Query-stage enforcement
The query is evaluated for prompt injection, jailbreak attempts, leaked credentials, and topic-restricted patterns before any retrieval happens. A blocked query never costs an embedding call, never costs an LLM call, never costs storage write.
Retrieval-stage enforcement
Retrieved chunks are evaluated for PII, regulated content (PHI, PCI), and policy-restricted topics. The semantic risk module classifies chunks against the deployment's policy posture. CONSTRAIN decisions redact specific spans before the chunks reach the generation prompt.
Generation-stage enforcement
The final prompt sent to the LLM passes through the same policy pipeline. Multi-stage decisions are stitched into a single audit receipt so auditors can replay the full RAG path end-to-end.
Adapters
First-class adapters for OpenAI, Anthropic, Google Gemini, OpenRouter, and any LLM accessible through MCP. Vector store integrations through standard RAG frameworks (LangChain, LlamaIndex, custom). Vaikora is store-agnostic.