Knowledge RAG — Permission-aware RAG for AI agents
por Info Inlet Inc.
Permission-aware RAG knowledge base API for AI agents — ingest, hybrid search, cited answers.
Knowledge RAG is a transactable API that gives AI agents a permission-aware RAG knowledge base. Ingest your documents into a workspace, then have agents or against them — with hard per-document access control so an agent only ever retrieves what the calling user is allowed to see.
Built for enterprise AI assistants where leaking the wrong document is a compliance event: customer support copilots, internal-knowledge chat, HR / finance / legal assistants, agentic workflows over private corpora.
Key features
- Permission-filtered retrieval — every chunk has a ; and intersect against caller server-side. Disallowed chunks never reach the LLM context.
- Hybrid retrieval — dense vectors (Qdrant) + lexical (Postgres FTS) fused with Reciprocal Rank Fusion; optional graph lane for entity-based queries.
- Cited, grounded answers — refuses outside knowledge; every claim carries a [n] citation back to the source chunk.
- Multi-tenant by workspace — chunks are isolated by your subscription; cross-tenant leakage is structurally impossible.
- Japanese-first text handling — CJK-aware chunking and lexical bigrams, so Japanese RAG works out-of-the-box (most global RAG APIs treat JP as one token per sentence).
- Open MCP client — install the free npm package and any MCP-aware client (Claude Desktop, Cursor, etc.) gets / / tools instantly.
How it works
- Subscribe to a plan on Microsoft Marketplace; you receive an API key by email.
- Call with workspace id + text → we chunk, embed, and index into both lanes.
- Call for top-K chunks, or for a grounded answer with citations.
- Caller's array is intersected with each chunk's on every call.
Plans
- Starter — $9 / month. 200 queries included. $0.05 per overage query. Ingest is free.
- Pro — $39 / month. 2,000 queries included. $0.03 per overage query.
Architecture — hosted on dedicated EU bare-metal (Hetzner, not in Azure); TLS 1.2+ via Cloudflare; API keys SHA-256 hashed at rest; no document content sent to or stored in Azure.
Interactive Swagger docs at api-rag.knowlez.com/docs; OpenAPI spec at .