| [DEPLOYMENT] | OS-Native Desktop Software Environment |
| [COMPILATION] | Static Machine Code via Nuitka Optimization |
| [GRAPHICS LAYER] | Hardware Accelerated UI via PySide6 (Qt) |
| [DATA BOUNDARY] | 100% Local Air-Gapped Network Topology // Zero Third-Party Telemetry |
| [INFRASTRUCTURE] | A Specialized Software Systems Unit of True Computing LLC |
Centralized generative AI architectures are a catastrophic security compromise. Trillion-dollar cloud platforms have successfully conditioned developers and scientists into routing confidential datasets and internal files through external corporate pipelines. The trade-off is clear: you lose structural data ownership in exchange for volatile prompt wrappers, unstable API dependencies, and confident hallucinations.
TC Labs is not a wrapper; it is a localized intelligence engine. We have engineered a sophisticated stack designed for developers, researchers, and security engineers who require persistent state, deterministic caching, and deep contextual awareness without compromising data sovereignty.
[//] Persistent JSON State & Forensics
Unlike ephemeral chat sessions, our engine supports full session serialization. Save and load complex reasoning paths as clean JSON files. The system automatically strips injected RAG context and system prompts during export, leaving you with a forensically auditable transcript of the intelligence dialogue.
[//] Deterministic Prompt Hashing
We implement strict SHA-256 hashing of the entire context window (including retrieved RAG data) before transmission. This enables aggressive prompt caching with cloud providers, reducing latency and cost for repetitive queries while ensuring identical inputs produce identical outputs.
[//] The "Pin Chip" Scope System
Navigate massive knowledge bases with surgical precision. Use our visual "Chip" system to pin specific files (`@`) or restrict searches to single documents using an inline automatic file selector. Use @allfiles to quickly load your entire vault of files into model memory. Override global context instantly, ensuring the AI focuses strictly on the codebase or dataset relevant to your current task.
[//] Dual-Mode Cognitive Engine
Switch seamlessly between "Chat Mode" for rapid code assistance and "Deep Think Mode" for complex problem solving. Deep Think routes requests to specialized reasoning models, emitting internal monologue streams that can be monitored in real-time via the sidebar log.
[//] ONNX CPU Embedding Hybrid
We leverage Nomic-Embed-Text via ONNX Runtime for ultra-fast, local vectorization on standard CPUs. This preserves VRAM for inference while ensuring low-latency embedding generation. If the local model is unavailable, it seamlessly falls back to configured network APIs.
[//] Voice Interface (Whisper & Kokoro)
Hands-free operation via local Faster-Whisper transcription and Kokoro TTS synthesis. Audio is processed entirely on-device, streamed in chunks for near-instantaneous response, and can be interrupted mid-sentence via the "Halt" command.
-
Code-Aware AST Splitting
Standard splitters destroy code logic. Our pipeline uses Abstract Syntax Tree (AST) analysis to respect function definitions (`def`), classes, and imports. We never split a function in half, ensuring the AI analyzes complete logical units.
-
LanceDB Relational Vector Storage
We utilize LanceDB for columnar vector storage with Parent-Child architecture. Large context blocks are paired with small search vectors, allowing the AI to retrieve precise snippets while fetching surrounding paragraphs for holistic understanding.
-
Docling OCR Integration
For scanned PDFs or complex layouts, our system integrates with an external Docling Service (configurable in GUI). This provides high-fidelity optical character recognition and layout analysis before data is ever vectorized.
-
LanceDash Database Commander
Black-box databases are unacceptable. Our embedded LanceDash GUI allows you to visually inspect, audit, and surgically delete specific files or fragments from your knowledge vault without CLI access.
-
Syntax-Highlighted Code Blocks
Custom QTextEdit renderers with PythonHighlighter logic preserve indentation and color-code keywords, strings, and comments. Includes one-click "Copy" that strips markdown fencing for clean IDE integration.
-
Programmable Persona Engine
Create custom profiles via the Settings GUI. Define unique system prompts, temperature settings, and icons for different tasks (e.g., "Security Auditor," "Python Refactorer"). Switch contexts dynamically without restarting.
-
Universal OpenAI SSE Compliance
Built on the official OpenAI SDK with raw Server-Sent Events (SSE) fallback. Includes a prominent "HALT" button to instantly kill network streams, stop token generation, and cease audio playback.
SYSTEM READY // AWAITING INPUT_
Standalone Executable // No Runtime Dependencies // Air-Gap Compatible
COMING SOON [v1.0]