How Deadnet Works
A technical deep-dive into the system behind the USB drive.
The Architecture
USB Drive (128GB) ├── Launcher Scripts ← One-click startup per platform ├── AI Engine │ ├── llama-server ← Runs the LLM locally (CPU-only) │ └── Llama 3 8B Model ← 4.5GB quantized model file ├── Knowledge Base │ ├── ChromaDB Index ← Pre-built vector embeddings │ ├── 19 Source PDFs ← Browsable original documents │ └── Manifest ← Document metadata + checksums ├── Offline Wikipedia ← ~20GB Kiwix ZIM file └── Web Interface ← Gradio app (localhost:7860)
The Query Pipeline
When you ask Deadnet a question, here's what happens under the hood:
Your question is embedded
Your natural language query is converted into a 768-dimensional vector using the nomic-embed-text model. This captures the semantic meaning, not just keywords.
Relevant chunks are retrieved
ChromaDB searches 6,300+ knowledge chunks and returns the 6 most semantically similar to your question. A relevance threshold filters out low-quality matches.
Context is assembled
The retrieved chunks, your conversation history, and a carefully designed system prompt are assembled into a single context window (8,192 tokens max).
The LLM generates a response
Llama 3 8B processes the assembled context and generates a response that directly answers your question using only the retrieved source material. Tokens stream to your browser in real-time.
Sources are cited
Every response includes citations linking back to the exact document and page number. You can click through to read the original PDF yourself.
Technical Specifications
Safety Guardrails
Grounded responses only. Deadnet can only answer from its curated knowledge base. It cannot fabricate information, access the internet, or draw from general AI training data.
Medical and legal disclaimers. Any response involving medical treatment, legal rights, or safety-critical information automatically includes a disclaimer that it is not a substitute for professional advice.
Honest uncertainty. When Deadnet doesn't have relevant information in its knowledge base, it says so clearly and suggests related topics that might help.