Your Knowledge.
Uncompromised.
SecureThink is an impenetrable, offline-first AI assistant. Encrypt your private documents, run edge models locally on your phone, and orchestrate native agents with zero server dependency. Experience seamless, end-to-end local RAG powered by our innovative on-device re-ranking architecture—zero data leaks.
End-to-End On-Device RAG
Traditional AI tools upload your sensitive documents to remote GPU clusters, incurring high subscription fees and severe privacy risks. SecureThink flips the paradigm: document parsing, overlapping chunk segmentation, ONNX embedding generation, and high-performance vector search all execute directly on your phone's processor.
| Dimension | Cloud-Based Server-RAG Stack | SecureThink local Edge-RAG |
|---|---|---|
| Deployment Cost | 💸 High Monthly Fees (requires GPU VMs, Vector databases, and search clusters) | 🔒 $0 Server Cost (Runs entirely on a single 6GB+ Android phone) |
| Data Privacy | ⚠️ Sends raw documents, queries, and contextual logs to cloud providers | 🛡️ Absolute Privacy (Keystore-locked AES-256 local storage) |
| Offline Capability | 📶 Dead without Internet connection; latency-prone under poor signals | 🚇 100% Offline-Ready (Instant retrieval in subways or airplanes) |
| Tokenization & Embedding | 🧠 Heavy Python workers + high-RAM GPU embedding servers (OpenAI / Cohere) | ⚡ On-device ONNX Mobile Runtime + Pure Kotlin WordPiece Tokenizer |
| Search & Re-Ranking | 💾 Distributing queries to ElasticSearch & Pinecone, plus heavy GPU re-rankers | 🔄 Room FTS4 + sqlite-vec KNN merged via Reciprocal Rank Fusion (RRF) |
100% Offline & Private
Endless Vaults & Local Ingestion
Import unlimited PDFs, Markdown, and notes into your Secure Vault. Everything is ingested locally without leaving your device. Your data remains strictly within the AES-GCM-256 encrypted sandbox.
On-Device Embedding & Precision Retrieval
Experience ultra-fast, offline RAG. SecureThink parses and embeds your documents using local ONNX Mobile Runtime and SQLite-Vec. Fine-tune your chunking and overlap parameters for maximum recall accuracy on each specific knowledge base.
Explicit RAG Scoping & Edge Inference
True RAG isn't just an attachment upload. Explicitly select your encrypted Knowledge Base scope to activate retrieval. Our native Reciprocal Rank Fusion (RRF) algorithm completely replaces heavy cloud re-rankers, allowing edge models (like Gemma 4) to securely synthesize your scored vectors completely offline.
Native Device Interactions
Edge intelligence goes beyond text. SecureThink's local models can safely orchestrate on-device tools—like setting timers or managing alarms—giving you a truly private, offline smart assistant completely immune to network latency or cloud outages.
But what when you need to solve complex problems and connect external resources?
Local models are incredibly secure, but their context windows are constrained by mobile hardware and complex reasoning can be limited. We built a bridge.
Extended Cloud Mind & Agents
Bring Your Own Intelligence
Connect SecureThink to your trusted desktop models (Ollama/LM Studio) or commercial APIs (OpenRouter/OpenAI). Expand your context window to handle massive document analysis.
Uncompromised Privacy via PII Redaction
When using Cloud Mind, your security remains intact. The local Privacy Guard automatically scrubs Personally Identifiable Information (emails, cards, IDs) from your RAG chunks before any data is sent over the network.
Limitless Agentic Capabilities via MCP
Empower your models with external tools without exposing your vault. Use built-in web search agents (Brave / Tavily) for live data, or connect external MCP servers in a tightly controlled environment.
Frequently Asked Questions
Learn more about our local security features, encryption framework, and offline models.
1. How does SecureThink protect user-imported documents? What file formats are supported?
We support importing PDFs, DOCX, CSV, TXT, and Markdown (.md) files from other apps, plus offline image text extraction (OCR). When imported, SecureThink creates a local replica of the original document within its private app sandbox, fully encrypted on-disk using AES-GCM-256 keys managed by the Android Keystore. This secure replica is dynamically used when you want to edit text, move files to different vaults, or adjust RAG chunk parameters for re-embedding. The extracted text chunks are stored in an encrypted SQLCipher database, keeping your data entirely locked on-device.
2. How does SecureThink protect user chat conversations?
Every chat session is fully encrypted and stored securely on your device. Accessing past chat histories requires biometric authentication (fingerprint/face recognition). If you are using the Local AI mode, the entire conversation and inference process happens completely offline on your device’s hardware.
3. How does SecureThink implement the RAG pipeline? Can I customize retrieval parameters?
SecureThink utilizes an innovative offline-first hybrid retrieval pipeline. It combines on-device cosine vector similarity search (via sqlite-vec) and full-text keyword matching (via Room FTS4), merged together using Reciprocal Rank Fusion (RRF) and dynamic active query keyword boosting for maximum recall accuracy. Furthermore, users can customize RAG parameters—including chunk size, chunk overlap, top-K, and similarity thresholds—separately for each individual vault session to optimize context windows.
4. What is the Cloud Mind interface? Does SecureThink send data to a proprietary cloud?
SecureThink is strictly offline-first and does not host, run, or bundle any proprietary cloud models. The 'Cloud Mind' page is an optional configuration utility that allows power users to connect their application to their own self-hosted local models (running on a desktop via Ollama or LM Studio) or custom cloud endpoints they explicitly trust (such as OpenRouter or OpenAI). To protect user privacy when communicating with external endpoints, SecureThink executes all prompts through an on-device Privacy Guard, scrubbing sensitive information before it leaves the device.
5. SecureThink's privacy policy mentions Firebase. What exactly is Firebase used for?
Firebase is utilized exclusively for non-identifiable, aggregated app telemetry, such as Crashlytics (to report app crashes and bugs) and basic Analytics (to understand app usage and stability). We do not use Firebase to store your chat logs, documents, or personal data.
6. What anonymous statistics does SecureThink collect? Is it enabled by default?
By default, telemetry and crash collection are completely disabled to ensure maximum privacy. Users can optionally opt-in to anonymous diagnostics via the settings menu. If enabled, the app only transmits standard non-identifiable trace data (like OS version and crash stack traces) to help us patch stability issues. Absolutely no user documents, notes, vault data, or chat histories are ever collected.
7. What Android System Tools does SecureThink support? Can users customize or import them?
SecureThink supports several secure system automation tools that run entirely offline. These include setting alarms, configuring timers, querying device status (battery, connectivity, and time), sharing text content via Android's native share sheet, and opening secure local URLs. Tools that interact with external services—such as drafting an email or creating a calendar meeting—never access your accounts silently; they strictly trigger standard, system-level intents that require your manual confirmation and approval before execution.
🚀 ROADMAP PREVIEW (V2): In our upcoming V2 release, we will introduce User-Defined System Tools! Users will be able to import custom JSON-defined tools or orchestrate Android Intents directly (e.g., creating a tool to launch mapping home in one tap, or firing local network webhooks) via a simple UI editor, significantly expanding local offline workflow automation.
8. If there is a crash or problem, how can I get help?
You can contact our support team directly by emailing us at [email protected]. If you are using the app, you can also navigate to Settings and click 'Email Technical Support', which will draft a message from your default mail client to help us resolve your issue.
9. Why does SecureThink only offer a few models? Aren't more models better?
We deliberately curate rather than accumulate. Every model in SecureThink must pass a rigorous evaluation: it must run reliably via LiteRT-LM on real Android devices, deliver genuine offline value that the default system cannot already provide, and meet our bar for inference quality, memory footprint, and stability. Many promising models — including on-device TTS, smaller LLMs, or experimental architectures — either require incompatible runtimes, consume disproportionate resources, or duplicate functionality already available through the OS. We would rather ship a few models that work flawlessly than a long list of models that crash, overheat, or drain your battery. When a model earns its place, you'll find it here.
Currently Supported
All models run fully offline via LiteRT-LM. Download directly from within the app. More models coming in future updates.
| Model | Size | Vision | Audio | Thinking | Min RAM |
|---|---|---|---|---|---|
| Qwen3.5-0.8B | 1.1 GB | ✗ | ✗ | ✗ | 4 GB |
| Gemma 4 E2B | 2.4 GB | ✓ | ✓ | ✓ | 6 GB |
| Gemma 4 E4B | 3.4 GB | ✓ | ✓ | ✓ | 12 GB |
We carefully evaluate each model before adding it — see FAQ #9 for our model selection philosophy.
Experience SecureThink Today
Run offline LLMs, encrypt your second brain, and execute secure agent commands. SecureThink is a paid-only, offline-first application with no ads, no in-app purchases, and no hidden subscriptions. All features are fully unlocked upon purchase. For any questions, please contact customer support: [email protected]