Pocket Models

Flutter source code for Pocket Models — a privacy-first, on-device AI assistant built on the DataSapien SDK. Run local LLMs against your own personal data ("MeData") on your phone, without sending anything to the cloud.

This repository contains the host-app source code. The DataSapien SDK itself is a commercial product and requires a subscription — see https://datasapien.com/pricing/.

How it works

flowchart TB
    Dev([Orchestrator Admin])
    User([App User])

    Orch[DataSapien Orchestrator<br/>web portal · Journeys · AI Models · MeData Definitions]

    subgraph App["Pocket Models app"]
        UI[Flutter UI<br/>screens · widgets]
        VM[ViewModels &amp; Services]
    end

    subgraph SDK["DataSapien SDK"]
        Int[IntelligenceService]
        Me[MeDataService]
        Jou[JourneyService]
    end

    subgraph Device["On-device — never leaves the phone"]
        Model[(LLM weights<br/>GGUF)]
        Store[(MeData<br/>DataVault)]
    end

    Backend[(DataSapien backend<br/>auth · AI Models · Journeys · MeData Definitions)]

    Dev -->|configures| Orch
    Orch -.->|publishes| Backend

    User <--> UI
    UI <--> VM
    VM -->|invokeModel · loadModel| Int
    VM -->|read · write personal data| Me
    VM -->|runJourney · syncJourneys| Jou
    Int <--> Model
    Me <--> Store
    Backend -.->|AI Models · Journeys · MeData Definitions| SDK

Loading

The diagram has two halves. Config-time (dashed): an Orchestrator Admin uses the DataSapien Orchestrator (web portal) to publish Journeys, register AI Models, and define MeData schemas — those flow into the DataSapien backend, which the SDK pulls from at runtime. Runtime (solid): the app's UI talks to ViewModels, which call into the three DataSapien SDK services. Inference and personal data stay on the device.

---

Table of contents

1. Integrate with DataSapien
2. DataSapien API surface used in this app
   • IntelligenceService
   • MeDataService
   • JourneyService
   • Diagnostics
3. Running the app
4. Documentation links
5. License

---

Integrate with DataSapien

1. Add the SDK dependency

In pubspec.yaml:

dependencies:
  datasapien_sdk: ^0.42.0
  datasapien_sdk_health: ^0.42.0   # optional — only if you use HealthKit / Health Connect

Then:

flutter pub get

2. Initialize and set up the SDK

The integration lives entirely in lib/main.dart. You build a DataSapienConfig, hand it to DataSapien.initialize(...), then call DataSapien.setup() once at startup.

import 'package:datasapien_sdk/datasapien_sdk.dart';

DataSapienConfig _buildDataSapienConfig() {
  return DataSapienConfig.builder()
      .setAuth(
        authUrl: 'YOUR_AUTH_URL',
        authClientId: 'YOUR_CLIENT_ID',
        authClientSecret: 'YOUR_CLIENT_SECRET',
        authScope: 'YOUR_AUTH_SCOPE',
      )
      .setHostUrl('YOUR_HOST_URL')
      .setMediaUrl('YOUR_MEDIA_URL')
      .setMainColor('#1464FA')
      .setDebug(true)
      .build();
}

void main() async {
  WidgetsFlutterBinding.ensureInitialized();

  final config = _buildDataSapienConfig();
  await DataSapien.initialize(config);

  // setup() validates credentials, hydrates managed-model metadata, etc.
  await DataSapien.setup();

  runApp(const MyApp());
}

Where do YOUR_* values come from?

You receive them when you subscribe to DataSapien (https://datasapien.com/pricing/). The placeholders map to:

Placeholder	What it is
YOUR_AUTH_URL	OAuth2 token endpoint for your DataSapien tenant
YOUR_CLIENT_ID	OAuth2 client ID
YOUR_CLIENT_SECRET	OAuth2 client secret (never commit to a public repo)
YOUR_AUTH_SCOPE	OAuth2 scope for the DataSapien API
YOUR_HOST_URL	Backend host URL (e.g. your tenant's MBE endpoint)
YOUR_MEDIA_URL	Media/CDN host used for managed-model artwork (also set in lib/utils/app_constants.dart)

For production you should keep these out of source — load them from environment variables, --dart-define flags, or a secure config file.

---

DataSapien API surface used in this app

All SDK access goes through DataSapien.getXxxService() accessors. The three services this app touches are listed below, grouped by purpose, with a short description and a pointer to where each call is used in this codebase.

IntelligenceService — DataSapien.getIntelligenceService()

On-device model management and inference.

Method	What it does	Used in
getManagedAIModels()	Returns the full list of models registered for the tenant (downloadable catalog). Each entry has name, multimodal, imageUrl, etc.	lib/services/model_selector_source.dart, lib/services/model_download_manager.dart
getDownloadedModelsList()	Lists names of models whose weights are already on disk.	lib/services/model_download_manager.dart
isModelFilesDownloaded(modelName)	true if every weight file for the given model is present locally.	lib/services/model_download_manager.dart
downloadModelFiles(modelName, {onProgress})	Downloads the model weights. onProgress: (double p) is called repeatedly with 0.0–1.0.	lib/services/model_download_manager.dart
deleteModelFiles(modelName)	Removes the model's weight files from disk.	lib/services/model_download_manager.dart
loadModel(modelName, modelKey, {modelParams})	Loads the model into memory under an arbitrary modelKey you choose. modelParams carries context size / GPU layers / etc.	lib/viewmodels/main_chat_view_model.dart
unloadModel({key})	Frees model memory for a previously loaded modelKey.	lib/viewmodels/main_chat_view_model.dart
invokeModel(modelKey, prompts, {inferenceParams, onStream})	Runs an inference against a loaded model. Returns the full generated string. If onStream: (String chunk) { ... } is supplied, it fires per-token while generating.	lib/viewmodels/main_chat_view_model.dart
stopModelInference(modelKey)	Cancels an in-flight invokeModel call for the given modelKey.	lib/viewmodels/main_chat_view_model.dart

MeDataService — DataSapien.getMeDataService()

The user's local "MeData" store — categorized personal data the on-device model can read at inference time. Everything stays on the device.

Method	What it does	Used in
getMeDataCategories()	Returns top-level categories (e.g. Health, Personalization, Identity).	lib/services/me_data_category_loader.dart
getMeDataDefinitions()	Returns every defined MeData key and its schema.	lib/screens/settings/data_privacy_screen.dart
getMeDataDefinitionsByCategory(categoryName)	Definitions filtered to one category.	lib/services/me_data_category_loader.dart
getMeDataDefinition(definitionName)	Single-definition lookup. Returns null if the definition isn't known.	lib/viewmodels/main_chat_view_model.dart
getMeDataRecords(definitionName)	All stored values for a given MeData definition (history, newest-last).	lib/services/me_data_category_loader.dart
getLastMeDataRecord(definitionName)	The most recently stored value for a definition.	lib/viewmodels/main_chat_view_model.dart
saveMeDataRecord(definitionName, value)	Persists a new value (appends a new history entry).	lib/theme/theme_manager.dart, lib/screens/settings/memory_settings_screen.dart
deleteMeData(definitionName)	Deletes every record for a definition.	lib/screens/profile/my_data_tab.dart
deleteMeDataRecord(definitionName, recordId)	Deletes one specific history entry by its record ID.	lib/screens/profile/my_data_history_screen.dart

JourneyService — DataSapien.getJourneyService()

Journeys are server-defined JS workflows that run on-device and produce MeData outputs (e.g. an onboarding sequence that fills in baseline data).

Method	What it does	Used in
getJourneys({tags, statuses, onlyInAudience})	Lists journeys available to the user, filterable by tags (e.g. ['ai']), statuses (JourneyStatus.notStarted, .completed, …), and onlyInAudience: true to hide journeys the user doesn't qualify for.	lib/screens/profile/journeys_tab.dart, lib/screens/chat/main_chat_screen.dart
runJourney(journeyName, {data})	Executes a journey by name. data is a Map<String, dynamic> of inputs the journey script may consume.	lib/screens/profile/journeys_tab.dart
syncJourneys()	Re-pulls the latest journey definitions from the backend and re-evaluates derived MeData. Call after any MeData mutation that may unlock new journeys.	lib/viewmodels/main_chat_view_model.dart

Diagnostics

DataSapienDiagnostics records SDK + UI events for support escalation. Setup example from lib/main.dart:

DataSapienDiagnostics.instance
  ..configure(
    const DataSapienDiagnosticsConfig(
      mode: DataSapienDiagnosticsMode.verboseSupport,
    ),
  )
  ..setEnabled(true)
  ..logUiEvent('Pocket Models diagnostics recording started');

---

Running the app

flutter pub get
flutter run

The app will start, but inference and MeData calls will fail until you replace the YOUR_* placeholders in lib/main.dart and lib/utils/app_constants.dart with real DataSapien credentials.

Pocket Models also uses Firebase Crashlytics. You'll need to run flutterfire configure once to generate lib/firebase_options.dart and the platform config files before the app will build for Android/iOS. Both are gitignored.

---

Documentation links

• DataSapien developer portal: https://dev.datasapien.com/
• DataSapien marketing site: https://datasapien.com/
• Pricing & SDK subscription: https://datasapien.com/pricing/
• Flutter SDK on pub.dev: datasapien_sdk · datasapien_sdk_health

---

License

MIT. See LICENSE.

Note: the MIT license covers only the host-app source code in this repository. The DataSapien SDK is a separate commercial product behind https://datasapien.com/pricing/.