feat: implement background LLM pattern analysis on Insights screen

Adds a prompt builder that serialises 7-day mood/activity data, wires InsightNotifier to call the on-device LLM directly (no user interaction), caches results by prompt hash, and renders the analysis as a passive card. Includes loading, empty (<2 entries), and error states with a force-refresh button.
2026-04-26 14:37:37 +10:00
parent d1b18c5296
commit ea412d6fe9
4 changed files with 322 additions and 48 deletions
@@ -84,6 +84,22 @@ class MoodDao extends DatabaseAccessor<AppDatabase> with _$MoodDaoMixin {
        .getSingleOrNull()) != null;
  }

+  // Entries with their linked activities — used by LLM prompt builder
+  Future<List<MoodEntryWithActivities>> getEntriesWithActivitiesInRange(
+      DateTime start, DateTime end) async {
+    final entries = await getEntriesInRange(start, end);
+    return Future.wait(entries.map((e) async {
+      final links = await (select(entryActivities)
+            ..where((t) => t.entryId.equals(e.id)))
+          .get();
+      final actIds = links.map((l) => l.activityId).toList();
+      final acts = actIds.isEmpty
+          ? <Activity>[]
+          : await (select(activities)..where((t) => t.id.isIn(actIds))).get();
+      return MoodEntryWithActivities(entry: e, activities: acts, tags: []);
+    }));
+  }
+
  // Average mood per day — used by calendar heatmap
  Future<Map<DateTime, double>> getDailyAverages(
      DateTime start, DateTime end) async {
@@ -0,0 +1,56 @@
+import '../database/daos/mood_dao.dart';
+
+abstract class PromptBuilder {
+  static String buildWeeklyPrompt(List<MoodEntryWithActivities> entries) {
+    final lines = entries.map(_entryLine).join('\n');
+    return '''
+You are a personal wellbeing assistant analysing a private mood journal.
+Analyse the following data and identify 2–3 patterns or correlations across mood, sleep, energy, positivity, and self-worth.
+End with one specific, actionable suggestion. Be concise (4–6 sentences total). Do not greet the user.
+
+$lines''';
+  }
+
+  static String _entryLine(MoodEntryWithActivities e) {
+    final dt = e.entry.timestamp;
+    final date =
+        '${dt.year}-${dt.month.toString().padLeft(2, '0')}-${dt.day.toString().padLeft(2, '0')}';
+    final time =
+        '${dt.hour.toString().padLeft(2, '0')}:${dt.minute.toString().padLeft(2, '0')}';
+
+    final parts = <String>[
+      '$date $time',
+      'Mood ${e.entry.moodLevel}/5',
+    ];
+
+    if (e.entry.sleepMinutes != null) {
+      final h = e.entry.sleepMinutes! ~/ 60;
+      final m = (e.entry.sleepMinutes! % 60).toString().padLeft(2, '0');
+      parts.add('Sleep ${h}h$m');
+    }
+    if (e.entry.energyLevel != null) {
+      parts.add('Energy ${_scale(e.entry.energyLevel!, _energyLabels)}');
+    }
+    if (e.entry.positivityLevel != null) {
+      parts.add('Positivity ${_scale(e.entry.positivityLevel!, _positivityLabels)}');
+    }
+    if (e.entry.selfWorthLevel != null) {
+      parts.add('Self-worth ${_scale(e.entry.selfWorthLevel!, _selfWorthLabels)}');
+    }
+    if (e.activities.isNotEmpty) {
+      parts.add('Activities: ${e.activities.map((a) => a.name).join(', ')}');
+    }
+    if (e.entry.note != null && e.entry.note!.isNotEmpty) {
+      parts.add('Note: ${e.entry.note}');
+    }
+
+    return parts.join(' | ');
+  }
+
+  static String _scale(int level, List<String> labels) =>
+      labels[(level - 1).clamp(0, labels.length - 1)];
+
+  static const _energyLabels = ['very low', 'low', 'ok', 'high', 'very high'];
+  static const _positivityLabels = ['very negative', 'negative', 'neutral', 'positive', 'very positive'];
+  static const _selfWorthLabels = ['very low', 'low', 'moderate', 'high', 'very high'];
+}
@@ -3,71 +3,87 @@ import 'package:crypto/crypto.dart';
 import 'package:drift/drift.dart';
 import 'package:flutter_riverpod/flutter_riverpod.dart';
 import '../database/app_database.dart';
+import '../llm/prompt_builder.dart';
 import 'database_provider.dart';
+import 'llm_provider.dart';

 // ── Insight type enum ─────────────────────────────────────────────────────────

 enum InsightType { weekly, monthly, query }

-// ── Cached insight lookup ─────────────────────────────────────────────────────
-
-final cachedInsightProvider =
-    FutureProvider.autoDispose.family<LlmInsight?, String>((ref, promptHash) {
-  return ref.watch(insightDaoProvider).getCached(promptHash);
-});
-
-// ── Insight notifier — triggers generation and caches results ─────────────────
+// ── Insight notifier — fetches data, builds prompt, calls LLM, caches result ──

 class InsightNotifier extends Notifier<AsyncValue<String?>> {
  @override
  AsyncValue<String?> build() => const AsyncData(null);

-  String _hashPrompt(String prompt) =>
-      sha256.convert(utf8.encode(prompt)).toString();
-
  Future<void> generateInsight({
-    required String prompt,
-    required InsightType type,
-    DateTime? rangeStart,
-    DateTime? rangeEnd,
+    InsightType type = InsightType.weekly,
+    bool forceRefresh = false,
  }) async {
    state = const AsyncLoading();

-    final hash = _hashPrompt(prompt);
-    final dao = ref.read(insightDaoProvider);
+    try {
+      // Ensure model is initialised before generating
+      await ref.read(llmReadyProvider.future);
+      if (!ref.mounted) return;

-    final cached = await dao.getCached(hash);
-    if (!ref.mounted) return;
-    if (cached != null) {
-      state = AsyncData(cached.responseText);
-      return;
+      final now = DateTime.now();
+      final end = DateTime(now.year, now.month, now.day, 23, 59, 59);
+      final start = end.subtract(const Duration(days: 6));
+
+      final dao = ref.read(moodDaoProvider);
+      final entries = await dao.getEntriesWithActivitiesInRange(start, end);
+      if (!ref.mounted) return;
+
+      if (entries.length < 2) {
+        state = const AsyncData(null);
+        return;
+      }
+
+      final prompt = PromptBuilder.buildWeeklyPrompt(entries);
+      final hash = _hashPrompt(prompt);
+
+      if (!forceRefresh) {
+        final cached = await ref.read(insightDaoProvider).getCached(hash);
+        if (!ref.mounted) return;
+        if (cached != null) {
+          state = AsyncData(cached.responseText);
+          return;
+        }
+      }
+
+      final llm = ref.read(llmServiceProvider);
+      final buffer = StringBuffer();
+      await for (final token in llm.generateStream(prompt)) {
+        if (!ref.mounted) return;
+        buffer.write(token);
+        state = AsyncData(buffer.toString());
+      }
+
+      if (!ref.mounted) return;
+      final result = buffer.toString().trim();
+      await ref.read(insightDaoProvider).upsertInsight(
+            LlmInsightsCompanion.insert(
+              type: type.name,
+              promptHash: hash,
+              responseText: result,
+              rangeStart: Value(start),
+              rangeEnd: Value(end),
+            ),
+          );
+      if (!ref.mounted) return;
+      state = AsyncData(result);
+    } catch (e, st) {
+      if (!ref.mounted) return;
+      state = AsyncError(e, st);
    }
-
-    // Generation happens in LlmNotifier (see llm_provider.dart)
-    state = const AsyncData(null);
-  }
-
-  Future<void> saveInsight({
-    required String promptHash,
-    required String responseText,
-    required InsightType type,
-    DateTime? rangeStart,
-    DateTime? rangeEnd,
-  }) async {
-    await ref.read(insightDaoProvider).upsertInsight(
-          LlmInsightsCompanion.insert(
-            type: type.name,
-            promptHash: promptHash,
-            responseText: responseText,
-            rangeStart: Value(rangeStart),
-            rangeEnd: Value(rangeEnd),
-          ),
-        );
-    if (!ref.mounted) return;
-    state = AsyncData(responseText);
  }

  void clear() => state = const AsyncData(null);
+
+  static String _hashPrompt(String prompt) =>
+      sha256.convert(utf8.encode(prompt)).toString();
 }

 final insightNotifierProvider =