Add new LLM API

Author: Mauve Signweaver

README.md

@@ -66,9 +66,9 @@ A minimal web browser for the distributed web
 	- Scan a QR code from the browser action window.
 	- Right click a link or image to generate a QR code for it
 - Built-in ad blocker (ublock origin)
+- Use local and cloud LLMs via `window.llm.chat({messages})` and `window.llm.complete("prompt")`
 - Built-in support for creating web archives via [ArchiveWeb.page](https://github.com/webrecorder/archiveweb.page/)
 - Open links in new windows (right click on element)
-- Find text on the page (`ctrl+f` to bring into focus, `esc` to hide)
 - Autocomplete URLs from history (type in the URL bar, up/down to navigate, right to autocomplete)
 - Persist open windows when quitting
 - Save files from pages (any protocol, right click it)

app/config.js

@@ -12,6 +12,15 @@ const DEFAULT_BT_DIR = path.join(USER_DATA, 'bt')
 const DEFAULT_PAGE = 'agregore://welcome'
 
 export default RC('agregore', {
+  llm: {
+    enabled: true,
+
+    baseURL: 'http://localhost:11434/v1/',
+    // Uncomment this to use OpenAI instead
+    // baseURL: 'https://api.openai.com/v1/'
+    apiKey: 'ollama',
+    model: 'phi3:3.8b-mini-4k-instruct-q4_0'
+  },
   accelerators: {
     OpenDevTools: 'CommandOrControl+Shift+I',
     NewWindow: 'CommandOrControl+N',

app/index.js

@@ -10,6 +10,7 @@ import { WindowManager } from './window.js'
 import { createExtensions } from './extensions/index.js'
 import * as history from './history.js'
 import { version } from './version.js'
+import * as llm from './llm.js'
 
 const IS_DEBUG = process.env.NODE_ENV === 'debug'
 
@@ -142,6 +143,8 @@ async function onready () {
 
   const webSession = session.fromPartition(WEB_PARTITION)
 
+  llm.addPreloads(webSession)
+
   const electronSection = /Electron.+ /i
   const existingAgent = webSession.getUserAgent()
   const newAgent = existingAgent.replace(electronSection, `AgregoreDesktop/${version} `)

app/llm-preload.js

@@ -0,0 +1,7 @@
+const { contextBridge, ipcRenderer } = require('electron')
+
+contextBridge.exposeInMainWorld('llm', {
+  chat: (args) => ipcRenderer.invoke('llm-chat', args),
+  complete: (args) => ipcRenderer.invoke('llm-complete', args),
+  isSupported: (prompt, args = {}) => ipcRenderer.invoke('llm-supported', { ...args, prompt })
+})

app/llm.js

@@ -0,0 +1,142 @@
+import config from './config.js'
+import { ipcMain } from 'electron'
+import path from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const __dirname = fileURLToPath(new URL('./', import.meta.url))
+
+const { baseURL, apiKey, model, enabled } = config.llm
+
+let isInitialized = false
+
+ipcMain.handle('llm-supported', async (event) => {
+  return isSupported()
+})
+
+ipcMain.handle('llm-chat', async (event, args) => {
+  return chat(args)
+})
+
+ipcMain.handle('llm-complete', async (event, args) => {
+  return complete(args)
+})
+
+export async function isSupported () {
+  if (!enabled) return false
+  const has = await hasModel()
+  if (has) return true
+  return (apiKey === 'ollama')
+}
+
+export function addPreloads (session) {
+  const preloadPath = path.join(__dirname, 'llm-preload.js')
+  const preloads = session.getPreloads()
+  preloads.push(preloadPath)
+  session.setPreloads(preloads)
+}
+
+export async function init () {
+  if (!enabled) throw new Error('LLM API is not enabled')
+  if (isInitialized) return
+  // TODO: prompt for download
+  if (apiKey === 'ollama') {
+    const has = await hasModel()
+    if (!has) {
+      await pullModel()
+    }
+  }
+  isInitialized = true
+}
+
+async function listModels () {
+  const { data } = await get('./models', 'Unable to list models')
+  return data
+}
+
+async function pullModel () {
+  await post('/api/pull', {
+    name: model
+  }, `Unable to pull model ${model}`)
+}
+
+async function hasModel () {
+  try {
+    const models = await listModels()
+
+    return !!models.find(({ id }) => id === model)
+  } catch {
+    return false
+  }
+}
+
+export async function chat ({
+  messages = [],
+  temperature,
+  maxTokens,
+  stop
+}) {
+  await init()
+  const { choices } = await post('./chat/completions', {
+    messages,
+    model,
+    temperature,
+    max_tokens: maxTokens,
+    stop
+  }, 'Unable to generate completion')
+
+  return choices[0].text
+}
+
+export async function complete ({
+  prompt,
+  temperature,
+  maxTokens,
+  stop
+}) {
+  await init()
+  const { choices } = await post('./completions', {
+    prompt,
+    model,
+    temperature,
+    max_tokens: maxTokens,
+    stop
+  }, 'Unable to generate completion')
+
+  return choices[0].text
+}
+
+async function get (path, errorMessage) {
+  const url = new URL(path, baseURL).href
+
+  const response = await fetch(url, {
+    method: 'GET',
+    headers: {
+      Authorization: `Bearer ${apiKey}`
+    }
+  })
+
+  if (!response.ok) {
+    throw new Error(`${errorMessage} ${await response.text()}`)
+  }
+
+  return await response.json()
+}
+
+async function post (path, data, errorMessage) {
+  const url = new URL(path, baseURL).href
+
+  const response = await fetch(url, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json; charset=utf8',
+      Authorization: `Bearer ${apiKey}`
+    },
+    body: JSON.stringify(data)
+  })
+
+  if (!response.ok) {
+    throw new Error(`${errorMessage} ${await response.text()}`)
+  }
+
+  return await response.json()
+}