Welcome to Geodesia G-1¶

Geodesia G-1 is a validating gateway that sits in front of any large language model (LLM) and provides a comprehensive quality and compliance layer. It is fully OpenAI-compatible — your existing application sends requests to Geodesia G-1 exactly as it would to OpenAI, and the gateway forwards them to your chosen underlying model (vLLM, Ollama, SGLang, OpenAI, TensorRT-LLM, and others) after enriching both the input and output with safety and reliability signals.

The platform is now Application-oriented — with G-1 Studio, one shared LLM and G1-Hummingbird detector can serve many isolated Applications, each with its own policy, thresholds, RAG collection, compliance posture, and cost center.

You do not need to retrain your model. You do not need to change your application code. You plug Geodesia G-1 in, and your LLM immediately gains:

• Hallucination detection — independent, separately calibrated signals that tell you whether the model's answer is grounded in the provided context or is a fabrication.
• Safety enforcement — real-time prompt screening and answer inspection to block unsafe, harmful, or jailbreak requests before they reach the model or the user.
• Regulatory compliance — a full audit trail, EU AI Act impact assessments, GDPR data retention, kill-switch, human oversight escalation, and more — across 13 global frameworks.
• Causal explainability — token-level attribution that shows exactly which words in the input caused the model's output, with no access to model internals required — painted as a heatmap on the text.
• Realtime voice guard — a streaming-Whisper layer in front of the detector screens spoken jailbreaks and unsafe prompts as they are said, and brakes mid-sentence.
• MCP security — starting the proxy starts an MCP security layer that vets every tool description, tool-call argument and tool result before an agent can act.
• Token & cost control — off-topic requests are refused before the upstream call (zero tokens billed) and easy prompts are routed to a cheaper model than hard ones, from the same detector pass.
• A security policy that evolves with you — Policy Lens simulates a threshold change on your own real traffic before you apply it, and every human correction feeds back into the detector.

How It Works¶

Every request flows through input validation → the LLM you choose → output validation, with each call written to the compliance ledger.

Every chat message goes through this pipeline:

1. Input validation — the prompt and conversation history are scored across prompt safety and jailbreak detection axes. If a threshold is exceeded in blocking mode, the request is refused before the model sees it. Spoken input is transcribed by a streaming Whisper layer and scored on the same axes, mid-utterance.
2. Context injection — if you supplied a grounding context (or uploaded documents to the knowledge base), it is injected into the upstream request.
3. Generation — the upstream LLM produces a response. For streaming requests, Geodesia monitors every 32 tokens (configurable cadence) and can halt generation mid-stream if dangerous content emerges.
4. Output validation — the completed answer is scored for hallucination and unsafe content. RAG answers additionally go through claim-level citation verification.
5. Compliance logging — the call is written to the audit database, watermarked, and linked to the running hash chain.
6. Response delivery — the original OpenAI-compatible response is returned with an additional geodesia field containing the full detection payload.

Quick Navigation¶