16 posts
We went looking for the benchmark behind our own published latency figure and could not find one. Measuring properly started at 302ms p95, surfaced three product bugs, and ended at a public, reproducible 16ms p50 / 21ms p95.
A RAG pipeline can enforce access inside the vector query or in the app after results return. Each has a distinct failure mode. Here is what breaks.
Embeddings throw away the permissions your source systems already track. Here is the recipe to carry document-level permissions into a RAG pipeline.
Gateco supports role, attribute, and relationship-based access control, and you can mix them in one policy set. Here is which model fits which pattern.
Gateco is not a RAG framework. It is the authorization layer you insert at the retrieval step of LangChain or LlamaIndex. Here is where it goes, and why.
An orchestrator decides which agents and steps run. Gateco decides what an agent can retrieve. Different layers, and retrieval is where RAG leaks.
When policy evaluation hits an error, Gateco denies the retrieval and logs it. Here is why fail-closed is the right default, and when fail-open fits.
Google has two retrieval products under the Vertex AI brand: Vector Search, a managed ANN index, and Vertex AI Search, Discovery Engine. When to use each.
The Gateco MCP server gives Claude Desktop, Cursor, and any MCP host policy-enforced access to your vector knowledge bases. Denied content never surfaces.
Gateco now supports 1-hop relationship-based access control: policies can check whether a principal owns or is assigned to a resource. How and when to use it.
Cerbos is a generic authorization engine. Gateco is a retrieval-specific security layer for RAG. They solve different problems, and can be used together.
pgvector Row Level Security is the most common DIY RAG auth pattern. When it works, when it breaks, and the five triggers that make teams outgrow it.
How much latency does an authorization layer add to RAG? The measured answer, with a public benchmark: 16ms p50 and 21ms p95 policy overhead, and what drives variance across connectors.
Azure AI Search is a managed search platform; pgvector, Pinecone, and Qdrant are retrieval primitives. The choice shapes your RAG architecture and governance.
Gateco now supports four distinct retrieval modes. Here's when to reach for each one, and why hybrid might be your new default.
DIY RAG authorization needs a policy engine, metadata resolution, audit logging, connector adapters, and identity sync. What it actually takes to build it.