *
+ * Report mode is the default and changes nothing, which makes it a check as much
+ * as a preview: a link to a path this tree does not carry fails the run instead
+ * of publishing a URL that resolves nowhere. Trees differ, so run it wherever
+ * the pages are published from.
+ *
+ * Dependency-free and run directly by node, so a fresh clone can use it with no
+ * install step.
+ */
+
+import { readdirSync, readFileSync, statSync, writeFileSync } from 'node:fs'
+import { dirname, join, relative, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const REPO_ROOT = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const DOCS_DIR = join(REPO_ROOT, 'docs')
+
+// The published site is read by people outside NVIDIA, so links resolve against
+// the public mirror's default branch rather than the internal branch being built.
+const GITHUB_BLOB_BASE = 'https://github.com/NVIDIA/Personal-AI-Router/blob/main'
+
+// Matches the inline markdown link and image target, which is the only link form
+// these pages use.
+const LINK_PATTERN = /(!?\[[^\]]*\]\()(\.\.\/[^)\s]+)(\))/g
+
+function mdxFiles(dir) {
+ return readdirSync(dir, { withFileTypes: true }).flatMap((entry) => {
+ const full = join(dir, entry.name)
+ if (entry.isDirectory()) return mdxFiles(full)
+ return entry.name.endsWith('.mdx') ? [full] : []
+ })
+}
+
+// A fenced block can contain example markdown, which must survive untouched.
+function fencedLineFlags(lines) {
+ let fenced = false
+ return lines.map((line) => {
+ if (/^\s*(```|~~~)/.test(line)) {
+ fenced = !fenced
+ return true
+ }
+ return fenced
+ })
+}
+
+function rewriteFile(file) {
+ const original = readFileSync(file, 'utf8')
+ const lines = original.split('\n')
+ const fenced = fencedLineFlags(lines)
+ const changes = []
+ const missing = []
+
+ const rewritten = lines.map((line, index) => {
+ if (fenced[index]) return line
+ return line.replace(LINK_PATTERN, (match, open, target, close) => {
+ const [path, anchor = ''] = splitAnchor(target)
+ const repoPath = relative(REPO_ROOT, resolve(DOCS_DIR, path))
+ if (!exists(join(REPO_ROOT, repoPath))) {
+ missing.push({ line: index + 1, target, repoPath })
+ return match
+ }
+ const url = `${GITHUB_BLOB_BASE}/${repoPath.split('\\').join('/')}${anchor}`
+ changes.push({ line: index + 1, target, url })
+ return `${open}${url}${close}`
+ })
+ })
+
+ return { original, rewritten: rewritten.join('\n'), changes, missing }
+}
+
+function splitAnchor(target) {
+ const hash = target.indexOf('#')
+ return hash === -1 ? [target, ''] : [target.slice(0, hash), target.slice(hash)]
+}
+
+function exists(path) {
+ try {
+ statSync(path)
+ return true
+ } catch {
+ return false
+ }
+}
+
+const write = process.argv.includes('--write')
+let rewrittenCount = 0
+let missingCount = 0
+
+for (const file of mdxFiles(DOCS_DIR).sort()) {
+ const { original, rewritten, changes, missing } = rewriteFile(file)
+ const shown = relative(REPO_ROOT, file)
+
+ for (const entry of missing) {
+ console.error(`${shown}:${entry.line}: ${entry.target} -> ${entry.repoPath} does not exist in the repository`)
+ }
+ missingCount += missing.length
+
+ if (changes.length === 0) continue
+ rewrittenCount += changes.length
+ console.log(`${shown} (${changes.length})`)
+ for (const entry of changes) console.log(` ${entry.line}: ${entry.target} -> ${entry.url}`)
+
+ if (write && rewritten !== original) writeFileSync(file, rewritten)
+}
+
+if (missingCount > 0) {
+ console.error(`\n${missingCount} link(s) point outside the repository. Fix them before publishing.`)
+ process.exit(1)
+}
+
+console.log(`\n${write ? 'Rewrote' : 'Would rewrite'} ${rewrittenCount} repository link(s) across docs/.`)
+
+// A link the pattern did not reach would publish dead, so make that loud rather
+// than leaving it to a reader to discover.
+if (write) {
+ const leftover = mdxFiles(DOCS_DIR)
+ .flatMap((file) => {
+ const lines = readFileSync(file, 'utf8').split('\n')
+ const fenced = fencedLineFlags(lines)
+ return lines
+ .map((text, index) => ({ file: relative(REPO_ROOT, file), line: index + 1, text }))
+ .filter((entry) => !fenced[entry.line - 1] && /\]\(\.\.\//.test(entry.text))
+ })
+ .map((entry) => `${entry.file}:${entry.line}: ${entry.text.trim()}`)
+
+ if (leftover.length > 0) {
+ console.error(`\nRelative repository links survived the rewrite:\n${leftover.join('\n')}`)
+ process.exit(1)
+ }
+}