The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
/**
* Is a Discover row already on this machine, and if so, which local model is it.
*
* Pulled out of DiscoverModels.tsx (GH #118) because the answer is the whole
* bug. a user's fresh Windows 11 install downloaded a chat model, the tile said
* Installed, a restart flipped it back to "Get", and the button then did
* nothing while the multi-gigabyte file sat on the disk the entire time.
*
* THE RULE, and the reason this file exists: installed is a question about the
* DISK, never about a port. Every input below is filesystem evidence:
*
* - `downloads`: this session's finished downloads, which are files that were
* written. Fastest path, gone after a restart, never the only source.
* - `installedModels`: what the model refresh read off the disk. For the
* built-in engine that list comes from `list_bundled_models`, a Rust
* directory scan that does not talk to the engine at all, and for Ollama and
* LM Studio from their own on-disk stores.
*
* Nothing here may ever consult engine health, a `/v1/models` request or a
* provider connection state. An engine that is down is a reason to start it,
* never a reason to tell the user their 8 GB file is not there.
*/
import { hfUrlToOllamaRef } from './hf-to-provider'
import { findLocalGgufInstalled, type InstalledModelLike } from './lmstudio-match'
/** The Discover-row fields this decision needs. */
export interface InstalledLookupModel {
/** GGUF file name of the row, when it names one. */
filename?: string
/** Direct download URL, used to derive the canonical Ollama reference. */
downloadUrl?: string
/** Ollama tag, for rows that are Ollama-native. */
ollamaModel?: string
}
/** The download-store shape this decision needs. */
export type DownloadStatuses = Record<string, { status?: string } | undefined>
/**
* The installed model behind a Discover row, or null.
*
* The returned entry carries the picker id (`name`), which is what turns the
* tile's badge into an action: a row that knows WHICH model it is can load it
* into the chat.
*/
export function findInstalledForDiscoverModel(
model: InstalledLookupModel,
downloads: DownloadStatuses,
installedModels: InstalledModelLike[],
): InstalledModelLike | null {
const installedOllamaTags = installedModels
.filter((m) => m.provider === 'ollama')
.map((m) => (m.model || m.name || '').toLowerCase())
const ollamaEntry = (tag: string): InstalledModelLike | null =>
installedModels.find(
(m) => m.provider === 'ollama' && (m.model || m.name || '').toLowerCase() === tag,
) ?? null
if (model.ollamaModel) {
const tag = model.ollamaModel.toLowerCase()
if (installedOllamaTags.includes(tag)) return ollamaEntry(tag)
// Ollama appends `:latest` to bare model names, accept either form
if (!tag.includes(':') && installedOllamaTags.includes(`${tag}:latest`)) {
return ollamaEntry(`${tag}:latest`)
}
}
if (model.filename && model.downloadUrl) {
const ref = hfUrlToOllamaRef(model.downloadUrl, model.filename)?.toLowerCase()
if (ref && installedOllamaTags.includes(ref)) return ollamaEntry(ref)
}
// LM Studio AND the built-in engine, both read off the disk. The matcher
// (lib/lmstudio-match.ts) bridges the id forms: LM Studio's basename ids and
// its modern quant-less publisher/short key, and the built-in engine's file
// stem.
if (model.filename) {
const local = findLocalGgufInstalled(model.filename, installedModels)
if (local) return local
}
// A download that finished in THIS session is a file on the disk too, and it
// is the only evidence that exists before the next model refresh runs. It
// comes last because it carries no picker id, so an entry from the disk scan
// is always the more useful answer.
if (model.filename && downloads[model.filename]?.status === 'complete') return {}
return null
}
/** The badge question, unchanged in meaning from the pre-2.6.8 closure. */
export function isDiscoverModelInstalled(
model: InstalledLookupModel,
downloads: DownloadStatuses,
installedModels: InstalledModelLike[],
): boolean {
return findInstalledForDiscoverModel(model, downloads, installedModels) !== null
}