Skip to content

Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more

feat(fi): add local backend selection

altic-dev89b461733f2f
diff --git a/Sources/Fluid/ContentView.swift b/Sources/Fluid/ContentView.swift
index 52472dd5..2653add5 100644
--- a/Sources/Fluid/ContentView.swift
+++ b/Sources/Fluid/ContentView.swift
@@ -2300,6 +2300,14 @@ struct ContentView: View {
finalText = normalizedTranscribedText
}
let postProcessingLatencyMs = Int((Date().timeIntervalSince(postProcessingStart) * 1000).rounded())
+ let postProcessingProviderName = postProcessingModelInfo.provider ?? "unknown"
+ let postProcessingModelName = postProcessingModelInfo.model ?? "unknown"
+ DebugLogger.shared.info(
+ "Dictation AI post-processing finished in \(postProcessingLatencyMs)ms "
+ + "provider=\(postProcessingProviderName) model=\(postProcessingModelName) "
+ + "inputChars=\(postProcessingInputChars) fallback=\(aiFallbackReason != nil)",
+ source: "ContentView"
+ )
AnalyticsService.shared.capture(
.dictationPostProcessingCompleted,
properties: [
diff --git a/Sources/Fluid/Persistence/BackupService.swift b/Sources/Fluid/Persistence/BackupService.swift
index 4be78d15..5acd775e 100644
--- a/Sources/Fluid/Persistence/BackupService.swift
+++ b/Sources/Fluid/Persistence/BackupService.swift
@@ -14,6 +14,7 @@ struct SettingsBackupPayload: Codable, Equatable {
let modelReasoningConfigs: [String: SettingsStore.ModelReasoningConfig]
let privateAIPrefixKVCacheEnabled: Bool?
let privateAIBoostEnabled: Bool?
+ let privateAIBackendPreference: SettingsStore.PrivateAIBackendPreference?
let privateAIContextTokenLimit: Int?
let selectedSpeechModel: SettingsStore.SpeechModel
let selectedCohereLanguage: SettingsStore.CohereLanguage
diff --git a/Sources/Fluid/Persistence/SettingsStore.swift b/Sources/Fluid/Persistence/SettingsStore.swift
index 35340cd7..278d8bfc 100644
--- a/Sources/Fluid/Persistence/SettingsStore.swift
+++ b/Sources/Fluid/Persistence/SettingsStore.swift
@@ -70,6 +70,39 @@ final class SettingsStore: ObservableObject {
return min(requestedOutputTokens, availableOutputTokens)
}
+ enum PrivateAIBackendPreference: String, Codable, CaseIterable, Identifiable {
+ case auto
+ case llama
+ case mlx
+
+ var id: String { self.rawValue }
+
+ /// Default backend when no preference is stored.
+ /// Apple Silicon → MLX (fastest Fluid-1 path). Intel → llama.cpp.
+ static var systemDefault: PrivateAIBackendPreference {
+ CPUArchitecture.isAppleSilicon ? .mlx : .llama
+ }
+
+ var displayName: String {
+ switch self {
+ case .auto: return "Auto"
+ case .llama: return "llama.cpp"
+ case .mlx: return "MLX"
+ }
+ }
+
+ var detail: String {
+ switch self {
+ case .auto:
+ return CPUArchitecture.isAppleSilicon ? "Uses MLX on Apple Silicon; errors if MLX is missing." : "Uses llama.cpp on Intel."
+ case .llama:
+ return "GGUF baseline; works on Intel and Apple Silicon."
+ case .mlx:
+ return "Apple Silicon only; fastest local Fluid-1 path."
+ }
+ }
+ }
+
// MARK: - Prompt Profiles (Unified)
enum PromptMode: String, Codable, CaseIterable, Identifiable {
@@ -1463,6 +1496,25 @@ final class SettingsStore: ObservableObject {
}
}
+ var privateAIBackendPreference: PrivateAIBackendPreference {
+ get {
+ let rawValue = self.defaults.string(forKey: Keys.privateAIBackendPreference)?
+ .trimmingCharacters(in: .whitespacesAndNewlines)
+ .lowercased()
+ let preference = rawValue.flatMap(PrivateAIBackendPreference.init(rawValue:))
+ ?? PrivateAIBackendPreference.systemDefault
+ if preference == .mlx, CPUArchitecture.isIntel {
+ return .llama
+ }
+ return preference
+ }
+ set {
+ objectWillChange.send()
+ let preference = newValue == .mlx && CPUArchitecture.isIntel ? .llama : newValue
+ self.defaults.set(preference.rawValue, forKey: Keys.privateAIBackendPreference)
+ }
+ }
+
var privateAIContextTokenLimit: Int {
get {
let value = self.defaults.integer(forKey: Keys.privateAIContextTokenLimit)
@@ -2833,6 +2885,7 @@ final class SettingsStore: ObservableObject {
modelReasoningConfigs: self.modelReasoningConfigs,
privateAIPrefixKVCacheEnabled: self.privateAIPrefixKVCacheEnabled,
privateAIBoostEnabled: self.privateAIBoostEnabled,
+ privateAIBackendPreference: self.privateAIBackendPreference,
privateAIContextTokenLimit: self.privateAIContextTokenLimit,
selectedSpeechModel: self.selectedSpeechModel,
selectedCohereLanguage: self.selectedCohereLanguage,
@@ -2932,6 +2985,9 @@ final class SettingsStore: ObservableObject {
if let privateAIBoostEnabled = payload.privateAIBoostEnabled {
self.privateAIBoostEnabled = privateAIBoostEnabled
}
+ if let privateAIBackendPreference = payload.privateAIBackendPreference {
+ self.privateAIBackendPreference = privateAIBackendPreference
+ }
if let privateAIContextTokenLimit = payload.privateAIContextTokenLimit {
self.privateAIContextTokenLimit = privateAIContextTokenLimit
}
@@ -4659,6 +4715,7 @@ private extension SettingsStore {
static let selectedProviderID = "SelectedProviderID"
static let privateAIPrefixKVCacheEnabled = "PrivateAIProviderPrefixKVCacheEnabled"
static let privateAIBoostEnabled = "PrivateAIProviderBoostEnabled"
+ static let privateAIBackendPreference = "FluidIntelligenceBackendPreference"
static let privateAIContextTokenLimit = "PrivateAIProviderContextTokenLimit"
static let privateAIContextDefaultMigratedTo4K = "PrivateAIProviderContextDefaultMigratedTo4K"
static let providerAPIKeys = "ProviderAPIKeys"
diff --git a/Sources/Fluid/Services/PrivateAIProvider.swift b/Sources/Fluid/Services/PrivateAIProvider.swift
index 4a0b57da..176d5c50 100644
--- a/Sources/Fluid/Services/PrivateAIProvider.swift
+++ b/Sources/Fluid/Services/PrivateAIProvider.swift
@@ -283,7 +283,20 @@ enum PrivateAIProviderFeature {
}
nonisolated static func verificationFingerprint(for modelID: String) -> String {
- "private-ai-provider|\(modelID)"
+ let backendPreference = UserDefaults.standard.string(forKey: "FluidIntelligenceBackendPreference")?
+ .trimmingCharacters(in: .whitespacesAndNewlines)
+ .lowercased()
+ let normalizedBackendPreference: String
+ if let backendPreference,
+ !backendPreference.isEmpty,
+ SettingsStore.PrivateAIBackendPreference(rawValue: backendPreference) != nil
+ {
+ normalizedBackendPreference = backendPreference
+ } else {
+ // Keep fingerprint in sync with SettingsStore.privateAIBackendPreference default.
+ normalizedBackendPreference = SettingsStore.PrivateAIBackendPreference.systemDefault.rawValue
+ }
+ return "private-ai-provider|\(modelID)|backend:\(normalizedBackendPreference)"
}
}
diff --git a/Sources/Fluid/UI/AISettings/AIEnhancementSettingsViewModel.swift b/Sources/Fluid/UI/AISettings/AIEnhancementSettingsViewModel.swift
index 4912d7ee..8c0fe73e 100644
--- a/Sources/Fluid/UI/AISettings/AIEnhancementSettingsViewModel.swift
+++ b/Sources/Fluid/UI/AISettings/AIEnhancementSettingsViewModel.swift
@@ -379,48 +379,47 @@ final class AIEnhancementSettingsViewModel: ObservableObject {
let providerID = PrivateAIProviderFeature.shared.providerID
let key = self.providerKey(for: providerID)
guard !self.isTestingConnection else { return false }
+ let currentModel = PrivateAIModelRegistry.model(id: model.id) ?? model
self.isTestingConnection = true
self.updateConnectionStatus(.testing, for: providerID)
- self.connectionErrorMessage = ""
defer {
self.isTestingConnection = false
}
- guard PrivateAIIntegrationService.isModelInstalled(model) else {
+ guard PrivateAIIntegrationService.isModelInstalled(currentModel) else {
self.updateConnectionStatus(.failed, for: providerID)
- self.connectionErrorMessage = "\(model.displayName) is not installed."
+ self.setConnectionError("\(currentModel.displayName) is not installed.", for: providerID)
return false
}
do {
- let status = try await PrivateAIIntegrationService.shared.loadModel(model)
+ let status = try await PrivateAIIntegrationService.shared.loadModel(currentModel)
switch status.state {
case .ready:
var fingerprints = self.settings.verifiedProviderFingerprints
- fingerprints[key] = self.privateAIFingerprint(for: model.id)
+ fingerprints[key] = self.privateAIFingerprint(for: currentModel.id)
self.settings.verifiedProviderFingerprints = fingerprints
- self.selectedModelByProvider[key] = model.id
+ self.selectedModelByProvider[key] = currentModel.id
self.settings.selectedModelByProvider = self.selectedModelByProvider
self.selectProviderForUse(providerID)
self.updateConnectionStatus(.success, for: providerID)
- self.connectionErrorMessage = ""
DebugLogger.shared.info(
- "Private AI Provider verification succeeded for \(model.id)",
+ "Private AI Provider verification succeeded for \(currentModel.id)",
source: "AISettingsView"
)
return true
default:
self.updateConnectionStatus(.failed, for: providerID)
- self.connectionErrorMessage = status.message ?? "\(model.displayName) did not report ready."
+ self.setConnectionError(status.message ?? "\(currentModel.displayName) did not report ready.", for: providerID)
return false
}
} catch {
self.updateConnectionStatus(.failed, for: providerID)
- self.connectionErrorMessage = self.privateAIErrorMessage(for: error)
+ self.setConnectionError(self.privateAIErrorMessage(for: error), for: providerID)
DebugLogger.shared.error(
- "Private AI Provider verification failed for \(model.id): \(self.connectionErrorMessage)",
+ "Private AI Provider verification failed for \(currentModel.id): \(self.connectionErrorMessage)",
source: "AISettingsView"
)
return false
diff --git a/Sources/Fluid/UI/AISettingsView+AIConfiguration.swift b/Sources/Fluid/UI/AISettingsView+AIConfiguration.swift
index 97cc13f2..9bf57e57 100644
--- a/Sources/Fluid/UI/AISettingsView+AIConfiguration.swift
+++ b/Sources/Fluid/UI/AISettingsView+AIConfiguration.swift
@@ -685,6 +685,8 @@ extension AIEnhancementSettingsView {
}
}
+ self.privateAIBackendRow(isBusy: isBusy)
+
if isDownloading || isLoading || isLoaded || hasLoadFailure || isVerified || !isInstalled {
self.privateAIModelStatusRow(
status: status,
@@ -695,7 +697,9 @@ extension AIEnhancementSettingsView {
}
self.privateAIPrefixCacheRow(isBusy: isBusy)
- self.privateAIBoostRow(isBusy: isBusy)
+ if self.privateAIShowsBoostRow {
+ self.privateAIBoostRow(isBusy: isBusy)
+ }
if self.viewModel.connectionStatus(for: PrivateAIProviderFeature.shared.providerID) == .failed,
!self.viewModel.connectionErrorMessage.isEmpty
@@ -775,6 +779,40 @@ extension AIEnhancementSettingsView {
}
}
+ private func privateAIBackendRow(isBusy: Bool) -> some View {
+ HStack(alignment: .center, spacing: 8) {
+ Text("Backend")
+ .font(.caption)
+ .frame(width: 124, alignment: .leading)
+
+ self.privateAIBackendPicker(isBusy: isBusy)
+ .frame(width: 140)
+
+ Text(self.settings.privateAIBackendPreference.detail)
+ .font(.caption2)
+ .foregroundStyle(.secondary)
+ .lineLimit(2)
+
+ Spacer(minLength: 0)
+ }
+ }
+
+ private func privateAIBackendPicker(isBusy: Bool) -> some View {
+ Picker("", selection: self.privateAIBackendBinding) {
+ ForEach(self.privateAISelectableBackendPreferences) { preference in
+ Text(preference.displayName).tag(preference)
+ }
+ }
+ .pickerStyle(.menu)
+ .labelsHidden()
+ .disabled(isBusy)
+ .help("Local Fluid-1 runtime. Default is MLX on Apple Silicon.")
+ }
+
+ private var privateAISelectableBackendPreferences: [SettingsStore.PrivateAIBackendPreference] {
+ CPUArchitecture.isIntel ? [.auto, .llama] : SettingsStore.PrivateAIBackendPreference.allCases
+ }
+
private func privateAIBoostRow(isBusy: Bool) -> some View {
HStack(alignment: .center, spacing: 8) {
Text("Faster results")
@@ -881,6 +919,26 @@ extension AIEnhancementSettingsView {
)
}
+ private var privateAIBackendBinding: Binding<SettingsStore.PrivateAIBackendPreference> {
+ Binding(
+ get: { self.settings.privateAIBackendPreference },
+ set: { preference in
+ self.setPrivateAIBackendPreference(preference)
+ }
+ )
+ }
+
+ private var privateAIShowsBoostRow: Bool {
+ switch self.settings.privateAIBackendPreference {
+ case .llama:
+ return true
+ case .auto:
+ return CPUArchitecture.isIntel
+ case .mlx:
+ return false
+ }
+ }
+
private var privateAIContextTokenLimitBinding: Binding<Int> {
Binding(
get: { self.settings.privateAIContextTokenLimit },
@@ -1213,6 +1271,30 @@ extension AIEnhancementSettingsView {
}
}
+ private func setPrivateAIBackendPreference(_ preference: SettingsStore.PrivateAIBackendPreference) {
+ guard self.settings.privateAIBackendPreference != preference else { return }
+ let modelID = self.privateAISelectedModelID
+
+ self.settings.privateAIBackendPreference = preference
+ UserDefaults.standard.removeObject(forKey: PrivateAIIntegrationService.localModelPathDefaultsKey)
+ self.privateAILoadState = .idle
+ self.viewModel.resetVerification(for: PrivateAIProviderFeature.shared.providerID)
+
+ Task { @MainActor in
+ await PrivateAIIntegrationService.shared.unloadCachedRuntime(
+ reason: "Fluid Intelligence backend changed to \(preference.displayName)"
+ )
+ guard self.privateAISelectedModelID == modelID else { return }
+ let model = self.selectedPrivateAIModel
+ if PrivateAIIntegrationService.isModelInstalled(model) {
+ self.verifyPrivateAIConnection(model)
+ } else {
+ self.privateAILoadState = .idle
+ self.viewModel.refreshProviderItems()
+ }
+ }
+ }
+
private func providerDetailsSection(for item: ProviderItem) -> AnyView {
let isAppleDisabled = item.id == "apple-intelligence-disabled"
let isApple = item.id == "apple-intelligence"
@@ -2002,6 +2084,20 @@ extension AIEnhancementSettingsView {
Spacer(minLength: 0)
}
+ HStack(alignment: .center, spacing: 12) {
+ self.privateAISettingLabel("Backend", systemImage: "cpu")
+
+ self.privateAIBackendPicker(isBusy: isBusy)
+ .frame(width: 160)
+
+ Text(self.settings.privateAIBackendPreference.detail)
+ .font(.caption2)
+ .foregroundStyle(.secondary)
+ .lineLimit(2)
+
+ Spacer(minLength: 0)
+ }
+
HStack(alignment: .center, spacing: 12) {
self.privateAISettingLabel("Dictation window", systemImage: "memorychip")
@@ -2021,7 +2117,9 @@ extension AIEnhancementSettingsView {
}
self.privateAIPrefixCacheRow(isBusy: isBusy)
- self.privateAIBoostRow(isBusy: isBusy)
+ if self.privateAIShowsBoostRow {
+ self.privateAIBoostRow(isBusy: isBusy)
+ }
}
HStack(spacing: 8) {