| 1 | import 'ollama_reasoning_decoder.dart'; |
| 2 | import 'dart:async'; |
| 3 | import 'dart:convert'; |
| 4 | import 'dart:io'; |
| 5 | import 'dart:isolate'; |
| 6 | import 'dart:typed_data'; |
| 7 | import 'package:dio/dio.dart'; |
| 8 | import '../../servers/data/models/server.dart'; |
| 9 | import 'models/message.dart'; |
| 10 | import 'models/chat_parameters.dart'; |
| 11 | import 'models/mcp_integration.dart'; |
| 12 | import 'tools/tool_definition.dart'; |
| 13 | import '../../../core/models/enums.dart'; |
| 14 | import '../../../core/logger/app_logger.dart'; |
| 15 | import '../../on_device/data/on_device_gemma_service.dart'; |
| 16 | import '../../on_device/data/on_device_chat_service.dart'; |
| 17 | import 'tools/adapters/tool_transport_adapter.dart'; |
| 18 | import '../utils/attachment_helpers.dart'; |
| 19 | import '../utils/image_upload_utils.dart'; |
| 20 | import 'tools/adapters/openai_tool_adapter.dart'; |
| 21 | import 'tools/adapters/lm_studio_tool_adapter.dart'; |
| 22 | import 'tools/adapters/openrouter_tool_adapter.dart'; |
| 23 | import 'tools/adapters/ollama_tool_adapter.dart'; |
| 24 | import 'chat_api_error.dart'; |
| 25 | import 'chat_error_formatter.dart'; |
| 26 | import '../providers/chat_reasoning_providers.dart'; |
| 27 | |
| 28 | abstract class ChatService { |
| 29 | Stream<ChatResponse> sendMessage({ |
| 30 | required Server server, |
| 31 | required String modelId, |
| 32 | required List<Message> messages, |
| 33 | required ChatParameters params, |
| 34 | List<McpIntegration>? integrations, |
| 35 | List<ToolDefinition>? tools, |
| 36 | bool continueGeneration = false, |
| 37 | }); |
| 38 | |
| 39 | void cancelStream(); |
| 40 | |
| 41 | static ChatService forServer( |
| 42 | ServerType type, |
| 43 | Dio dio, { |
| 44 | OnDeviceGemmaService? onDeviceGemma, |
| 45 | bool imageCompressionEnabled = true, |
| 46 | ImageCompressionLevel imageCompressionLevel = ImageCompressionLevel.medium, |
| 47 | }) { |
| 48 | switch (type) { |
| 49 | case ServerType.lmStudio: |
| 50 | return LMStudioChatService( |
| 51 | dio, |
| 52 | imageCompressionEnabled: imageCompressionEnabled, |
| 53 | imageCompressionLevel: imageCompressionLevel, |
| 54 | ); |
| 55 | case ServerType.openAICompatible: |
| 56 | case ServerType.requesty: |
| 57 | return OpenAICompatibleChatService( |
| 58 | dio, |
| 59 | imageCompressionEnabled: imageCompressionEnabled, |
| 60 | imageCompressionLevel: imageCompressionLevel, |
| 61 | ); |
| 62 | case ServerType.ollama: |
| 63 | case ServerType.ollamaCloud: |
| 64 | return OllamaChatService( |
| 65 | dio, |
| 66 | imageCompressionEnabled: imageCompressionEnabled, |
| 67 | imageCompressionLevel: imageCompressionLevel, |
| 68 | ); |
| 69 | case ServerType.openRouter: |
| 70 | return OpenRouterChatService( |
| 71 | dio, |
| 72 | imageCompressionEnabled: imageCompressionEnabled, |
| 73 | imageCompressionLevel: imageCompressionLevel, |
| 74 | ); |
| 75 | case ServerType.onDevice: |
| 76 | if (onDeviceGemma == null) { |
| 77 | throw StateError( |
| 78 | 'OnDeviceGemmaService is required for onDevice server type', |
| 79 | ); |
| 80 | } |
| 81 | return OnDeviceChatService( |
| 82 | onDeviceGemma, |
| 83 | imageCompressionEnabled: imageCompressionEnabled, |
| 84 | imageCompressionLevel: imageCompressionLevel, |
| 85 | ); |
| 86 | } |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | class ChatResponse { |
| 91 | final ChatResponseType type; |
| 92 | final String? content; |
| 93 | final String? reasoningContent; |
| 94 | final ToolCallData? toolCall; |
| 95 | final InvalidToolCallData? invalidToolCall; |
| 96 | final ChatStats? stats; |
| 97 | |
| 98 | const ChatResponse({ |
| 99 | required this.type, |
| 100 | this.content, |
| 101 | this.reasoningContent, |
| 102 | this.toolCall, |
| 103 | this.invalidToolCall, |
| 104 | this.stats, |
| 105 | }); |
| 106 | } |
| 107 | |
| 108 | enum ChatResponseType { |
| 109 | message, |
| 110 | reasoning, |
| 111 | toolCall, |
| 112 | invalidToolCall, |
| 113 | done, |
| 114 | processing, |
| 115 | timeoutError, |
| 116 | error, |
| 117 | } |
| 118 | |
| 119 | class ToolCallData { |
| 120 | final String tool; |
| 121 | final Map<String, dynamic> arguments; |
| 122 | final String? output; |
| 123 | final ToolProviderInfo? providerInfo; |
| 124 | |
| 125 | const ToolCallData({ |
| 126 | required this.tool, |
| 127 | required this.arguments, |
| 128 | this.output, |
| 129 | this.providerInfo, |
| 130 | }); |
| 131 | } |
| 132 | |
| 133 | class ToolProviderInfo { |
| 134 | final String type; |
| 135 | final String? pluginId; |
| 136 | final String? serverLabel; |
| 137 | |
| 138 | const ToolProviderInfo({required this.type, this.pluginId, this.serverLabel}); |
| 139 | } |
| 140 | |
| 141 | class InvalidToolCallData { |
| 142 | final String reason; |
| 143 | final String? metadataType; |
| 144 | final String? toolName; |
| 145 | final Map<String, dynamic>? arguments; |
| 146 | final ToolProviderInfo? providerInfo; |
| 147 | |
| 148 | const InvalidToolCallData({ |
| 149 | required this.reason, |
| 150 | this.metadataType, |
| 151 | this.toolName, |
| 152 | this.arguments, |
| 153 | this.providerInfo, |
| 154 | }); |
| 155 | } |
| 156 | |
| 157 | class ChatStats { |
| 158 | final int inputTokens; |
| 159 | final int totalOutputTokens; |
| 160 | final int? reasoningOutputTokens; |
| 161 | final double? tokensPerSecond; |
| 162 | final double? timeToFirstTokenSeconds; |
| 163 | final double? modelLoadTimeSeconds; |
| 164 | |
| 165 | const ChatStats({ |
| 166 | required this.inputTokens, |
| 167 | required this.totalOutputTokens, |
| 168 | this.reasoningOutputTokens, |
| 169 | this.tokensPerSecond, |
| 170 | this.timeToFirstTokenSeconds, |
| 171 | this.modelLoadTimeSeconds, |
| 172 | }); |
| 173 | } |
| 174 | |
| 175 | class LMStudioChatService implements ChatService { |
| 176 | final Dio _dio; |
| 177 | final bool imageCompressionEnabled; |
| 178 | final ImageCompressionLevel imageCompressionLevel; |
| 179 | CancelToken? _cancelToken; |
| 180 | |
| 181 | LMStudioChatService( |
| 182 | this._dio, { |
| 183 | this.imageCompressionEnabled = true, |
| 184 | this.imageCompressionLevel = ImageCompressionLevel.medium, |
| 185 | }); |
| 186 | |
| 187 | @override |
| 188 | Stream<ChatResponse> sendMessage({ |
| 189 | required Server server, |
| 190 | required String modelId, |
| 191 | required List<Message> messages, |
| 192 | required ChatParameters params, |
| 193 | List<McpIntegration>? integrations, |
| 194 | List<ToolDefinition>? tools, |
| 195 | bool continueGeneration = false, |
| 196 | }) async* { |
| 197 | _cancelToken = CancelToken(); |
| 198 | |
| 199 | // LM Studio's native /api/v1/chat is stateless from the client's |
| 200 | // perspective: it does not accept an OpenAI-style `messages` array. We |
| 201 | // collapse history into a single string `input` for plain text chats, |
| 202 | // or into a typed-input array for chats with image attachments. The |
| 203 | // MCP integrations array carries the server-side tool definitions. |
| 204 | final useTypedInput = messages.any( |
| 205 | (m) => |
| 206 | m.attachmentPaths != null && |
| 207 | m.attachmentPaths!.any(AttachmentHelpers.isImagePath), |
| 208 | ); |
| 209 | |
| 210 | final String systemPrompt; |
| 211 | final input = useTypedInput |
| 212 | ? await _buildNativeTypedInput(messages, continueGeneration) |
| 213 | : await _buildNativeInputString(messages, continueGeneration); |
| 214 | systemPrompt = _extractSystemPrompt(messages) ?? params.systemPrompt ?? ''; |
| 215 | |
| 216 | final body = <String, dynamic>{ |
| 217 | 'model': modelId, |
| 218 | 'input': input, |
| 219 | if (systemPrompt.isNotEmpty) 'system_prompt': systemPrompt, |
| 220 | if (params.sendTemperature) 'temperature': params.temperature, |
| 221 | if (params.sendTopP) 'top_p': params.topP, |
| 222 | 'max_output_tokens': params.maxTokens, |
| 223 | 'stream': true, |
| 224 | }; |
| 225 | |
| 226 | if (params.topK != null) body['top_k'] = params.topK; |
| 227 | if (params.minP != null) body['min_p'] = params.minP; |
| 228 | if (params.repeatPenalty != null) { |
| 229 | body['repeat_penalty'] = params.repeatPenalty; |
| 230 | } |
| 231 | // LM Studio's native /api/v1/chat rejects the OpenAI-compat shape |
| 232 | // [_applyReasoningControl] writes (object `reasoning`, |
| 233 | // `reasoning_effort`, `think`, `enable_thinking`) with HTTP 400. It |
| 234 | // wants `reasoning` as a string. The exact allowed values are per-model |
| 235 | // (`capabilities.reasoning.allowed_options`, e.g. `["off","on"]` for |
| 236 | // binary models or `["low","medium","high","xhigh"]` for |
| 237 | // meta/muse-glimmer which rejects both 'on' and 'off'). We therefore |
| 238 | // resolve via [resolveLmStudioReasoningValue] and omit the key when it |
| 239 | // returns null (unsupported model, or disabled while `off` isn't |
| 240 | // allowed — omitting falls back to the server default instead of 400). |
| 241 | final reasoningValue = resolveLmStudioReasoningValue( |
| 242 | enabled: params.reasoningEnabled, |
| 243 | effort: params.reasoningEffort, |
| 244 | allowedOptions: params.reasoningAllowedOptions, |
| 245 | defaultOption: params.reasoningDefaultOption, |
| 246 | ); |
| 247 | if (reasoningValue != null) { |
| 248 | body['reasoning'] = reasoningValue; |
| 249 | } |
| 250 | if (params.contextLength > 0) { |
| 251 | body['context_length'] = params.contextLength; |
| 252 | } |
| 253 | |
| 254 | // MCP integrations are the only tool path supported by /api/v1/chat. |
| 255 | // `tools` (OpenAI-style function definitions) are NOT supported here — |
| 256 | // sending them would be silently ignored. The local tool registry is |
| 257 | // therefore only useful for non-LM-Studio server types on this build. |
| 258 | if (integrations != null && integrations.isNotEmpty) { |
| 259 | body['integrations'] = integrations.map((i) => i.toApiJson()).toList(); |
| 260 | } |
| 261 | |
| 262 | try { |
| 263 | final response = await _dio.post<ResponseBody>( |
| 264 | server.chatEndpoint, |
| 265 | data: body, |
| 266 | options: Options( |
| 267 | responseType: ResponseType.stream, |
| 268 | headers: { |
| 269 | 'Content-Type': 'application/json', |
| 270 | ...buildServerAuthHeaders(server), |
| 271 | }, |
| 272 | ), |
| 273 | cancelToken: _cancelToken, |
| 274 | ); |
| 275 | |
| 276 | final stream = response.data!.stream.cast<List<int>>().transform( |
| 277 | utf8.decoder, |
| 278 | ); |
| 279 | String buffer = ''; |
| 280 | String? currentEventName; |
| 281 | final lmStudioToolAdapter = LmStudioToolAdapter(); |
| 282 | |
| 283 | await for (final chunk in stream) { |
| 284 | buffer += chunk; |
| 285 | final lines = buffer.split('\n'); |
| 286 | buffer = lines.removeLast(); |
| 287 | |
| 288 | for (final line in lines) { |
| 289 | if (line.isEmpty) { |
| 290 | // Blank line = SSE event boundary. Flush any buffered event. |
| 291 | currentEventName = null; |
| 292 | continue; |
| 293 | } |
| 294 | if (line.startsWith(':')) { |
| 295 | // SSE comment / keep-alive. |
| 296 | continue; |
| 297 | } |
| 298 | if (line.startsWith('event:')) { |
| 299 | currentEventName = line.substring(6).trim(); |
| 300 | continue; |
| 301 | } |
| 302 | if (!line.startsWith('data:')) continue; |
| 303 | final data = line.substring(5).trim(); |
| 304 | if (data.isEmpty) continue; |
| 305 | |
| 306 | Map<String, dynamic>? json; |
| 307 | try { |
| 308 | json = jsonDecode(data) as Map<String, dynamic>; |
| 309 | } catch (e) { |
| 310 | Log.error('LMStudio chat parsing error: $e'); |
| 311 | continue; |
| 312 | } |
| 313 | |
| 314 | // Prefer the SSE `event:` line, fall back to the JSON `type` field. |
| 315 | final eventType = currentEventName ?? json['type'] as String? ?? ''; |
| 316 | |
| 317 | switch (eventType) { |
| 318 | case 'chat.start': |
| 319 | // Nothing to do — server is acknowledging the stream. |
| 320 | break; |
| 321 | case 'message.delta': |
| 322 | final content = json['content'] as String?; |
| 323 | if (content != null && content.isNotEmpty) { |
| 324 | yield ChatResponse( |
| 325 | type: ChatResponseType.message, |
| 326 | content: content, |
| 327 | ); |
| 328 | } |
| 329 | break; |
| 330 | case 'message.start': |
| 331 | case 'message.end': |
| 332 | case 'reasoning.start': |
| 333 | case 'reasoning.end': |
| 334 | // Boundaries only — content arrives in `*.delta`. |
| 335 | break; |
| 336 | case 'reasoning.delta': |
| 337 | final content = json['content'] as String?; |
| 338 | if (content != null && content.isNotEmpty) { |
| 339 | yield ChatResponse( |
| 340 | type: ChatResponseType.reasoning, |
| 341 | reasoningContent: content, |
| 342 | ); |
| 343 | } |
| 344 | break; |
| 345 | case 'tool_call.start': |
| 346 | case 'tool_call.arguments': |
| 347 | // Adapter aggregates; final completion is emitted below. |
| 348 | lmStudioToolAdapter.consumeEvent(eventType, json); |
| 349 | break; |
| 350 | case 'tool_call.success': |
| 351 | lmStudioToolAdapter.consumeEvent(eventType, json); |
| 352 | for (final call |
| 353 | in lmStudioToolAdapter.takeServerExecutedCalls()) { |
| 354 | yield ChatResponse( |
| 355 | type: ChatResponseType.toolCall, |
| 356 | toolCall: ToolCallData( |
| 357 | tool: call.name, |
| 358 | arguments: call.arguments, |
| 359 | output: call.output, |
| 360 | providerInfo: _providerInfoFromMap(call.providerInfo), |
| 361 | ), |
| 362 | ); |
| 363 | } |
| 364 | break; |
| 365 | case 'tool_call.failure': |
| 366 | lmStudioToolAdapter.consumeEvent(eventType, json); |
| 367 | for (final call |
| 368 | in lmStudioToolAdapter.takeServerExecutedCalls()) { |
| 369 | yield ChatResponse( |
| 370 | type: ChatResponseType.toolCall, |
| 371 | toolCall: ToolCallData( |
| 372 | tool: call.name, |
| 373 | arguments: call.arguments, |
| 374 | output: call.output, |
| 375 | providerInfo: _providerInfoFromMap(call.providerInfo), |
| 376 | ), |
| 377 | ); |
| 378 | } |
| 379 | break; |
| 380 | case 'error': |
| 381 | final error = json['error']; |
| 382 | final message = error is Map |
| 383 | ? (error['message'] as String? ?? error.toString()) |
| 384 | : (error?.toString() ?? 'LM Studio error'); |
| 385 | yield ChatResponse( |
| 386 | type: ChatResponseType.error, |
| 387 | content: ChatApiError( |
| 388 | message: message, |
| 389 | type: error is Map ? error['type'] as String? : null, |
| 390 | ).encode(), |
| 391 | ); |
| 392 | return; |
| 393 | case 'chat.end': |
| 394 | final result = json['result'] as Map<String, dynamic>?; |
| 395 | if (result != null) { |
| 396 | yield ChatResponse( |
| 397 | type: ChatResponseType.done, |
| 398 | stats: _extractStats(result), |
| 399 | ); |
| 400 | // Surface any tool_call / invalid_tool_call items present |
| 401 | // in the final aggregated output that we may have missed |
| 402 | // in the stream. |
| 403 | final output = result['output']; |
| 404 | if (output is List) { |
| 405 | for (final item in output) { |
| 406 | if (item is! Map) continue; |
| 407 | final m = Map<String, dynamic>.from(item); |
| 408 | final t = m['type']; |
| 409 | if (t == 'tool_call') { |
| 410 | yield ChatResponse( |
| 411 | type: ChatResponseType.toolCall, |
| 412 | toolCall: ToolCallData( |
| 413 | tool: m['tool'] as String? ?? '', |
| 414 | arguments: |
| 415 | (m['arguments'] as Map?) |
| 416 | ?.cast<String, dynamic>() ?? |
| 417 | {}, |
| 418 | output: m['output'] as String?, |
| 419 | providerInfo: _providerInfoFromMap( |
| 420 | (m['provider_info'] as Map?) |
| 421 | ?.cast<String, dynamic>(), |
| 422 | ), |
| 423 | ), |
| 424 | ); |
| 425 | } else if (t == 'invalid_tool_call') { |
| 426 | yield ChatResponse( |
| 427 | type: ChatResponseType.invalidToolCall, |
| 428 | invalidToolCall: InvalidToolCallData( |
| 429 | reason: m['reason'] as String? ?? '', |
| 430 | metadataType: (m['metadata'] is Map) |
| 431 | ? (m['metadata']['type'] as String?) |
| 432 | : null, |
| 433 | toolName: (m['metadata'] is Map) |
| 434 | ? (m['metadata']['tool_name'] as String?) |
| 435 | : null, |
| 436 | arguments: (m['arguments'] as Map?) |
| 437 | ?.cast<String, dynamic>(), |
| 438 | providerInfo: _providerInfoFromMap( |
| 439 | (m['provider_info'] as Map?) |
| 440 | ?.cast<String, dynamic>(), |
| 441 | ), |
| 442 | ), |
| 443 | ); |
| 444 | } |
| 445 | } |
| 446 | } |
| 447 | } else { |
| 448 | yield const ChatResponse(type: ChatResponseType.done); |
| 449 | } |
| 450 | return; |
| 451 | default: |
| 452 | // Model load / prompt processing events are surfaced only in |
| 453 | // the debug log — they don't carry user-visible content. |
| 454 | if (eventType.isNotEmpty) { |
| 455 | Log.debug('LMStudio stream event: $eventType'); |
| 456 | } |
| 457 | } |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | // Stream ended without an explicit chat.end — treat as a normal |
| 462 | // completion so the UI doesn't stay stuck in "processing". |
| 463 | yield const ChatResponse(type: ChatResponseType.done); |
| 464 | } catch (e) { |
| 465 | Log.error('LMStudio connection error: $e'); |
| 466 | yield ChatResponse( |
| 467 | type: ChatResponseType.error, |
| 468 | content: _handleChatError(e), |
| 469 | ); |
| 470 | return; |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | /// Builds the native `input` field as a single string, suitable for |
| 475 | /// plain-text chats. Conversation history is collapsed into a transcript |
| 476 | /// because `/api/v1/chat` does not accept an OpenAI-style `messages` |
| 477 | /// array; system prompt is hoisted to a top-level field by the caller. |
| 478 | Future<String> _buildNativeInputString( |
| 479 | List<Message> messages, |
| 480 | bool continueGeneration, |
| 481 | ) async { |
| 482 | final visible = <Message>[]; |
| 483 | for (final m in messages) { |
| 484 | if (m.role == MessageRole.system) continue; |
| 485 | visible.add(m); |
| 486 | } |
| 487 | // Drop trailing empty assistant (prefill) unless we're continuing it. |
| 488 | if (!continueGeneration && visible.isNotEmpty) { |
| 489 | while (visible.isNotEmpty && |
| 490 | visible.last.role == MessageRole.assistant && |
| 491 | visible.last.content.isEmpty) { |
| 492 | visible.removeLast(); |
| 493 | } |
| 494 | } |
| 495 | final buf = StringBuffer(); |
| 496 | for (final m in visible) { |
| 497 | final tag = switch (m.role) { |
| 498 | MessageRole.user => 'User', |
| 499 | MessageRole.assistant => 'Assistant', |
| 500 | MessageRole.tool => 'Tool', |
| 501 | MessageRole.system => 'System', |
| 502 | }; |
| 503 | final textContent = await _messageTextWithAttachments(m); |
| 504 | buf.writeln('$tag: $textContent'); |
| 505 | } |
| 506 | return buf.toString().trimRight(); |
| 507 | } |
| 508 | |
| 509 | /// Builds the native `input` field as an array of typed input items. |
| 510 | /// Used when at least one image attachment is present. |
| 511 | Future<List<Map<String, dynamic>>> _buildNativeTypedInput( |
| 512 | List<Message> messages, |
| 513 | bool continueGeneration, |
| 514 | ) async { |
| 515 | final items = <Map<String, dynamic>>[]; |
| 516 | final visible = <Message>[]; |
| 517 | for (final m in messages) { |
| 518 | if (m.role == MessageRole.system) continue; |
| 519 | visible.add(m); |
| 520 | } |
| 521 | if (!continueGeneration && visible.isNotEmpty) { |
| 522 | while (visible.isNotEmpty && |
| 523 | visible.last.role == MessageRole.assistant && |
| 524 | visible.last.content.isEmpty) { |
| 525 | visible.removeLast(); |
| 526 | } |
| 527 | } |
| 528 | for (final m in visible) { |
| 529 | final textContent = await _messageTextWithAttachments(m); |
| 530 | if (textContent.isNotEmpty) { |
| 531 | items.add({'type': 'text', 'content': textContent}); |
| 532 | } |
| 533 | items.addAll(await _imageItemsFor(m)); |
| 534 | } |
| 535 | return items; |
| 536 | } |
| 537 | |
| 538 | Future<List<Map<String, dynamic>>> _imageItemsFor(Message m) async { |
| 539 | final paths = m.attachmentPaths ?? const <String>[]; |
| 540 | final items = <Map<String, dynamic>>[]; |
| 541 | for (final path in paths) { |
| 542 | if (!AttachmentHelpers.isImagePath(path)) continue; |
| 543 | try { |
| 544 | final file = File(path); |
| 545 | if (!await file.exists()) continue; |
| 546 | final fileBytes = await ImageUploadUtils.prepareImageBytes( |
| 547 | file, |
| 548 | enabled: imageCompressionEnabled, |
| 549 | level: imageCompressionLevel, |
| 550 | ); |
| 551 | final base64Image = await Isolate.run(() { |
| 552 | try { |
| 553 | return base64Encode(fileBytes); |
| 554 | } catch (_) { |
| 555 | return null; |
| 556 | } |
| 557 | }); |
| 558 | if (base64Image != null) { |
| 559 | items.add({ |
| 560 | 'type': 'image', |
| 561 | 'data_url': 'data:image/png;base64,$base64Image', |
| 562 | }); |
| 563 | } |
| 564 | } catch (e) { |
| 565 | Log.error('Failed to read attachment for LMStudio native format: $e'); |
| 566 | } |
| 567 | } |
| 568 | return items; |
| 569 | } |
| 570 | |
| 571 | /// Builds the text portion of a typed input item, extracting supported |
| 572 | /// document contents into the message text so the model actually receives |
| 573 | /// them. Image attachments are handled separately by [_imageItemsFor]. |
| 574 | Future<String> _messageTextWithAttachments(Message m) async { |
| 575 | var textContent = m.content; |
| 576 | final paths = m.attachmentPaths ?? const <String>[]; |
| 577 | for (final path in paths) { |
| 578 | if (!AttachmentHelpers.isDocumentPath(path)) continue; |
| 579 | final text = await AttachmentHelpers.readDocumentFile(path); |
| 580 | if (text == null) continue; |
| 581 | final fileName = AttachmentHelpers.fileNameOf(path); |
| 582 | if (fileName.isEmpty) continue; |
| 583 | textContent = AttachmentHelpers.appendTextAttachment( |
| 584 | textContent, |
| 585 | fileName, |
| 586 | text, |
| 587 | ); |
| 588 | } |
| 589 | return textContent; |
| 590 | } |
| 591 | |
| 592 | String? _extractSystemPrompt(List<Message> messages) { |
| 593 | for (final m in messages) { |
| 594 | if (m.role == MessageRole.system) return m.content; |
| 595 | } |
| 596 | return null; |
| 597 | } |
| 598 | |
| 599 | ChatStats? _extractStats(Map<String, dynamic> result) { |
| 600 | final stats = result['stats'] as Map<String, dynamic>?; |
| 601 | if (stats == null) return null; |
| 602 | return ChatStats( |
| 603 | inputTokens: (stats['input_tokens'] as num?)?.toInt() ?? 0, |
| 604 | totalOutputTokens: (stats['total_output_tokens'] as num?)?.toInt() ?? 0, |
| 605 | reasoningOutputTokens: (stats['reasoning_output_tokens'] as num?) |
| 606 | ?.toInt(), |
| 607 | tokensPerSecond: (stats['tokens_per_second'] as num?)?.toDouble(), |
| 608 | timeToFirstTokenSeconds: (stats['time_to_first_token_seconds'] as num?) |
| 609 | ?.toDouble(), |
| 610 | modelLoadTimeSeconds: (stats['model_load_time_seconds'] as num?) |
| 611 | ?.toDouble(), |
| 612 | ); |
| 613 | } |
| 614 | |
| 615 | ToolProviderInfo? _providerInfoFromMap(Map<String, dynamic>? map) { |
| 616 | if (map == null) return null; |
| 617 | return ToolProviderInfo( |
| 618 | type: map['type'] as String? ?? '', |
| 619 | pluginId: map['plugin_id'] as String?, |
| 620 | serverLabel: map['server_label'] as String?, |
| 621 | ); |
| 622 | } |
| 623 | |
| 624 | @override |
| 625 | void cancelStream() { |
| 626 | _cancelToken?.cancel('User cancelled'); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | class OpenAICompatibleChatService implements ChatService { |
| 631 | final Dio _dio; |
| 632 | final bool imageCompressionEnabled; |
| 633 | final ImageCompressionLevel imageCompressionLevel; |
| 634 | CancelToken? _cancelToken; |
| 635 | Timer? _timeoutTimer; |
| 636 | |
| 637 | OpenAICompatibleChatService( |
| 638 | this._dio, { |
| 639 | this.imageCompressionEnabled = true, |
| 640 | this.imageCompressionLevel = ImageCompressionLevel.medium, |
| 641 | }); |
| 642 | |
| 643 | @override |
| 644 | Stream<ChatResponse> sendMessage({ |
| 645 | required Server server, |
| 646 | required String modelId, |
| 647 | required List<Message> messages, |
| 648 | required ChatParameters params, |
| 649 | List<McpIntegration>? integrations, |
| 650 | List<ToolDefinition>? tools, |
| 651 | bool continueGeneration = false, |
| 652 | }) async* { |
| 653 | _cancelToken = CancelToken(); |
| 654 | final toolAdapter = OpenAiToolAdapter(); |
| 655 | |
| 656 | final apiMessages = await Future.wait( |
| 657 | messages.map( |
| 658 | (m) => _buildOpenAiApiMessage( |
| 659 | m, |
| 660 | imageCompressionEnabled: imageCompressionEnabled, |
| 661 | imageCompressionLevel: imageCompressionLevel, |
| 662 | ), |
| 663 | ), |
| 664 | ); |
| 665 | // Strip trailing empty assistant prefill (or normalise it when |
| 666 | // continuing generation) to avoid "prefill incompatible with |
| 667 | // enable_thinking" errors from servers running thinking models. |
| 668 | _normalizeAssistantPrefill(apiMessages, continueGeneration); |
| 669 | |
| 670 | final body = <String, dynamic>{ |
| 671 | 'model': modelId, |
| 672 | 'messages': apiMessages, |
| 673 | if (params.sendTemperature) 'temperature': params.temperature, |
| 674 | if (params.sendTopP) 'top_p': params.topP, |
| 675 | 'max_tokens': params.maxTokens, |
| 676 | 'stream': true, |
| 677 | }; |
| 678 | _applyOpenAICompatibleReasoningControl(body, params); |
| 679 | |
| 680 | if (integrations != null && integrations.isNotEmpty) { |
| 681 | body['integrations'] = integrations.map((i) => i.toApiJson()).toList(); |
| 682 | } |
| 683 | |
| 684 | if (tools != null && tools.isNotEmpty) { |
| 685 | final toolsPayload = toolAdapter.buildToolDefinitionPayload(tools); |
| 686 | if (toolsPayload.containsKey('tools')) { |
| 687 | body['tools'] = toolsPayload['tools']; |
| 688 | } |
| 689 | } |
| 690 | |
| 691 | Log.debug( |
| 692 | 'OpenAICompatible: Sending request to ${server.chatEndpoint} with model: $modelId', |
| 693 | ); |
| 694 | |
| 695 | try { |
| 696 | final response = await _dio.post<ResponseBody>( |
| 697 | server.chatEndpoint, |
| 698 | data: body, |
| 699 | options: Options( |
| 700 | responseType: ResponseType.stream, |
| 701 | headers: { |
| 702 | 'Content-Type': 'application/json', |
| 703 | ...buildServerAuthHeaders(server), |
| 704 | }, |
| 705 | ), |
| 706 | cancelToken: _cancelToken, |
| 707 | ); |
| 708 | |
| 709 | final stream = response.data!.stream.cast<List<int>>().transform( |
| 710 | utf8.decoder, |
| 711 | ); |
| 712 | String buffer = ''; |
| 713 | DateTime? lastContentReceived; |
| 714 | int emptyDeltaCount = 0; |
| 715 | int logCounter = 0; |
| 716 | bool timeoutTriggered = false; |
| 717 | |
| 718 | // Start timeout timer |
| 719 | _timeoutTimer = Timer(const Duration(seconds: 45), () { |
| 720 | timeoutTriggered = true; |
| 721 | }); |
| 722 | |
| 723 | Log.debug('OpenAICompatible: Starting to receive stream...'); |
| 724 | |
| 725 | try { |
| 726 | await for (final chunk in stream) { |
| 727 | // Check if timeout was triggered |
| 728 | if (timeoutTriggered && lastContentReceived == null) { |
| 729 | yield const ChatResponse( |
| 730 | type: ChatResponseType.timeoutError, |
| 731 | content: |
| 732 | 'Model is taking too long to respond. This may happen with free tier models.', |
| 733 | ); |
| 734 | return; |
| 735 | } |
| 736 | |
| 737 | buffer += chunk; |
| 738 | final lines = buffer.split('\n'); |
| 739 | buffer = lines.removeLast(); |
| 740 | |
| 741 | for (final line in lines) { |
| 742 | logCounter++; |
| 743 | if (logCounter % 10 == 0) { |
| 744 | Log.debug('OpenAICompatible SSE line #$logCounter'); |
| 745 | } |
| 746 | |
| 747 | if (line.startsWith('data: ')) { |
| 748 | final data = line.substring(6); |
| 749 | if (data == '[DONE]') { |
| 750 | _timeoutTimer?.cancel(); |
| 751 | Log.debug('OpenAICompatible: Received [DONE]'); |
| 752 | for (final call in toolAdapter.takeCompletedCalls()) { |
| 753 | yield ChatResponse( |
| 754 | type: ChatResponseType.toolCall, |
| 755 | toolCall: ToolCallData( |
| 756 | tool: call.name, |
| 757 | arguments: call.arguments, |
| 758 | ), |
| 759 | ); |
| 760 | } |
| 761 | yield const ChatResponse(type: ChatResponseType.done); |
| 762 | return; |
| 763 | } |
| 764 | try { |
| 765 | final json = jsonDecode(data) as Map<String, dynamic>; |
| 766 | toolAdapter.consumeDynamicChunk(json); |
| 767 | |
| 768 | // Check for errors in the response |
| 769 | if (json['error'] != null) { |
| 770 | _timeoutTimer?.cancel(); |
| 771 | final errorContent = _formatApiErrorContent(json['error']); |
| 772 | Log.error('OpenAICompatible mid-stream error: $errorContent'); |
| 773 | yield ChatResponse( |
| 774 | type: ChatResponseType.error, |
| 775 | content: errorContent, |
| 776 | ); |
| 777 | return; |
| 778 | } |
| 779 | |
| 780 | final choices = json['choices'] as List<dynamic>?; |
| 781 | if (choices == null || choices.isEmpty) { |
| 782 | continue; |
| 783 | } |
| 784 | |
| 785 | final firstChoice = choices[0] as Map<String, dynamic>?; |
| 786 | if (firstChoice == null) continue; |
| 787 | |
| 788 | final delta = firstChoice['delta']; |
| 789 | if (delta == null || delta is! Map<String, dynamic>) continue; |
| 790 | |
| 791 | final content = (delta['content'] ?? delta['text']) as String?; |
| 792 | final reasoning = |
| 793 | (delta['reasoning'] ?? delta['reasoning_content']) |
| 794 | as String?; |
| 795 | final refusal = delta['refusal'] as String?; |
| 796 | |
| 797 | // Check if content is empty/null (processing but not outputting) |
| 798 | final hasContent = content != null && content.isNotEmpty; |
| 799 | final hasReasoning = reasoning != null && reasoning.isNotEmpty; |
| 800 | final hasRefusal = refusal != null && refusal.isNotEmpty; |
| 801 | |
| 802 | if (!hasContent && !hasReasoning && !hasRefusal) { |
| 803 | emptyDeltaCount++; |
| 804 | if (emptyDeltaCount % 10 == 0) { |
| 805 | yield const ChatResponse(type: ChatResponseType.processing); |
| 806 | } |
| 807 | } else { |
| 808 | emptyDeltaCount = 0; |
| 809 | lastContentReceived = DateTime.now(); |
| 810 | |
| 811 | if (hasRefusal) { |
| 812 | yield ChatResponse( |
| 813 | type: ChatResponseType.error, |
| 814 | content: 'Refusal: $refusal', |
| 815 | ); |
| 816 | return; |
| 817 | } |
| 818 | if (hasContent) { |
| 819 | yield ChatResponse( |
| 820 | type: ChatResponseType.message, |
| 821 | content: content, |
| 822 | ); |
| 823 | } |
| 824 | if (hasReasoning) { |
| 825 | yield ChatResponse( |
| 826 | type: ChatResponseType.reasoning, |
| 827 | reasoningContent: reasoning, |
| 828 | ); |
| 829 | } |
| 830 | } |
| 831 | } catch (e) { |
| 832 | Log.error('OpenAICompatible parsing error: $e'); |
| 833 | } |
| 834 | } |
| 835 | } |
| 836 | } |
| 837 | } finally { |
| 838 | _timeoutTimer?.cancel(); |
| 839 | } |
| 840 | |
| 841 | // Flush any tool calls accumulated without a [DONE] marker (e.g. |
| 842 | // connection drop or server crash before sending [DONE]). |
| 843 | for (final call in toolAdapter.takeCompletedCalls()) { |
| 844 | yield ChatResponse( |
| 845 | type: ChatResponseType.toolCall, |
| 846 | toolCall: ToolCallData(tool: call.name, arguments: call.arguments), |
| 847 | ); |
| 848 | } |
| 849 | } catch (e) { |
| 850 | Log.error('OpenAICompatible connection error: $e'); |
| 851 | yield ChatResponse( |
| 852 | type: ChatResponseType.error, |
| 853 | content: _handleChatError(e), |
| 854 | ); |
| 855 | return; |
| 856 | } finally { |
| 857 | _timeoutTimer?.cancel(); |
| 858 | } |
| 859 | |
| 860 | yield const ChatResponse(type: ChatResponseType.done); |
| 861 | Log.debug('OpenAICompatible: Stream ended'); |
| 862 | } |
| 863 | |
| 864 | @override |
| 865 | void cancelStream() { |
| 866 | _timeoutTimer?.cancel(); |
| 867 | _cancelToken?.cancel('User cancelled'); |
| 868 | } |
| 869 | } |
| 870 | |
| 871 | String _formatOllamaErrorContent(dynamic error) { |
| 872 | if (error is String) { |
| 873 | return ChatErrorFormatter.formatBody(jsonEncode({'error': error})); |
| 874 | } |
| 875 | if (error is Map) { |
| 876 | return ChatErrorFormatter.formatBody(jsonEncode({'error': error})); |
| 877 | } |
| 878 | return ChatApiError(message: error.toString()).encode(); |
| 879 | } |
| 880 | |
| 881 | Future<String> _handleOllamaChatError(DioException e) async { |
| 882 | return ChatErrorFormatter.formatDioException( |
| 883 | e, |
| 884 | readBody: (data) async { |
| 885 | if (data is ResponseBody) { |
| 886 | return data.stream.cast<List<int>>().transform(utf8.decoder).join(); |
| 887 | } |
| 888 | if (data is String) return data; |
| 889 | return ''; |
| 890 | }, |
| 891 | ); |
| 892 | } |
| 893 | |
| 894 | String? _extractOllamaErrorMessage(dynamic data) { |
| 895 | if (data == null) return null; |
| 896 | final raw = data is String ? data : jsonEncode(data); |
| 897 | if (raw.trim().isEmpty) return null; |
| 898 | return ChatErrorFormatter.formatBody(raw); |
| 899 | } |
| 900 | |
| 901 | class OllamaChatService implements ChatService { |
| 902 | final Dio _dio; |
| 903 | final bool imageCompressionEnabled; |
| 904 | final ImageCompressionLevel imageCompressionLevel; |
| 905 | CancelToken? _cancelToken; |
| 906 | |
| 907 | OllamaChatService( |
| 908 | this._dio, { |
| 909 | this.imageCompressionEnabled = true, |
| 910 | this.imageCompressionLevel = ImageCompressionLevel.medium, |
| 911 | }); |
| 912 | |
| 913 | @override |
| 914 | Stream<ChatResponse> sendMessage({ |
| 915 | required Server server, |
| 916 | required String modelId, |
| 917 | required List<Message> messages, |
| 918 | required ChatParameters params, |
| 919 | List<McpIntegration>? integrations, |
| 920 | List<ToolDefinition>? tools, |
| 921 | bool continueGeneration = false, |
| 922 | }) async* { |
| 923 | _cancelToken = CancelToken(); |
| 924 | final toolAdapter = OllamaToolAdapter(); |
| 925 | final contentDecoder = OllamaReasoningDecoder(); |
| 926 | var hasNativeThinking = false; |
| 927 | |
| 928 | final apiMessages = <Map<String, dynamic>>[]; |
| 929 | for (final message in messages) { |
| 930 | apiMessages.add(await _messageToOllamaMapWithImages(message)); |
| 931 | } |
| 932 | // Strip the trailing empty assistant placeholder created for streaming. |
| 933 | // Ollama's Qwen 3.5 parser treats any trailing assistant turn as an open |
| 934 | // prefill; an empty one misclassifies the generated continuation as |
| 935 | // `message.thinking` (see ollama/ollama#17187 and localmind#78). |
| 936 | _normalizeAssistantPrefill(apiMessages, continueGeneration); |
| 937 | |
| 938 | final body = <String, dynamic>{ |
| 939 | 'model': modelId, |
| 940 | 'messages': apiMessages, |
| 941 | 'stream': true, |
| 942 | 'options': { |
| 943 | if (params.sendTemperature) 'temperature': params.temperature, |
| 944 | if (params.sendTopP) 'top_p': params.topP, |
| 945 | 'num_predict': params.maxTokens, |
| 946 | }, |
| 947 | }; |
| 948 | _applyOllamaReasoningControl(body, params); |
| 949 | |
| 950 | if (tools != null && tools.isNotEmpty) { |
| 951 | final toolsPayload = toolAdapter.buildToolDefinitionPayload(tools); |
| 952 | if (toolsPayload.containsKey('tools')) { |
| 953 | body['tools'] = toolsPayload['tools']; |
| 954 | } |
| 955 | } |
| 956 | |
| 957 | final imageCount = apiMessages.fold<int>(0, (count, message) { |
| 958 | final images = message['images']; |
| 959 | return count + (images is List ? images.length : 0); |
| 960 | }); |
| 961 | Log.debug( |
| 962 | 'Ollama request: endpoint=${server.chatEndpoint}, model=$modelId, ' |
| 963 | 'messages=${apiMessages.length}, images=$imageCount, ' |
| 964 | 'tools=${tools?.length ?? 0}', |
| 965 | ); |
| 966 | |
| 967 | try { |
| 968 | final response = await _dio.post<ResponseBody>( |
| 969 | server.chatEndpoint, |
| 970 | data: body, |
| 971 | options: Options( |
| 972 | responseType: ResponseType.stream, |
| 973 | // Ollama returns useful JSON error bodies for 4xx responses. Keep |
| 974 | // the stream available so the actual error reaches the UI. |
| 975 | validateStatus: (status) => status != null, |
| 976 | headers: { |
| 977 | 'Content-Type': 'application/json', |
| 978 | ...buildServerAuthHeaders(server), |
| 979 | }, |
| 980 | ), |
| 981 | cancelToken: _cancelToken, |
| 982 | ); |
| 983 | |
| 984 | final responseBody = response.data!; |
| 985 | final statusCode = response.statusCode ?? 0; |
| 986 | if (statusCode < 200 || statusCode >= 300) { |
| 987 | final rawError = await responseBody.stream |
| 988 | .cast<List<int>>() |
| 989 | .transform(utf8.decoder) |
| 990 | .join(); |
| 991 | final parsedError = _extractOllamaErrorMessage(rawError); |
| 992 | Log.error('Ollama HTTP $statusCode: ${parsedError ?? rawError}'); |
| 993 | yield ChatResponse( |
| 994 | type: ChatResponseType.error, |
| 995 | content: |
| 996 | parsedError ?? |
| 997 | (rawError.trim().isNotEmpty |
| 998 | ? rawError.trim() |
| 999 | : 'Ollama request failed (HTTP $statusCode).'), |
| 1000 | ); |
| 1001 | return; |
| 1002 | } |
| 1003 | |
| 1004 | final lines = responseBody.stream |
| 1005 | .cast<List<int>>() |
| 1006 | .transform(utf8.decoder) |
| 1007 | .transform(const LineSplitter()); |
| 1008 | |
| 1009 | await for (final line in lines) { |
| 1010 | if (line.isNotEmpty) { |
| 1011 | try { |
| 1012 | final json = jsonDecode(line) as Map<String, dynamic>; |
| 1013 | toolAdapter.consumeDynamicChunk(json); |
| 1014 | |
| 1015 | // Ollama surfaces mid-stream errors as a top-level `error` |
| 1016 | // field, often alongside HTTP 200. Surface them so the UI |
| 1017 | // shows something more specific than the generic |
| 1018 | // "unknown error" fallback when — for instance — the |
| 1019 | // selected model doesn't actually support images. |
| 1020 | final errorField = json['error']; |
| 1021 | if (errorField != null) { |
| 1022 | yield ChatResponse( |
| 1023 | type: ChatResponseType.error, |
| 1024 | content: _formatOllamaErrorContent(errorField), |
| 1025 | ); |
| 1026 | return; |
| 1027 | } |
| 1028 | |
| 1029 | final message = json['message']; |
| 1030 | if (message != null && message is Map<String, dynamic>) { |
| 1031 | final content = message['content'] as String?; |
| 1032 | final thinking = message['thinking'] as String?; |
| 1033 | hasNativeThinking = hasNativeThinking || thinking != null; |
| 1034 | |
| 1035 | if (thinking != null && thinking.isNotEmpty) { |
| 1036 | yield ChatResponse( |
| 1037 | type: ChatResponseType.reasoning, |
| 1038 | reasoningContent: thinking, |
| 1039 | ); |
| 1040 | } |
| 1041 | if (content != null && content.isNotEmpty) { |
| 1042 | for (final part in contentDecoder.add( |
| 1043 | content, |
| 1044 | nativeThinking: hasNativeThinking, |
| 1045 | )) { |
| 1046 | yield ChatResponse( |
| 1047 | type: part.isReasoning |
| 1048 | ? ChatResponseType.reasoning |
| 1049 | : ChatResponseType.message, |
| 1050 | content: part.isReasoning ? null : part.text, |
| 1051 | reasoningContent: part.isReasoning ? part.text : null, |
| 1052 | ); |
| 1053 | } |
| 1054 | } |
| 1055 | } |
| 1056 | if (json['done'] == true) { |
| 1057 | for (final part in contentDecoder.flush()) { |
| 1058 | yield ChatResponse( |
| 1059 | type: part.isReasoning |
| 1060 | ? ChatResponseType.reasoning |
| 1061 | : ChatResponseType.message, |
| 1062 | content: part.isReasoning ? null : part.text, |
| 1063 | reasoningContent: part.isReasoning ? part.text : null, |
| 1064 | ); |
| 1065 | } |
| 1066 | |
| 1067 | for (final call in toolAdapter.takeCompletedCalls()) { |
| 1068 | yield ChatResponse( |
| 1069 | type: ChatResponseType.toolCall, |
| 1070 | toolCall: ToolCallData( |
| 1071 | tool: call.name, |
| 1072 | arguments: call.arguments, |
| 1073 | ), |
| 1074 | ); |
| 1075 | } |
| 1076 | yield const ChatResponse(type: ChatResponseType.done); |
| 1077 | return; |
| 1078 | } |
| 1079 | } catch (e) { |
| 1080 | Log.error('Ollama chunk parsing error: $e'); |
| 1081 | } |
| 1082 | } |
| 1083 | } |
| 1084 | } catch (e) { |
| 1085 | Log.error('Ollama connection error: $e'); |
| 1086 | final content = e is DioException |
| 1087 | ? await _handleOllamaChatError(e) |
| 1088 | : _handleChatError(e); |
| 1089 | yield ChatResponse(type: ChatResponseType.error, content: content); |
| 1090 | return; |
| 1091 | } |
| 1092 | for (final part in contentDecoder.flush()) { |
| 1093 | yield ChatResponse( |
| 1094 | type: part.isReasoning |
| 1095 | ? ChatResponseType.reasoning |
| 1096 | : ChatResponseType.message, |
| 1097 | content: part.isReasoning ? null : part.text, |
| 1098 | reasoningContent: part.isReasoning ? part.text : null, |
| 1099 | ); |
| 1100 | } |
| 1101 | yield const ChatResponse(type: ChatResponseType.done); |
| 1102 | } |
| 1103 | |
| 1104 | Future<Map<String, dynamic>> _messageToOllamaMapWithImages(Message m) async { |
| 1105 | var textContent = m.content; |
| 1106 | final images = <String>[]; |
| 1107 | |
| 1108 | for (final path in m.attachmentPaths ?? const <String>[]) { |
| 1109 | if (AttachmentHelpers.isDocumentPath(path)) { |
| 1110 | final text = await AttachmentHelpers.readDocumentFile(path); |
| 1111 | if (text != null) { |
| 1112 | textContent = AttachmentHelpers.appendTextAttachment( |
| 1113 | textContent, |
| 1114 | AttachmentHelpers.fileNameOf(path), |
| 1115 | text, |
| 1116 | ); |
| 1117 | } |
| 1118 | continue; |
| 1119 | } |
| 1120 | |
| 1121 | if (!AttachmentHelpers.isImagePath(path)) continue; |
| 1122 | |
| 1123 | try { |
| 1124 | final file = File(path); |
| 1125 | if (!await file.exists()) continue; |
| 1126 | |
| 1127 | final fileBytes = await ImageUploadUtils.prepareImageBytes( |
| 1128 | file, |
| 1129 | enabled: imageCompressionEnabled, |
| 1130 | level: imageCompressionLevel, |
| 1131 | ); |
| 1132 | final base64Image = await Isolate.run(() { |
| 1133 | try { |
| 1134 | return base64Encode(fileBytes); |
| 1135 | } catch (_) { |
| 1136 | return null; |
| 1137 | } |
| 1138 | }); |
| 1139 | |
| 1140 | if (base64Image != null) images.add(base64Image); |
| 1141 | } catch (e) { |
| 1142 | Log.error('Failed to read attachment for Ollama format: $e'); |
| 1143 | } |
| 1144 | } |
| 1145 | |
| 1146 | final map = <String, dynamic>{ |
| 1147 | 'role': _roleToString(m.role), |
| 1148 | 'content': textContent, |
| 1149 | }; |
| 1150 | if (images.isNotEmpty) map['images'] = images; |
| 1151 | |
| 1152 | if (m.role == MessageRole.tool && m.toolCallId != null) { |
| 1153 | map['tool_call_id'] = m.toolCallId; |
| 1154 | } |
| 1155 | if (m.role == MessageRole.assistant && |
| 1156 | m.toolCalls != null && |
| 1157 | m.toolCalls!.isNotEmpty) { |
| 1158 | map['tool_calls'] = m.toolCalls!.map((tc) { |
| 1159 | return { |
| 1160 | 'id': tc.id, |
| 1161 | 'type': 'function', |
| 1162 | 'function': { |
| 1163 | 'name': tc.toolName, |
| 1164 | 'arguments': jsonEncode(tc.arguments), |
| 1165 | }, |
| 1166 | }; |
| 1167 | }).toList(); |
| 1168 | } |
| 1169 | return map; |
| 1170 | } |
| 1171 | |
| 1172 | @override |
| 1173 | void cancelStream() { |
| 1174 | _cancelToken?.cancel('User cancelled'); |
| 1175 | } |
| 1176 | } |
| 1177 | |
| 1178 | class OpenRouterChatService implements ChatService { |
| 1179 | final Dio _dio; |
| 1180 | final bool imageCompressionEnabled; |
| 1181 | final ImageCompressionLevel imageCompressionLevel; |
| 1182 | CancelToken? _cancelToken; |
| 1183 | Timer? _timeoutTimer; |
| 1184 | |
| 1185 | OpenRouterChatService( |
| 1186 | this._dio, { |
| 1187 | this.imageCompressionEnabled = true, |
| 1188 | this.imageCompressionLevel = ImageCompressionLevel.medium, |
| 1189 | }); |
| 1190 | |
| 1191 | @override |
| 1192 | Stream<ChatResponse> sendMessage({ |
| 1193 | required Server server, |
| 1194 | required String modelId, |
| 1195 | required List<Message> messages, |
| 1196 | required ChatParameters params, |
| 1197 | List<McpIntegration>? integrations, |
| 1198 | List<ToolDefinition>? tools, |
| 1199 | bool continueGeneration = false, |
| 1200 | }) async* { |
| 1201 | _cancelToken = CancelToken(); |
| 1202 | final toolAdapter = OpenRouterToolAdapter(); |
| 1203 | |
| 1204 | final apiMessages = await Future.wait( |
| 1205 | messages.map( |
| 1206 | (m) => _buildOpenAiApiMessage( |
| 1207 | m, |
| 1208 | imageCompressionEnabled: imageCompressionEnabled, |
| 1209 | imageCompressionLevel: imageCompressionLevel, |
| 1210 | ), |
| 1211 | ), |
| 1212 | ); |
| 1213 | _normalizeAssistantPrefill(apiMessages, continueGeneration); |
| 1214 | |
| 1215 | final body = <String, dynamic>{ |
| 1216 | 'model': modelId, |
| 1217 | 'messages': apiMessages, |
| 1218 | if (params.sendTemperature) 'temperature': params.temperature, |
| 1219 | if (params.sendTopP) 'top_p': params.topP, |
| 1220 | 'max_tokens': params.maxTokens, |
| 1221 | 'stream': true, |
| 1222 | }; |
| 1223 | _applyOpenRouterReasoningControl(body, params); |
| 1224 | |
| 1225 | if (tools != null && tools.isNotEmpty) { |
| 1226 | final toolsPayload = toolAdapter.buildToolDefinitionPayload(tools); |
| 1227 | if (toolsPayload.containsKey('tools')) { |
| 1228 | body['tools'] = toolsPayload['tools']; |
| 1229 | } |
| 1230 | } |
| 1231 | |
| 1232 | Log.debug( |
| 1233 | 'OpenRouter: Sending request to ${server.chatEndpoint} with model: $modelId', |
| 1234 | ); |
| 1235 | |
| 1236 | try { |
| 1237 | final response = await _dio.post<ResponseBody>( |
| 1238 | server.chatEndpoint, |
| 1239 | data: body, |
| 1240 | options: Options( |
| 1241 | responseType: ResponseType.stream, |
| 1242 | headers: { |
| 1243 | 'Content-Type': 'application/json', |
| 1244 | ...buildServerAuthHeaders(server), |
| 1245 | 'HTTP-Referer': 'https://localmind.app', |
| 1246 | 'X-Title': 'LocalMind', |
| 1247 | }, |
| 1248 | ), |
| 1249 | cancelToken: _cancelToken, |
| 1250 | ); |
| 1251 | |
| 1252 | final stream = response.data!.stream.cast<List<int>>().transform( |
| 1253 | utf8.decoder, |
| 1254 | ); |
| 1255 | String buffer = ''; |
| 1256 | DateTime? lastContentReceived; |
| 1257 | int emptyDeltaCount = 0; |
| 1258 | int logCounter = 0; |
| 1259 | bool timeoutTriggered = false; |
| 1260 | |
| 1261 | // Start timeout timer |
| 1262 | _timeoutTimer = Timer(const Duration(seconds: 45), () { |
| 1263 | timeoutTriggered = true; |
| 1264 | }); |
| 1265 | |
| 1266 | Log.debug('OpenRouter: Starting to receive stream...'); |
| 1267 | |
| 1268 | try { |
| 1269 | await for (final chunk in stream) { |
| 1270 | // Check if timeout was triggered |
| 1271 | if (timeoutTriggered && lastContentReceived == null) { |
| 1272 | yield const ChatResponse( |
| 1273 | type: ChatResponseType.timeoutError, |
| 1274 | content: |
| 1275 | 'Model is taking too long to respond. This may happen with free tier models.', |
| 1276 | ); |
| 1277 | return; |
| 1278 | } |
| 1279 | |
| 1280 | buffer += chunk; |
| 1281 | final lines = buffer.split('\n'); |
| 1282 | buffer = lines.removeLast(); |
| 1283 | |
| 1284 | for (final line in lines) { |
| 1285 | logCounter++; |
| 1286 | if (logCounter % 10 == 0) { |
| 1287 | Log.debug('OpenRouter SSE line #$logCounter: $line'); |
| 1288 | } |
| 1289 | |
| 1290 | if (line.startsWith('data: ')) { |
| 1291 | final data = line.substring(6); |
| 1292 | if (data == '[DONE]') { |
| 1293 | _timeoutTimer?.cancel(); |
| 1294 | Log.debug('OpenRouter: Received [DONE]'); |
| 1295 | for (final call in toolAdapter.takeCompletedCalls()) { |
| 1296 | yield ChatResponse( |
| 1297 | type: ChatResponseType.toolCall, |
| 1298 | toolCall: ToolCallData( |
| 1299 | tool: call.name, |
| 1300 | arguments: call.arguments, |
| 1301 | ), |
| 1302 | ); |
| 1303 | } |
| 1304 | yield const ChatResponse(type: ChatResponseType.done); |
| 1305 | return; |
| 1306 | } |
| 1307 | try { |
| 1308 | final json = jsonDecode(data) as Map<String, dynamic>; |
| 1309 | toolAdapter.consumeDynamicChunk(json); |
| 1310 | |
| 1311 | // Check for errors in the response |
| 1312 | if (json['error'] != null) { |
| 1313 | _timeoutTimer?.cancel(); |
| 1314 | final errorContent = _formatApiErrorContent(json['error']); |
| 1315 | Log.error('OpenRouter mid-stream error: $errorContent'); |
| 1316 | yield ChatResponse( |
| 1317 | type: ChatResponseType.error, |
| 1318 | content: errorContent, |
| 1319 | ); |
| 1320 | return; |
| 1321 | } |
| 1322 | |
| 1323 | final choices = json['choices'] as List<dynamic>?; |
| 1324 | if (choices == null || choices.isEmpty) { |
| 1325 | continue; |
| 1326 | } |
| 1327 | |
| 1328 | final firstChoice = choices[0] as Map<String, dynamic>?; |
| 1329 | if (firstChoice == null) continue; |
| 1330 | |
| 1331 | final delta = firstChoice['delta']; |
| 1332 | if (delta == null || delta is! Map<String, dynamic>) continue; |
| 1333 | |
| 1334 | final content = (delta['content'] ?? delta['text']) as String?; |
| 1335 | final reasoning = |
| 1336 | (delta['reasoning'] ?? delta['reasoning_content']) |
| 1337 | as String?; |
| 1338 | final refusal = delta['refusal'] as String?; |
| 1339 | |
| 1340 | // Check if content is empty/null (processing but not outputting) |
| 1341 | final hasContent = content != null && content.isNotEmpty; |
| 1342 | final hasReasoning = reasoning != null && reasoning.isNotEmpty; |
| 1343 | final hasRefusal = refusal != null && refusal.isNotEmpty; |
| 1344 | |
| 1345 | if (!hasContent && !hasReasoning && !hasRefusal) { |
| 1346 | emptyDeltaCount++; |
| 1347 | if (emptyDeltaCount % 10 == 0) { |
| 1348 | yield const ChatResponse(type: ChatResponseType.processing); |
| 1349 | } |
| 1350 | } else { |
| 1351 | emptyDeltaCount = 0; |
| 1352 | lastContentReceived = DateTime.now(); |
| 1353 | |
| 1354 | if (hasRefusal) { |
| 1355 | yield ChatResponse( |
| 1356 | type: ChatResponseType.error, |
| 1357 | content: 'Refusal: $refusal', |
| 1358 | ); |
| 1359 | return; |
| 1360 | } |
| 1361 | if (hasContent) { |
| 1362 | yield ChatResponse( |
| 1363 | type: ChatResponseType.message, |
| 1364 | content: content, |
| 1365 | ); |
| 1366 | } |
| 1367 | if (hasReasoning) { |
| 1368 | yield ChatResponse( |
| 1369 | type: ChatResponseType.reasoning, |
| 1370 | reasoningContent: reasoning, |
| 1371 | ); |
| 1372 | } |
| 1373 | } |
| 1374 | } catch (e) { |
| 1375 | Log.error('OpenRouter parsing error: $e'); |
| 1376 | } |
| 1377 | } else if (line.startsWith(': ')) { |
| 1378 | // SSE comment (keep-alive) |
| 1379 | if (logCounter % 50 == 0) { |
| 1380 | Log.debug('OpenRouter SSE comment: $line'); |
| 1381 | } |
| 1382 | } |
| 1383 | } |
| 1384 | } |
| 1385 | } finally { |
| 1386 | _timeoutTimer?.cancel(); |
| 1387 | } |
| 1388 | |
| 1389 | // Flush any tool calls accumulated without a [DONE] marker (e.g. |
| 1390 | // connection drop or server crash before sending [DONE]). |
| 1391 | for (final call in toolAdapter.takeCompletedCalls()) { |
| 1392 | yield ChatResponse( |
| 1393 | type: ChatResponseType.toolCall, |
| 1394 | toolCall: ToolCallData(tool: call.name, arguments: call.arguments), |
| 1395 | ); |
| 1396 | } |
| 1397 | } catch (e) { |
| 1398 | Log.error('OpenRouter connection error: $e'); |
| 1399 | yield ChatResponse( |
| 1400 | type: ChatResponseType.error, |
| 1401 | content: _handleChatError(e), |
| 1402 | ); |
| 1403 | return; |
| 1404 | } finally { |
| 1405 | _timeoutTimer?.cancel(); |
| 1406 | } |
| 1407 | |
| 1408 | yield const ChatResponse(type: ChatResponseType.done); |
| 1409 | Log.debug('OpenRouter: Stream ended'); |
| 1410 | } |
| 1411 | |
| 1412 | @override |
| 1413 | void cancelStream() { |
| 1414 | _timeoutTimer?.cancel(); |
| 1415 | _cancelToken?.cancel('User cancelled'); |
| 1416 | } |
| 1417 | } |
| 1418 | |
| 1419 | /// Applies the standard OpenAI-compatible `reasoning_effort` field. Avoid |
| 1420 | /// broadcasting provider-specific aliases because strict compatible servers |
| 1421 | /// reject unknown request keys. |
| 1422 | void _applyOpenAICompatibleReasoningControl( |
| 1423 | Map<String, dynamic> body, |
| 1424 | ChatParameters params, |
| 1425 | ) { |
| 1426 | if (params.reasoningEnabled == false) { |
| 1427 | body['reasoning_effort'] = 'none'; |
| 1428 | } else if (params.reasoningEnabled == true) { |
| 1429 | body['reasoning_effort'] = params.reasoningEffort.apiValue; |
| 1430 | } |
| 1431 | } |
| 1432 | |
| 1433 | /// OpenRouter defines a `reasoning` object. Use that object alone instead of |
| 1434 | /// mixing it with the shorthand `reasoning_effort` or unrelated local-server |
| 1435 | /// flags. |
| 1436 | void _applyOpenRouterReasoningControl( |
| 1437 | Map<String, dynamic> body, |
| 1438 | ChatParameters params, |
| 1439 | ) { |
| 1440 | if (params.reasoningEnabled == false) { |
| 1441 | body['reasoning'] = {'enabled': false}; |
| 1442 | } else if (params.reasoningEnabled == true) { |
| 1443 | body['reasoning'] = {'effort': params.reasoningEffort.apiValue}; |
| 1444 | } |
| 1445 | } |
| 1446 | |
| 1447 | /// Applies the Think toggle using only Ollama's native `/api/chat` `think` |
| 1448 | /// field (`true`/`false` or `"low"`/`"medium"`/`"high"`/`"max"`). Omits the |
| 1449 | /// key when the model doesn't support reasoning or can't run with thinking |
| 1450 | /// off, so strict servers and proxies never see unrelated `reasoning`, |
| 1451 | /// `reasoning_effort`, or `enable_thinking` keys. |
| 1452 | void _applyOllamaReasoningControl( |
| 1453 | Map<String, dynamic> body, |
| 1454 | ChatParameters params, |
| 1455 | ) { |
| 1456 | final think = resolveOllamaThinkValue( |
| 1457 | enabled: params.reasoningEnabled, |
| 1458 | effort: params.reasoningEffort, |
| 1459 | allowedOptions: params.reasoningAllowedOptions, |
| 1460 | defaultOption: params.reasoningDefaultOption, |
| 1461 | ); |
| 1462 | if (think != null) body['think'] = think; |
| 1463 | } |
| 1464 | |
| 1465 | String _formatApiErrorContent(dynamic error) { |
| 1466 | if (error is Map) { |
| 1467 | final map = Map<String, dynamic>.from(error); |
| 1468 | final parsed = ChatApiError.fromErrorMap(map); |
| 1469 | if (parsed != null) { |
| 1470 | // Use the unified formatter so the bubble benefits from humanising |
| 1471 | // and structured metadata even when the body is already shaped like |
| 1472 | // an OpenAI error. |
| 1473 | try { |
| 1474 | return ChatErrorFormatter.formatBody(jsonEncode(map)); |
| 1475 | } catch (_) { |
| 1476 | return parsed.encode(); |
| 1477 | } |
| 1478 | } |
| 1479 | } |
| 1480 | return ChatApiError(message: error.toString()).encode(); |
| 1481 | } |
| 1482 | |
| 1483 | String _handleChatError(dynamic e) { |
| 1484 | if (e is DioException) { |
| 1485 | if (e.response != null) { |
| 1486 | final status = e.response!.statusCode; |
| 1487 | final data = e.response!.data; |
| 1488 | if (data is String && data.trim().isNotEmpty) { |
| 1489 | return ChatErrorFormatter.formatBody(data, statusCode: status); |
| 1490 | } |
| 1491 | final parsed = ChatApiError.fromResponseBody(data); |
| 1492 | if (parsed != null) { |
| 1493 | return parsed.copyWithCodeIfMissing(status?.toString()).encode(); |
| 1494 | } |
| 1495 | // Let the unified formatter map known HTTP status codes. |
| 1496 | return ChatErrorFormatter.formatBody('', statusCode: status); |
| 1497 | } |
| 1498 | |
| 1499 | switch (e.type) { |
| 1500 | case DioExceptionType.connectionTimeout: |
| 1501 | case DioExceptionType.sendTimeout: |
| 1502 | case DioExceptionType.receiveTimeout: |
| 1503 | return const ChatApiError( |
| 1504 | message: 'The model took too long to respond.', |
| 1505 | type: 'timeout', |
| 1506 | ).encode(); |
| 1507 | case DioExceptionType.connectionError: |
| 1508 | return const ChatApiError( |
| 1509 | message: |
| 1510 | 'Could not reach the AI server. Confirm it is running and the address in Settings is correct.', |
| 1511 | type: 'connection_error', |
| 1512 | ).encode(); |
| 1513 | case DioExceptionType.cancel: |
| 1514 | return const ChatApiError( |
| 1515 | message: 'Request cancelled.', |
| 1516 | type: 'cancelled', |
| 1517 | ).encode(); |
| 1518 | default: |
| 1519 | return ChatApiError( |
| 1520 | message: e.message ?? 'Connection error.', |
| 1521 | type: e.type.name, |
| 1522 | ).encode(); |
| 1523 | } |
| 1524 | } |
| 1525 | return ChatApiError(message: e.toString()).encode(); |
| 1526 | } |
| 1527 | |
| 1528 | String _roleToString(MessageRole role) { |
| 1529 | switch (role) { |
| 1530 | case MessageRole.user: |
| 1531 | return 'user'; |
| 1532 | case MessageRole.assistant: |
| 1533 | return 'assistant'; |
| 1534 | case MessageRole.system: |
| 1535 | return 'system'; |
| 1536 | case MessageRole.tool: |
| 1537 | return 'tool'; |
| 1538 | } |
| 1539 | } |
| 1540 | |
| 1541 | /// Normalises the trailing assistant message in a chat `messages` array. |
| 1542 | /// When NOT continuing generation, drops trailing empty assistant prefill |
| 1543 | /// messages. Besides triggering "prefill incompatible with enable_thinking" |
| 1544 | /// errors on OpenAI-style thinking models, an empty trailing assistant turn |
| 1545 | /// makes Ollama's Qwen 3.5 parser misclassify the generated continuation as |
| 1546 | /// `message.thinking`. When continuing generation, keeps the trailing |
| 1547 | /// assistant prefill but ensures its `content` is a plain string (image |
| 1548 | /// content-part arrays are only valid on user messages, and |
| 1549 | /// [_buildOpenAiApiMessage] never produces them for assistants anyway). |
| 1550 | void _normalizeAssistantPrefill( |
| 1551 | List<Map<String, dynamic>> apiMessages, |
| 1552 | bool continueGeneration, |
| 1553 | ) { |
| 1554 | if (!continueGeneration) { |
| 1555 | while (apiMessages.isNotEmpty && |
| 1556 | apiMessages.last['role'] == 'assistant' && |
| 1557 | (apiMessages.last['content'] == null || |
| 1558 | (apiMessages.last['content'] is String && |
| 1559 | (apiMessages.last['content'] as String).isEmpty))) { |
| 1560 | apiMessages.removeLast(); |
| 1561 | } |
| 1562 | } else if (apiMessages.isNotEmpty && |
| 1563 | apiMessages.last['role'] == 'assistant') { |
| 1564 | final last = apiMessages.last; |
| 1565 | last['content'] = (last['content'] as String?) ?? ''; |
| 1566 | } |
| 1567 | } |
| 1568 | |
| 1569 | /// Builds the OpenAI-style request payload for a single [Message], including |
| 1570 | /// attachments. Supported documents are read and inlined into the text content |
| 1571 | /// (mirroring the Ollama path), and image attachments are emitted as OpenAI |
| 1572 | /// multimodal `image_url` content parts carrying base64 data URLs. |
| 1573 | /// |
| 1574 | /// Image parts are only emitted for user messages — OpenAI rejects content |
| 1575 | /// part arrays on other roles — so non-user messages keep a plain string |
| 1576 | /// `content` (which still receives inlined text attachments). |
| 1577 | Future<Map<String, dynamic>> _buildOpenAiApiMessage( |
| 1578 | Message m, { |
| 1579 | required bool imageCompressionEnabled, |
| 1580 | required ImageCompressionLevel imageCompressionLevel, |
| 1581 | }) async { |
| 1582 | final map = <String, dynamic>{ |
| 1583 | 'role': _roleToString(m.role), |
| 1584 | 'content': m.content, |
| 1585 | }; |
| 1586 | |
| 1587 | var textContent = m.content; |
| 1588 | final imageParts = <Map<String, dynamic>>[]; |
| 1589 | |
| 1590 | for (final path in m.attachmentPaths ?? const <String>[]) { |
| 1591 | if (AttachmentHelpers.isDocumentPath(path)) { |
| 1592 | final text = await AttachmentHelpers.readDocumentFile(path); |
| 1593 | if (text != null) { |
| 1594 | textContent = AttachmentHelpers.appendTextAttachment( |
| 1595 | textContent, |
| 1596 | AttachmentHelpers.fileNameOf(path), |
| 1597 | text, |
| 1598 | ); |
| 1599 | } |
| 1600 | continue; |
| 1601 | } |
| 1602 | |
| 1603 | if (!AttachmentHelpers.isImagePath(path)) continue; |
| 1604 | |
| 1605 | try { |
| 1606 | final file = File(path); |
| 1607 | if (!await file.exists()) continue; |
| 1608 | |
| 1609 | final fileBytes = await ImageUploadUtils.prepareImageBytes( |
| 1610 | file, |
| 1611 | enabled: imageCompressionEnabled, |
| 1612 | level: imageCompressionLevel, |
| 1613 | ); |
| 1614 | final base64Image = await Isolate.run(() { |
| 1615 | try { |
| 1616 | return base64Encode(fileBytes); |
| 1617 | } catch (_) { |
| 1618 | return null; |
| 1619 | } |
| 1620 | }); |
| 1621 | |
| 1622 | if (base64Image != null) { |
| 1623 | // Compression re-encodes resized images to PNG; detect that so the |
| 1624 | // data URL advertises the bytes actually being sent. |
| 1625 | final mimeType = _looksLikePng(fileBytes) |
| 1626 | ? 'image/png' |
| 1627 | : AttachmentHelpers.mimeTypeForImage(path); |
| 1628 | imageParts.add({ |
| 1629 | 'type': 'image_url', |
| 1630 | 'image_url': {'url': 'data:$mimeType;base64,$base64Image'}, |
| 1631 | }); |
| 1632 | } |
| 1633 | } catch (e) { |
| 1634 | Log.error('Failed to read attachment for OpenAI format: $e'); |
| 1635 | } |
| 1636 | } |
| 1637 | |
| 1638 | if (imageParts.isNotEmpty && m.role == MessageRole.user) { |
| 1639 | map['content'] = [ |
| 1640 | if (textContent.isNotEmpty) {'type': 'text', 'text': textContent}, |
| 1641 | ...imageParts, |
| 1642 | ]; |
| 1643 | } else if (textContent != m.content) { |
| 1644 | map['content'] = textContent; |
| 1645 | } |
| 1646 | |
| 1647 | if (m.role == MessageRole.tool && m.toolCallId != null) { |
| 1648 | map['tool_call_id'] = m.toolCallId; |
| 1649 | } |
| 1650 | if (m.role == MessageRole.assistant && |
| 1651 | m.toolCalls != null && |
| 1652 | m.toolCalls!.isNotEmpty) { |
| 1653 | map['tool_calls'] = m.toolCalls!.map((tc) { |
| 1654 | return { |
| 1655 | 'id': tc.id, |
| 1656 | 'type': 'function', |
| 1657 | 'function': { |
| 1658 | 'name': tc.toolName, |
| 1659 | 'arguments': jsonEncode(tc.arguments), |
| 1660 | }, |
| 1661 | }; |
| 1662 | }).toList(); |
| 1663 | } |
| 1664 | return map; |
| 1665 | } |
| 1666 | |
| 1667 | bool _looksLikePng(Uint8List bytes) { |
| 1668 | return bytes.length >= 4 && |
| 1669 | bytes[0] == 0x89 && |
| 1670 | bytes[1] == 0x50 && |
| 1671 | bytes[2] == 0x4e && |
| 1672 | bytes[3] == 0x47; |
| 1673 | } |
| 1674 | |
| 1675 | /// Returns the streaming-parsing adapter for a server type. |
| 1676 | /// |
| 1677 | /// OpenAI-compatible and Ollama servers use adapter-based tool call parsing |
| 1678 | /// from streaming chunks. LM Studio and on-device servers collect tool calls |
| 1679 | /// directly from [ChatResponseType.toolCall] events emitted during streaming |
| 1680 | /// and feed them to the loop via [ToolExecutionLoop.run]'s [preParsedCalls]. |
| 1681 | ToolTransportAdapter createAdapterForServerType(ServerType type) => |
| 1682 | switch (type) { |
| 1683 | ServerType.openAICompatible => OpenAiToolAdapter(), |
| 1684 | ServerType.openRouter => OpenRouterToolAdapter(), |
| 1685 | ServerType.requesty => OpenAiToolAdapter(), |
| 1686 | ServerType.lmStudio => OpenAiToolAdapter(), |
| 1687 | ServerType.onDevice => OpenAiToolAdapter(), |
| 1688 | ServerType.ollama => OllamaToolAdapter(), |
| 1689 | ServerType.ollamaCloud => OllamaToolAdapter(), |
| 1690 | }; |