Every chunk trust-scored before it enters the prompt.
Trust every retrieved chunk before it reaches the model
Retrieval is the new attack surface: a single poisoned document or injected instruction in your vector store can hijack an answer. DeepintShield treats every retrieved chunk as untrusted input – scoring it for injection, poisoning, secrets, and source trust, then allowing, redacting, rejecting, or quarantining it inline. One line wraps any LangChain, LlamaIndex, or custom retriever, with no pipeline rewrite.
Key Features
Per-chunk Decisions
Allow, redact, reject, or quarantine each chunk individually, with indirect-prompt-injection and corpus-poisoning detection.
One-line retriever guards
Guard retriever and Guard Embedder wrap existing retrievers and embedders to filter both ingestion and retrieval.
Source trust & quarantine
ACL tags and trust levels per source, with automatic corpus quarantine the moment poisoning markers appear.
Verifiable citations
A cross-encoder doubles as a relevance and injection filter - no third-party reranking API in the data path.
Verifiable citations
Emit source/document/chunk/offset citations and block ungrounded answers when citations are required.
Provenance into agent policy
Verified-provenance requirements gate autonomous tool calls, so poisoned context can’t reach an agent’s reasoning.
RAG security, indirect prompt injection protection, RAG data poisoning detection, chunk-level access control, citation enforcement, RAG provenance