Team E Final Changes

LearningLens2025/.gitignore

@@ -49,8 +49,12 @@ terraform/terraformout.txt
 # npm
 
 /lambda/code_eval/node_modules
-/lambda/gettoken/node_modules
-/lambda/gettoken/gettoken.zip
+/lambda/ai_log/node_modules
+/lambda/game_data/node_modules
+/lambda/reflections/node_modules
+/lambda/ai_log/ai_log.zip
 /lambda/code_eval/code_eval.zip
+/lambda/game_data/game_data.zip
+/lambda/reflections/reflections.zip
 # Any extra javascript files for lambda functions
 !lambda/code_eval/*.js

LearningLens2025/frontend/.example.env

@@ -13,4 +13,6 @@ deepseek_apiKey=
 GOOGLE_CLIENT_ID=
 AI_LOGGING_URL=
 CODE_EVAL_URL=
+GAME_URL=
+REFLECTIONS_URL=
 LOCAL_MODEL_DOWNLOAD_URL_PATH=

LearningLens2025/frontend/lib/Api/experimental/assistant/textbased_function_caller_view.dart

@@ -10,6 +10,8 @@ import 'package:learninglens_app/Api/llm/perplexity_api.dart';
 import 'package:learninglens_app/Api/lms/factory/lms_factory.dart';
 import 'package:learninglens_app/Controller/custom_appbar.dart';
 import 'package:learninglens_app/services/local_storage_service.dart';
+import 'package:learninglens_app/Api/llm/local_llm_service.dart'; // local llm
+import 'package:flutter/foundation.dart';
 
 class TextBasedFunctionCallerView extends StatefulWidget {
   const TextBasedFunctionCallerView({super.key});
@@ -30,6 +32,8 @@ class _TextBasedFunctionCallerViewState
   LlmType selectedLLM = LlmType.GROK; //default to chatgpt
   late LLM llm;
 
+  bool _localLlmAvail = !kIsWeb;
+
   @override
   void initState() {
     super.initState();
@@ -54,6 +58,8 @@ class _TextBasedFunctionCallerViewState
       aiModel = PerplexityLLM(LocalStorageService.getPerplexityKey());
     } else if (selectedLLM == LlmType.DEEPSEEK) {
       aiModel = DeepseekLLM(LocalStorageService.getDeepseekKey());
+    } else if (selectedLLM == LlmType.LOCAL) {
+      aiModel = LocalLLMService();
     } else {
       // default
       aiModel = OpenAiLLM(LocalStorageService.getOpenAIKey());
@@ -199,11 +205,18 @@ class _TextBasedFunctionCallerViewState
                     items: LlmType.values.map((LlmType llm) {
                       return DropdownMenuItem<LlmType>(
                         value: llm,
-                        enabled: LocalStorageService.userHasLlmKey(llm),
+                        enabled: (llm == LlmType.LOCAL &&
+                                LocalStorageService.getLocalLLMPath() != "" &&
+                                _localLlmAvail) ||
+                            LocalStorageService.userHasLlmKey(llm),
                         child: Text(
                           llm.displayName,
                           style: TextStyle(
-                            color: LocalStorageService.userHasLlmKey(llm)
+                            color: (llm == LlmType.LOCAL &&
+                                        LocalStorageService.getLocalLLMPath() !=
+                                            "" &&
+                                        _localLlmAvail) ||
+                                    LocalStorageService.userHasLlmKey(llm)
                                 ? Colors.black87
                                 : Colors.grey,
                           ),

LearningLens2025/frontend/lib/Api/experimental/assistant/textbased_llm_client.dart

@@ -3,6 +3,7 @@ import 'package:learninglens_app/Api/experimental/assistant/textbased_function_c
 import 'package:learninglens_app/Api/llm/llm_api_modules_base.dart';
 import 'package:learninglens_app/Api/llm/prompt_engine.dart';
 import 'package:learninglens_app/services/api_service.dart';
+import 'package:learninglens_app/Api/llm/local_llm_service.dart';
 
 //  Replicate the functionality used in chatgpt_client, but swap over to prompt engineering instead of the function caller.
 //  This code gears the development for the assistant to be more generic in terms of which llm is used rather than relying on
@@ -61,34 +62,45 @@ class TextBasedLLMClient {
     }
   }
 
-  /// Calls the OpenAI Chat Completion API with the entire conversation so far
+  /// Calls the OpenAI Chat Completion API with the entire conversation so far or call locl LLM.
   Future<String?> _callLLM(List<Map<String, String>> conversation) async {
-    final response = await ApiService().httpPost(
-      Uri.parse(llm.url),
-      headers: {
-        "Authorization": 'Bearer ${llm.apiKey}',
-        "Content-Type": "application/json",
-      },
-      body: jsonEncode({
-        "model": llm.model,
-        "messages": conversation.map((m) {
-          return {"role": m["role"], "content": m["content"]};
-        }).toList(),
-        "temperature": 0.7,
-        "top_p": 0.9,
-      }),
-    );
-
-    if (response.statusCode != 200) {
-      return "Error from LLM: ${response.statusCode} => ${response.body}";
-    }
+    if (llm.apiKey == "") {
+      final response = await LocalLLMService().runModel2(conversation.map((m) {
+        return {"role": m["role"], "content": m["content"]};
+      }).toList());
 
-    final jsonData = jsonDecode(response.body);
-    if (jsonData["choices"] == null || jsonData["choices"].isEmpty) {
-      return null;
-    }
+      if (response == "") {
+        return null;
+      }
+      return response;
+    } else {
+      final response = await ApiService().httpPost(
+        Uri.parse(llm.url),
+        headers: {
+          "Authorization": 'Bearer ${llm.apiKey}',
+          "Content-Type": "application/json",
+        },
+        body: jsonEncode({
+          "model": llm.model,
+          "messages": conversation.map((m) {
+            return {"role": m["role"], "content": m["content"]};
+          }).toList(),
+          "temperature": 0.7,
+          "top_p": 0.9,
+        }),
+      );
+
+      if (response.statusCode != 200) {
+        return "Error from LLM: ${response.statusCode} => ${response.body}";
+      }
 
-    return jsonData["choices"][0]["message"]["content"];
+      final jsonData = jsonDecode(response.body);
+      if (jsonData["choices"] == null || jsonData["choices"].isEmpty) {
+        return null;
+      }
+
+      return jsonData["choices"][0]["message"]["content"];
+    }
   }
 
   /// Parses a "CALL functionName(...)" string, calls the local function, returns its result

LearningLens2025/frontend/lib/Api/llm/DeepSeek_api.dart

@@ -352,7 +352,6 @@ $inputText
       'max_tokens': maxOutputTokens,
       'stream': true, // critical for SSE
     };
-    print("Streaming body: $body");
     final client = http.Client();
     try {
       final req = http.Request('POST', parsedUrl)

LearningLens2025/frontend/lib/Api/llm/local_llm_service.dart

@@ -158,7 +158,7 @@ class LocalLLMService implements LLM {
 
     int requestId = await fllamaChat(request, (response, responseJson, done) {
       responseBuff.add(response);
-      //print(response);
+      print(response);
       if (done) {
         _runningRequestId = null;
         finalResponse = _getFinalResponse(responseBuff);
@@ -197,6 +197,7 @@ class LocalLLMService implements LLM {
     return "ERROR: Local LLM failed to produce a response";
   }
 
+  // method to run model with previous contexts
   Future<String> runModel2(
     List<Map<String, dynamic>>? context, {
     int? maxTokenSet,
@@ -343,8 +344,8 @@ class LocalLLMService implements LLM {
         builder: (context) => AlertDialog(
           title: const Text('⚠️Warning:'),
           content: Text(
-            'The currently loaded Local LLM does not consistently generate a valid XML needed for rubric geneartion.\n'
-            'Proceeding could result in invalid XML output error.\n\n'
+            'The currently loaded Local LLM does not consistently generate a valid XML or json files used in Learning Management System.\n'
+            'Proceeding could result in invalid XML or json output error.\n\n'
             'The recommended model for this task is 7B or higher reasoning models (Qwen).\n'
             'Do you want to continue anyway?',
           ),
@@ -409,6 +410,7 @@ class LocalLLMService implements LLM {
     return await runModel2(context);
   }
 
+  // unused in local LLM;
   @override
   Future<String> generate(String prompt) {
     // TODO: implement generate
@@ -498,43 +500,139 @@ class LocalLLMService implements LLM {
   }
 
   Future<List<String>> fetchModelKeys() async {
-    final url = Uri.parse(
-      'https://raw.githubusercontent.com/ssung13/SWEN670F2025/main/models.csv',
-    );
+    final downloadUrl = LocalStorageService.getLocalLLMDownloadURLPath();
+    if (downloadUrl != '') {
+      final url = Uri.parse(
+        downloadUrl,
+      );
 
-    final List<String> modelKeys = [];
+      final List<String> modelKeys = [];
 
-    try {
-      final response = await http.get(url);
-      if (response.statusCode == 200) {
-        final csvContent = response.body;
-
-        for (var line in LineSplitter.split(csvContent)) {
-          if (line.trim().isEmpty) continue;
-          final parts = line.split(',');
-          if (parts.isNotEmpty) {
-            modelKeys.add(parts[0].trim());
+      try {
+        final response = await http.get(url);
+        if (response.statusCode == 200) {
+          final csvContent = response.body;
+
+          for (var line in LineSplitter.split(csvContent)) {
+            if (line.trim().isEmpty) continue;
+            final parts = line.split(',');
+            if (parts.isNotEmpty) {
+              modelKeys.add(parts[0].trim());
+            }
           }
+        } else {
+          print('Failed to fetch CSV: ${response.statusCode}');
         }
-      } else {
-        print('Failed to fetch CSV: ${response.statusCode}');
+      } catch (e) {
+        print('Error fetching CSV: $e');
       }
-    } catch (e) {
-      print('Error fetching CSV: $e');
-    }
 
-    return modelKeys;
+      return modelKeys;
+    } else {
+      return [];
+    }
   }
 
+  // method for outputting stream.
   @override
   Stream<String> chatStream(
       {List<Map<String, dynamic>>? context,
       String? prompt,
-      double temperature = 0.7,
-      double topP = 1.0,
-      double frequencyPenalty = 0.0,
-      double presencePenalty = 0.0}) {
-    // TODO: implement chatStream
-    throw UnimplementedError();
+      double? temperature,
+      double? topP,
+      double? frequencyPenalty,
+      double? presencePenalty}) async* {
+    try {
+      model = LocalStorageService.getLocalLLMPath();
+
+      List<Message>? convertedContext = context?.map((entry) {
+        final role = Role.values.firstWhere((r) => r.name == entry['role']);
+        final message = entry['content'] as String;
+        return Message(role, message);
+      }).toList();
+
+      double temperatureInput = temperature ?? 0.7;
+
+      // model id for the web build (DO NOT use WEB for now / WIP).
+      String mlcModelId = MlcModelId.qwen05b;
+
+      // model path for the desktop build.
+      String mmprojPath = "";
+
+      final request = OpenAiRequest(
+        tools: [
+          if (_tool != null)
+            Tool(
+              name: _tool!.name,
+              jsonSchema: _tool!.parametersAsString,
+            ),
+        ],
+        maxTokens: maxOutputTokens.round(),
+        messages: convertedContext!,
+        numGpuLayers: 99,
+        /* this seems to have no adverse effects in environments w/o GPU support, ex. Android and web */
+        modelPath: kIsWeb ? mlcModelId : model,
+        mmprojPath: mmprojPath,
+        frequencyPenalty: 0.0,
+        // Don't use below 1.1, LLMs without a repeat penalty
+        // will repeat the same token.
+        presencePenalty: 1.1,
+        topP: _topP,
+        // 22.9s for 249 input tokens with 20K context for SmolLM3.
+        // 22.9s for 249 input tokens with 4K context for SmolLM3.
+        contextSize: contextSize,
+        // Don't use 0.0, some models will repeat
+        // the same token.
+        temperature: temperatureInput,
+        logger: (log) {
+          if (log.contains('<unused')) {
+            // 25-03-11: Added because Gemma 3 outputs so many that it
+            // can break the VS Code log viewer.
+            return;
+          }
+          if (log.contains('ggml_')) {
+            // 25-03-11: Added because that's the biggest clutter-er left
+            // when trying to get logs reduced down to compare Gemma 3 working vs.
+            // not-working cases.
+            return;
+          }
+          // ignore: avoid_print
+          print('[llama.cpp] $log');
+        },
+      );
+
+      // This is required, otherwise the fllama will crash if using fllamaChat
+      await fllamaChatTemplateGet(model);
+
+      final controller = StreamController<String>();
+      String previous = '';
+      int requestId = await fllamaChat(request, (response, responseJson, done) {
+        if (!controller.isClosed) {
+          if (response.isNotEmpty) {
+            if (response.startsWith(previous)) {
+              controller.add(response.substring(previous.length));
+              previous = response;
+            } else {
+              // In case it doesn't follow pattern (rare)
+              controller.add(response);
+              previous = response;
+            }
+          }
+          if (done) {
+            controller.close();
+            _runningRequestId = null;
+          }
+        }
+      });
+
+      _runningRequestId = requestId;
+      yield* controller.stream;
+
+      while (_runningRequestId != null) {
+        await Future.delayed(const Duration(milliseconds: 100));
+      }
+    } catch (e) {
+      print(e);
+    }
   }
 }