Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 28 additions & 0 deletions apple/TablePro/Core/AI/AIProviderDraftRules.swift
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,34 @@ internal enum AIProviderDraftRules {
return fetched.count == 1 ? fetched.first?.id : nil
}

// Without an effort, a model that always reasons runs the server's default, which can be its costliest.
internal static func effort(
afterPicking current: ReasoningEffort?,
reasoning: AIReasoningSupport,
curated: CuratedModel?
) -> ReasoningEffort? {
let levels = reasoning.effortLevels
guard !levels.isEmpty else { return nil }
if let current, levels.contains(current) { return current }
guard reasoning.isMandatory else { return curated?.defaultEffort }
return current.flatMap(reasoning.clampedEffort)
?? curated?.defaultEffort
?? reasoning.defaultEffort
?? levels.first
}

// A typed id is incomplete until the last keystroke; the send path fits the effort to the model anyway.
internal static func effort(
keeping current: ReasoningEffort?,
reasoning: AIReasoningSupport
) -> ReasoningEffort? {
guard reasoning.isMandatory, !reasoning.effortLevels.isEmpty else { return current }
if let current, reasoning.effortLevels.contains(current) { return current }
return current.flatMap(reasoning.clampedEffort)
?? reasoning.defaultEffort
?? reasoning.effortLevels.first
}

/// Save waits for a model only while the provider's own list is there to pick one from. With
/// no list, because the fetch is blocked or failed or the provider has not been signed in to
/// yet, an empty model is saved as before and can be set later.
Expand Down
25 changes: 23 additions & 2 deletions apple/TablePro/Core/AI/AIProviderFactory.swift
Original file line number Diff line number Diff line change
Expand Up @@ -69,18 +69,40 @@ enum AIProviderFactory {
}
}

// `onlyWhereRequired` is for inline suggestions: they send an effort only where a model always reasons.
static func reasoningEffort(
_ requested: ReasoningEffort?,
config: AIProviderConfig,
model: String,
onlyWhereRequired: Bool = false,
catalog: AIModelCatalog = .shared
) -> ReasoningEffort? {
guard let descriptor = AIProviderRegistry.shared.descriptor(for: config.type.rawValue) else {
return onlyWhereRequired ? nil : requested
}
let fetched = catalog.fetchedInfo(providerID: config.id, modelID: model)
if onlyWhereRequired,
!descriptor.reasoning(forModelID: model, fetched: fetched, preset: config.preset).isMandatory {
return nil
}
return descriptor.effort(requested, forModelID: model, fetched: fetched, preset: config.preset)
}

static func resolve(
settings: AISettings,
overrideProviderId: UUID? = nil,
overrideModel: String? = nil
) -> ResolvedProvider? {
guard settings.enabled else { return nil }
let config: AIProviderConfig?
// A model picked for a provider that was since removed belongs to that provider, not the active one.
var model = overrideModel
if let overrideProviderId,
let match = settings.providers.first(where: { $0.id == overrideProviderId }) {
config = match
} else {
config = settings.activeProvider
if overrideProviderId != nil { model = nil }
}
guard let config else { return nil }
let apiKey: String?
Expand All @@ -91,7 +113,6 @@ enum AIProviderFactory {
apiKey = nil
}
let provider = createProvider(for: config, apiKey: apiKey)
let model = overrideModel ?? config.model
return ResolvedProvider(provider: provider, model: model, config: config)
return ResolvedProvider(provider: provider, model: model ?? config.model, config: config)
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -72,6 +72,12 @@ final class AIChatInlineSource: InlineSuggestionSource {
options: ChatTransportOptions(
model: resolved.model,
systemPrompt: systemPrompt,
reasoningEffort: AIProviderFactory.reasoningEffort(
resolved.config.reasoningEffort,
config: resolved.config,
model: resolved.model,
onlyWhereRequired: true
),
sessionId: sessionId
)
)
Expand Down
22 changes: 22 additions & 0 deletions apple/TablePro/Core/AI/Models/AIModelCatalog.swift
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ final class AIModelCatalog: @unchecked Sendable {
private let lock = NSLock()
private var fetched: [UUID: [String: AIModelInfo]] = [:]
private var refreshTokens: [UUID: UUID] = [:]
private var revisions: [UUID: Int] = [:]

init() {}

Expand All @@ -24,13 +25,32 @@ final class AIModelCatalog: @unchecked Sendable {
defer { lock.unlock() }
refreshTokens.removeValue(forKey: providerID)
fetched[providerID] = Self.indexed(models)
revisions[providerID, default: 0] += 1
}

func remove(providerID: UUID) {
lock.lock()
defer { lock.unlock() }
refreshTokens.removeValue(forKey: providerID)
fetched.removeValue(forKey: providerID)
revisions[providerID, default: 0] += 1
}

func revision(providerID: UUID) -> Int {
lock.lock()
defer { lock.unlock() }
return revisions[providerID, default: 0]
}

// For a reader such as the chat panel: it never overtakes a refresh a Save started, and its
// answer is dropped once anything newer was written or begun since it read `revision`.
func store(providerID: UUID, models: [AIModelInfo], ifUnchangedSince revision: Int) {
guard !models.isEmpty else { return }
lock.lock()
defer { lock.unlock() }
guard revisions[providerID, default: 0] == revision else { return }
fetched[providerID] = Self.indexed(models)
revisions[providerID, default: 0] += 1
}

/// Loads a provider's list again after Save dropped the one on record. Open chat windows read
Expand All @@ -49,6 +69,7 @@ final class AIModelCatalog: @unchecked Sendable {
lock.lock()
defer { lock.unlock() }
refreshTokens[providerID] = token
revisions[providerID, default: 0] += 1
return token
}

Expand All @@ -59,6 +80,7 @@ final class AIModelCatalog: @unchecked Sendable {
refreshTokens.removeValue(forKey: providerID)
guard !models.isEmpty else { return }
fetched[providerID] = Self.indexed(models)
revisions[providerID, default: 0] += 1
}

private static func indexed(_ models: [AIModelInfo]) -> [String: AIModelInfo] {
Expand Down
15 changes: 10 additions & 5 deletions apple/TablePro/Core/AI/Models/AIModelOverlay.swift
Original file line number Diff line number Diff line change
Expand Up @@ -47,11 +47,9 @@ enum AIModelOverlay {
id: modelID,
maxOutputTokens: entry.maxOutputTokens,
modalities: [.text, .image],
reasoning: AIReasoningSupport(
mode: .effortOnly,
effortLevels: entry.effortLevels,
defaultEffort: .medium
)
reasoning: entry.effortLevels.isEmpty
? .unsupported
: AIReasoningSupport(mode: .effortOnly, effortLevels: entry.effortLevels, defaultEffort: .medium)
)
}

Expand All @@ -78,7 +76,14 @@ enum AIModelOverlay {
let maxOutputTokens: Int?
}

// First match wins. The models without reasoning come first: OpenAI rejects a reasoning
// effort for them, and `gpt-5-chat-latest` would otherwise match `gpt-5`.
private static let openAITable: [OpenAIEntry] = [
OpenAIEntry(
patterns: ["chat-latest", "chatgpt-4o", "gpt-4", "gpt-3.5"],
effortLevels: [],
maxOutputTokens: nil
),
OpenAIEntry(
patterns: ["gpt-5.6", "gpt-5-6"],
effortLevels: ReasoningEffort.allCases,
Expand Down
20 changes: 19 additions & 1 deletion apple/TablePro/Core/AI/OpenAICompatibleProvider.swift
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@ final class OpenAICompatibleProvider: ChatTransport {
private let model: String
private let maxOutputTokens: Int?
private let treatsForbiddenAsAuthFailure: Bool
private let modelIDs: AIModelIDFilter
private let providerID: UUID?
private let catalog: AIModelCatalog
private let session: URLSession
Expand All @@ -36,6 +37,7 @@ final class OpenAICompatibleProvider: ChatTransport {
model: String = "",
maxOutputTokens: Int? = nil,
treatsForbiddenAsAuthFailure: Bool = false,
modelIDs: AIModelIDFilter = .all,
providerID: UUID? = nil,
catalog: AIModelCatalog = .shared,
session: URLSession = URLSession(configuration: .ephemeral)
Expand All @@ -49,6 +51,7 @@ final class OpenAICompatibleProvider: ChatTransport {
self.model = model.trimmingCharacters(in: .whitespacesAndNewlines)
self.maxOutputTokens = maxOutputTokens
self.treatsForbiddenAsAuthFailure = treatsForbiddenAsAuthFailure
self.modelIDs = modelIDs
self.providerID = providerID
self.catalog = catalog
self.session = session
Expand All @@ -66,6 +69,7 @@ final class OpenAICompatibleProvider: ChatTransport {
model: config.model,
maxOutputTokens: config.maxOutputTokens,
treatsForbiddenAsAuthFailure: config.preset?.rejectsBadKeyWithForbidden ?? false,
modelIDs: config.preset?.modelIDFilter(forEndpoint: config.endpoint) ?? .all,
providerID: config.id,
session: session
)
Expand Down Expand Up @@ -319,6 +323,11 @@ final class OpenAICompatibleProvider: ChatTransport {
return Self.looksLikeAnAPIResponse(data: data, response: httpResponse)
}

// The placeholder model is unknown by design; that answer proves the route and the key.
if model.isEmpty, statusCode == 404, Self.isUnknownModelError(data) {
return true
}

let errorBody = String(data: data, encoding: .utf8) ?? ""
throw AIProviderError.mapHTTPError(
statusCode: statusCode,
Expand All @@ -329,6 +338,13 @@ final class OpenAICompatibleProvider: ChatTransport {
}
}

private static func isUnknownModelError(_ data: Data) -> Bool {
guard let json = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
let error = json["error"] as? [String: Any]
else { return false }
return error["code"] as? String == "model_not_found"
}

/// A wrong Base URL often lands on a reverse proxy's login page or a single-page app's
/// fallback route, both of which answer 200 with HTML that the chat stream cannot read.
private static func looksLikeAnAPIResponse(data: Data, response: HTTPURLResponse) -> Bool {
Expand Down Expand Up @@ -600,7 +616,9 @@ final class OpenAICompatibleProvider: ChatTransport {
)
}

return modelsArray.compactMap(Self.decodeModel(_:)).sorted { $0.id < $1.id }
return modelsArray.compactMap(Self.decodeModel(_:))
.filter { modelIDs.lists($0.id) }
.sorted { $0.id < $1.id }
}

private func fetchOllamaModels() async throws -> [String] {
Expand Down
67 changes: 62 additions & 5 deletions apple/TablePro/Core/AI/OpenAIResponsesProvider.swift
Original file line number Diff line number Diff line change
Expand Up @@ -66,11 +66,68 @@ final class OpenAIResponsesProvider: ChatTransport {
}

func fetchAvailableModels() async throws -> [AIModelInfo] {
try await fetchModelIDs().map { AIModelInfo(id: $0) }
if dialect == .xai, let languageModels = await fetchXAILanguageModels() {
return languageModels
}
let host = resolvedEndpoint?.apiBase.host
return try await fetchModelEntries(resource: AIEndpointStyle.responses.modelsResource, key: "data")
.filter { Self.listsChatModel($0, dialect: dialect, host: host) }
.compactMap { $0["id"] as? String }
.sorted()
.map { AIModelInfo(id: $0) }
}

// `/v1/models` carries no modality; a gateway in xAI's place may not serve `language-models`.
private func fetchXAILanguageModels() async -> [AIModelInfo]? {
guard let entries = try? await fetchModelEntries(resource: "language-models", key: "models") else {
return nil
}
let models = entries.compactMap(Self.decodeXAILanguageModel(_:)).sorted { $0.id < $1.id }
return models.isEmpty ? nil : models
}

static func decodeXAILanguageModel(_ json: [String: Any]) -> AIModelInfo? {
guard let id = json["id"] as? String, !id.isEmpty,
let outputs = json["output_modalities"] as? [String], outputs.contains("text")
else { return nil }
let inputs = json["input_modalities"] as? [String] ?? []
return AIModelInfo(id: id, modalities: inputs.contains("image") ? [.text, .image] : [.text])
}

static func listsChatModel(_ json: [String: Any], dialect: ResponsesDialect, host: String?) -> Bool {
guard let id = json["id"] as? String, !id.isEmpty else { return false }
switch dialect {
case .xai:
if json["image_price"] != nil { return false }
return !["imagine", "-image", "voice"].contains(where: { id.contains($0) })
case .openAI:
guard let host, isOpenAIHost(host) else { return true }
return isOpenAIChatModel(id)
}
}

// Only OpenAI's own hosts are known to list its non-chat models; a gateway lists what it routes.
private static func isOpenAIHost(_ host: String) -> Bool {
let host = host.lowercased()
return host == "api.openai.com" || host.hasSuffix(".api.openai.com")
}

// Checked against the Responses column of every live model on platform.openai.com/docs/models.
// A fine-tune is `ft:<base>:<org>:<suffix>:<id>`; only its base model says what it is.
static func isOpenAIChatModel(_ id: String) -> Bool {
let base = id.hasPrefix("ft:") ? String(id.dropFirst(3).prefix { $0 != ":" }) : id
let families = ["gpt-", "chatgpt-", "chat-latest", "codex-"]
let isOSeries = base.first == "o" && base.dropFirst().first?.isNumber == true
guard isOSeries || families.contains(where: { base.hasPrefix($0) }) else { return false }
let nonChat = [
"image", "audio", "realtime", "transcribe", "tts", "whisper", "-live", "search",
"embedding", "moderation", "instruct", "deep-research", "computer-use"
]
return !nonChat.contains(where: { base.contains($0) })
}

private func fetchModelIDs() async throws -> [String] {
let url = try requestURL(AIEndpointStyle.responses.modelsResource)
private func fetchModelEntries(resource: String, key: String) async throws -> [[String: Any]] {
let url = try requestURL(resource)
var request = URLRequest(url: url)
request.timeoutInterval = AIProvider.modelListTimeout
if let apiKey, !apiKey.isEmpty {
Expand All @@ -87,11 +144,11 @@ final class OpenAIResponsesProvider: ChatTransport {
guard let httpResponse = response as? HTTPURLResponse,
httpResponse.statusCode == 200,
let json = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
let modelsArray = json["data"] as? [[String: Any]]
let entries = json[key] as? [[String: Any]]
else {
throw AIProviderError.networkError("Failed to fetch models")
}
return modelsArray.compactMap { $0["id"] as? String }.sorted()
return entries
}

func testConnection() async throws -> Bool {
Expand Down
Loading
Loading