[mob][photos] Add text embeddings cache service (#7130)

## Description

Add text embeddings cache service to prevent recomputes for:
- Memories
- Magic cache

## Tests

Tested in debug mode on my pixel phone.

mobile/apps/photos/CLAUDE.md

@@ -196,6 +196,11 @@ lib/
 - Ensure documentation reflects the current implementation
 - Update examples in specs if behavior changes
 
+### 5. Database Methods - BEST PRACTICE
+**Prioritize readability in database methods**
+- For small result sets (e.g., 1-2 stale entries), prefer filtering in Dart for cleaner, more readable code
+- For large datasets, use SQL WHERE clauses for performance - they're much more efficient in SQLite
+
 ## Important Notes
 
 - Large service files (some 70k+ lines) - consider file context when editing

mobile/apps/photos/lib/core/error-reporting/super_logging.dart

@@ -230,7 +230,6 @@ class SuperLogging {
     } catch (e) {
       $.warning("Failed to initialize log viewer: $e");
     }
-  }
 
     if (appConfig.body == null) return;
 
@@ -313,14 +312,14 @@ class SuperLogging {
     saveLogString(str, rec.error);
     // Hook for external log viewer (if available)
     // This allows the log_viewer package to capture logs without creating a dependency
-    if(_logViewerCallback != null) {
-    try {
-      if (_logViewerCallback != null) {
-        _logViewerCallback!(rec, config.prefix);
+    if (_logViewerCallback != null) {
+      try {
+        if (_logViewerCallback != null) {
+          _logViewerCallback!(rec, config.prefix);
+        }
+      } catch (_) {
+        // Silently ignore any errors from the log viewer
       }
-    } catch (_) {
-      // Silently ignore any errors from the log viewer
-    }
     }
   }
 

mobile/apps/photos/lib/db/ml/db.dart

@@ -62,6 +62,7 @@ class MLDataDB with SqlDbBase implements IMLDataDB<int> {
     createClipEmbeddingsTable,
     createFileDataTable,
     createFaceCacheTable,
+    createTextEmbeddingsCacheTable,
   ];
 
   // only have a single app-wide reference to the database
@@ -1429,6 +1430,56 @@ class MLDataDB with SqlDbBase implements IMLDataDB<int> {
     Bus.instance.fire(EmbeddingUpdatedEvent());
   }
 
+  /// WARNING: don't confuse this with [putClip]. If you're not sure, use [putClip]
+  Future<void> putRepeatedTextEmbeddingCache(
+    String query,
+    List<double> embedding,
+  ) async {
+    final db = await asyncDB;
+    await db.execute(
+      'INSERT OR REPLACE INTO $textEmbeddingsCacheTable '
+      '($textQueryColumn, $embeddingColumn, $mlVersionColumn, $createdAtColumn) '
+      'VALUES (?, ?, ?, ?)',
+      [
+        query,
+        Float32List.fromList(embedding).buffer.asUint8List(),
+        clipMlVersion,
+        DateTime.now().millisecondsSinceEpoch,
+      ],
+    );
+  }
+
+  /// WARNING: don't confuse this with [getAllClipVectors]. If you're not sure, use [getAllClipVectors]
+  Future<List<double>?> getRepeatedTextEmbeddingCache(String query) async {
+    final db = await asyncDB;
+    final results = await db.getAll(
+      'SELECT $embeddingColumn, $mlVersionColumn, $createdAtColumn '
+      'FROM $textEmbeddingsCacheTable '
+      'WHERE $textQueryColumn = ?',
+      [query],
+    );
+
+    if (results.isEmpty) return null;
+
+    final threeMonthsAgo =
+        DateTime.now().millisecondsSinceEpoch - (90 * 24 * 60 * 60 * 1000);
+
+    // Find first valid entry
+    for (final result in results) {
+      if (result[mlVersionColumn] == clipMlVersion &&
+          result[createdAtColumn] as int > threeMonthsAgo) {
+        return Float32List.view((result[embeddingColumn] as Uint8List).buffer);
+      }
+    }
+
+    // No valid entry found, clean up
+    await db.execute(
+      'DELETE FROM $textEmbeddingsCacheTable WHERE $textQueryColumn = ?',
+      [query],
+    );
+    return null;
+  }
+
   @override
   Future<void> deleteClipEmbeddings(List<int> fileIDs) async {
     final db = await instance.asyncDB;

mobile/apps/photos/lib/db/ml/schema.dart

@@ -16,6 +16,8 @@ const mlVersionColumn = 'ml_version';
 const personIdColumn = 'person_id';
 const clusterIDColumn = 'cluster_id';
 const personOrClusterIdColumn = 'person_or_cluster_id';
+const textQueryColumn = 'text_query';
+const createdAtColumn = 'created_at';
 
 const createFacesTable = '''CREATE TABLE IF NOT EXISTS $facesTable (
   $fileIDColumn	INTEGER NOT NULL,
@@ -137,3 +139,18 @@ CREATE TABLE IF NOT EXISTS $faceCacheTable (
 ''';
 
 const deleteFaceCacheTable = 'DELETE FROM $faceCacheTable';
+
+// ## TEXT EMBEDDINGS CACHE TABLE
+const textEmbeddingsCacheTable = 'text_embeddings_cache';
+
+const createTextEmbeddingsCacheTable = '''
+CREATE TABLE IF NOT EXISTS $textEmbeddingsCacheTable (
+  $textQueryColumn TEXT NOT NULL,
+  $embeddingColumn BLOB NOT NULL,
+  $mlVersionColumn INTEGER NOT NULL,
+  $createdAtColumn INTEGER NOT NULL,
+  PRIMARY KEY ($textQueryColumn)
+);
+''';
+
+const deleteTextEmbeddingsCacheTable = 'DELETE FROM $textEmbeddingsCacheTable';

mobile/apps/photos/lib/main.dart

@@ -1,4 +1,4 @@
-  import 'dart:async';
+import 'dart:async';
 import 'dart:io';
 
 import "package:adaptive_theme/adaptive_theme.dart";

mobile/apps/photos/lib/service_locator.dart

@@ -19,6 +19,7 @@ import "package:photos/services/smart_albums_service.dart";
 import "package:photos/services/smart_memories_service.dart";
 import "package:photos/services/storage_bonus_service.dart";
 import "package:photos/services/sync/trash_sync_service.dart";
+import "package:photos/services/text_embeddings_cache_service.dart";
 import "package:photos/services/update_service.dart";
 import "package:photos/utils/local_settings.dart";
 import "package:shared_preferences/shared_preferences.dart";
@@ -136,6 +137,12 @@ SmartMemoriesService get smartMemoriesService {
   return _smartMemoriesService!;
 }
 
+TextEmbeddingsCacheService? _textEmbeddingsCacheService;
+TextEmbeddingsCacheService get textEmbeddingsCacheService {
+  _textEmbeddingsCacheService ??= TextEmbeddingsCacheService.instance;
+  return _textEmbeddingsCacheService!;
+}
+
 BillingService? _billingService;
 BillingService get billingService {
   _billingService ??= BillingService(

mobile/apps/photos/lib/services/machine_learning/semantic_search/semantic_search_service.dart

@@ -193,15 +193,22 @@ class SemanticSearchService {
     return results;
   }
 
-  Future<Map<String, List<int>>> getMatchingFileIDs(
+  /// Get matching file IDs for common repeated queries like smart memories and magic cache.
+  /// WARNING: Use this method carefully - it uses persistent caching which is only
+  /// beneficial for queries that are repeated across app sessions.
+  /// For regular user searches, use getMatchingFiles instead.
+  Future<Map<String, List<int>>> getMatchingFileIDsForCommonQueries(
     Map<String, double> queryToScore,
   ) async {
     final textEmbeddings = <String, List<double>>{};
     final minimumSimilarityMap = <String, double>{};
+
     for (final entry in queryToScore.entries) {
       final query = entry.key;
       final score = entry.value;
-      final textEmbedding = await _getTextEmbedding(query);
+      // Use cache service instead of _getTextEmbedding
+      final textEmbedding =
+          await textEmbeddingsCacheService.getEmbedding(query);
       textEmbeddings[query] = textEmbedding;
       minimumSimilarityMap[query] = score;
     }
@@ -210,6 +217,7 @@ class SemanticSearchService {
       textEmbeddings,
       minimumSimilarityMap: minimumSimilarityMap,
     );
+
     final result = <String, List<int>>{};
     for (final entry in queryResults.entries) {
       final query = entry.key;

mobile/apps/photos/lib/services/magic_cache_service.dart

@@ -401,8 +401,8 @@ class MagicCacheService {
     for (Prompt prompt in magicPromptsData) {
       queryToScore[prompt.query] = prompt.minScore;
     }
-    final clipResults =
-        await SemanticSearchService.instance.getMatchingFileIDs(queryToScore);
+    final clipResults = await SemanticSearchService.instance
+        .getMatchingFileIDsForCommonQueries(queryToScore);
     for (Prompt prompt in magicPromptsData) {
       final List<int> fileUploadedIDs = clipResults[prompt.query] ?? [];
       if (fileUploadedIDs.isNotEmpty) {

mobile/apps/photos/lib/services/smart_memories_service.dart

@@ -37,7 +37,6 @@ import "package:photos/services/location_service.dart";
 import "package:photos/services/machine_learning/face_ml/person/person_service.dart";
 import "package:photos/services/machine_learning/ml_result.dart";
 import "package:photos/services/search_service.dart";
-import "package:photos/utils/text_embeddings_util.dart";
 
 class MemoriesResult {
   final List<SmartMemory> memories;
@@ -103,18 +102,29 @@ class SmartMemoriesService {
         'allImageEmbeddings has ${allImageEmbeddings.length} entries $t',
       );
 
-      // Load pre-computed text embeddings from assets
-      final textEmbeddings = await loadTextEmbeddingsFromAssets();
-      if (textEmbeddings == null) {
-        _logger.severe('Failed to load pre-computed text embeddings');
-        throw Exception(
-          'Failed to load pre-computed text embeddings',
+      _logger.info('Loading text embeddings via cache service');
+      final clipPositiveTextVector = Vector.fromList(
+        await textEmbeddingsCacheService.getEmbedding(
+          "Photo of a precious and nostalgic memory radiating warmth, vibrant energy, or quiet beauty — alive with color, light, or emotion",
+        ),
+      );
+
+      final clipPeopleActivityVectors = <PeopleActivity, Vector>{};
+      for (final activity in PeopleActivity.values) {
+        final query = activityQuery(activity);
+        clipPeopleActivityVectors[activity] = Vector.fromList(
+          await textEmbeddingsCacheService.getEmbedding(query),
         );
       }
-      _logger.info('Using pre-computed text embeddings from assets');
-      final clipPositiveTextVector = textEmbeddings.clipPositiveVector;
-      final clipPeopleActivityVectors = textEmbeddings.peopleActivityVectors;
-      final clipMemoryTypeVectors = textEmbeddings.clipMemoryTypeVectors;
+
+      final clipMemoryTypeVectors = <ClipMemoryType, Vector>{};
+      for (final memoryType in ClipMemoryType.values) {
+        final query = clipQuery(memoryType);
+        clipMemoryTypeVectors[memoryType] = Vector.fromList(
+          await textEmbeddingsCacheService.getEmbedding(query),
+        );
+      }
+      _logger.info('Text embeddings loaded via cache service');
 
       final local = await getLocale();
       final languageCode = local?.languageCode ?? "en";

mobile/apps/photos/lib/services/text_embeddings_cache_service.dart

@@ -0,0 +1,29 @@
+import 'package:logging/logging.dart';
+import 'package:photos/db/ml/db.dart';
+import 'package:photos/services/machine_learning/ml_computer.dart';
+
+class TextEmbeddingsCacheService {
+  static final _logger = Logger('TextEmbeddingsCacheService');
+
+  TextEmbeddingsCacheService._privateConstructor();
+  static final instance = TextEmbeddingsCacheService._privateConstructor();
+
+  Future<List<double>> getEmbedding(String query) async {
+    // 1. Check database cache
+    final dbResult =
+        await MLDataDB.instance.getRepeatedTextEmbeddingCache(query);
+    if (dbResult != null) {
+      _logger.info('Text embedding cache hit for query');
+      return dbResult;
+    }
+
+    // 2. Compute new embedding
+    _logger.info('Computing new text embedding for query');
+    final embedding = await MLComputer.instance.runClipText(query);
+
+    // 3. Store in database cache
+    await MLDataDB.instance.putRepeatedTextEmbeddingCache(query, embedding);
+
+    return embedding;
+  }
+}

mobile/apps/photos/lib/utils/text_embeddings_util.dart

@@ -1,172 +0,0 @@
-import 'dart:convert';
-import "dart:developer" as dev show log;
-import "dart:io" show File;
-
-import 'package:flutter/services.dart';
-import 'package:logging/logging.dart';
-import 'package:ml_linalg/vector.dart';
-import "package:path_provider/path_provider.dart"
-    show getExternalStorageDirectory;
-import 'package:photos/models/memories/clip_memory.dart';
-import 'package:photos/models/memories/people_memory.dart';
-import "package:photos/services/machine_learning/ml_computer.dart"
-    show MLComputer;
-
-final _logger = Logger('TextEmbeddingsUtil');
-
-/// Loads pre-computed text embeddings from assets
-Future<TextEmbeddings?> loadTextEmbeddingsFromAssets() async {
-  try {
-    _logger.info('Loading text embeddings from assets');
-    final jsonString =
-        await rootBundle.loadString('assets/ml/text_embeddings.json');
-    final data = json.decode(jsonString) as Map<String, dynamic>;
-
-    final embeddings = data['embeddings'] as Map<String, dynamic>;
-
-    // Parse clip positive embedding
-    Vector? clipPositiveVector;
-    final clipPositive = embeddings['clip_positive'] as Map<String, dynamic>;
-    final clipPositiveVectorData =
-        (clipPositive['vector'] as List).cast<double>();
-    if (clipPositiveVectorData.isNotEmpty) {
-      clipPositiveVector = Vector.fromList(clipPositiveVectorData);
-    }
-
-    // Parse people activities embeddings
-    final Map<PeopleActivity, Vector> peopleActivityVectors = {};
-    final peopleActivities =
-        embeddings['people_activities'] as Map<String, dynamic>;
-    for (final activity in PeopleActivity.values) {
-      final activityName = activity.toString().split('.').last;
-      if (peopleActivities.containsKey(activityName)) {
-        final activityData =
-            peopleActivities[activityName] as Map<String, dynamic>;
-        final vector = (activityData['vector'] as List).cast<double>();
-        if (vector.isNotEmpty) {
-          peopleActivityVectors[activity] = Vector.fromList(vector);
-        }
-      }
-    }
-
-    // Parse clip memory types embeddings
-    final Map<ClipMemoryType, Vector> clipMemoryTypeVectors = {};
-    final clipMemoryTypes =
-        embeddings['clip_memory_types'] as Map<String, dynamic>;
-    for (final memoryType in ClipMemoryType.values) {
-      final typeName = memoryType.toString().split('.').last;
-      if (clipMemoryTypes.containsKey(typeName)) {
-        final typeData = clipMemoryTypes[typeName] as Map<String, dynamic>;
-        final vector = (typeData['vector'] as List).cast<double>();
-        if (vector.isNotEmpty) {
-          clipMemoryTypeVectors[memoryType] = Vector.fromList(vector);
-        }
-      }
-    }
-
-    // Check if we have all required embeddings
-    if (clipPositiveVector == null) {
-      _logger.severe('Clip positive vector is missing');
-      throw Exception('Clip positive vector is missing');
-    }
-
-    if (peopleActivityVectors.length != PeopleActivity.values.length) {
-      _logger.severe('Some people activity vectors are missing');
-      throw Exception('Some people activity vectors are missing');
-    }
-
-    if (clipMemoryTypeVectors.length != ClipMemoryType.values.length) {
-      _logger.severe('Some clip memory type vectors are missing');
-      throw Exception('Some clip memory type vectors are missing');
-    }
-
-    _logger.info('Text embeddings loaded successfully from JSON assets');
-    return TextEmbeddings(
-      clipPositiveVector: clipPositiveVector,
-      peopleActivityVectors: peopleActivityVectors,
-      clipMemoryTypeVectors: clipMemoryTypeVectors,
-    );
-  } catch (e, stackTrace) {
-    _logger.severe('Failed to load text embeddings from JSON', e, stackTrace);
-    return null;
-  }
-}
-
-class TextEmbeddings {
-  final Vector clipPositiveVector;
-  final Map<PeopleActivity, Vector> peopleActivityVectors;
-  final Map<ClipMemoryType, Vector> clipMemoryTypeVectors;
-
-  const TextEmbeddings({
-    required this.clipPositiveVector,
-    required this.peopleActivityVectors,
-    required this.clipMemoryTypeVectors,
-  });
-}
-
-/// Helper function to generate text embeddings and save them to a JSON file
-/// Run this once to generate the embeddings, then copy the output
-/// to assets/ml/text_embeddings.json
-Future<void> generateAndSaveTextEmbeddings() async {
-  final Map<String, dynamic> embeddingsData = {
-    'version': '1.0.0',
-    'embeddings': {
-      'clip_positive': {},
-      'people_activities': {},
-      'clip_memory_types': {},
-    },
-  };
-
-  // Generate clip positive embedding
-  const String clipPositiveQuery =
-      'Photo of a precious and nostalgic memory radiating warmth, vibrant energy, or quiet beauty — alive with color, light, or emotion';
-  final clipPositiveVector =
-      await MLComputer.instance.runClipText(clipPositiveQuery);
-  embeddingsData['embeddings']['clip_positive'] = {
-    'prompt': clipPositiveQuery,
-    'vector': clipPositiveVector,
-  };
-
-  // Generate people activity embeddings
-  final peopleActivities = <String, dynamic>{};
-  for (final activity in PeopleActivity.values) {
-    final activityName = activity.toString().split('.').last;
-    final prompt = activityQuery(activity);
-    final vector = await MLComputer.instance.runClipText(prompt);
-    peopleActivities[activityName] = {
-      'prompt': prompt,
-      'vector': vector,
-    };
-  }
-  embeddingsData['embeddings']['people_activities'] = peopleActivities;
-
-  // Generate clip memory type embeddings
-  final clipMemoryTypes = <String, dynamic>{};
-  for (final memoryType in ClipMemoryType.values) {
-    final typeName = memoryType.toString().split('.').last;
-    final prompt = clipQuery(memoryType);
-    final vector = await MLComputer.instance.runClipText(prompt);
-    clipMemoryTypes[typeName] = {
-      'prompt': prompt,
-      'vector': vector,
-    };
-  }
-  embeddingsData['embeddings']['clip_memory_types'] = clipMemoryTypes;
-
-  // Convert to JSON and log it
-  final jsonString = const JsonEncoder.withIndent('  ').convert(embeddingsData);
-  dev.log(
-    '_generateAndSaveTextEmbeddings: Generated text embeddings JSON',
-  );
-
-  final tempDir = await getExternalStorageDirectory();
-  final file = File('${tempDir!.path}/text_embeddings.json');
-  await file.writeAsString(jsonString);
-  dev.log(
-    '_generateAndSaveTextEmbeddings: Saved text embeddings to ${file.path}',
-  );
-
-  dev.log(
-    '_generateAndSaveTextEmbeddings: Text embeddings generation complete! Copy the JSON output above to assets/ml/text_embeddings.json',
-  );
-}

mobile/apps/photos/pubspec.yaml

@@ -349,7 +349,6 @@ flutter:
     - assets/image-editor/
     - assets/icons/
     - assets/launcher_icon/
-    - assets/ml/
   fonts:
     - family: Inter
       fonts:

mobile/apps/photos/scripts/internal_changes.txt

@@ -1,2 +1,3 @@
-- Neeraj: Potential fix for ios in-app payment
-- Neeraj: (i) Debug option to enable logViewer
+- Laurens: text embedding caching for memories and discover
+- Neeraj: (i) Debug option to enable logViewer
+- Neeraj: Potential fix for ios in-app payment