Your MCP Server Returns the Source. Does the Answer Still Cite It?
A benchmark for what happens to citations after a retrieval server hands its results to an AI agent it doesn't control. Part 1 of 6 in the series Citations at the MCP Boundary . This is where the series starts. The parts after this one go deeper into each finding; the full list is at the end of this post. Here's a situation that's becoming normal very quickly. A company turns its documentation…
A benchmark tests how citations behave when a Retrieval Augmented Generation (RAG) system, accessed via the Model Context Protocol (MCP), interfaces with an AI agent. The goal is to determine if citations survive the interaction between the retrieval server and the AI agent. The source material outlines a detailed experiment conducted to measure the "Citation Survival Rate" (CSR), which assesses whether the retrieved sources are accurately cited in the final AI-generated answer.
The experiment involves three AI agents - a LangGraph agent, an OpenAI Agents SDK agent, and a single-shot Retrieval Augmented Generation (RAG) system, all using the same model (gpt-4o-mini), documentation (FastAPI), system prompt, and citation format ([url] after each claim). The CSR metric required two checks: the support check, which verifies that the cited chunk actually supports the generated claim, and the trace check, ensuring that the URL cited was indeed retrieved by the agent.
The results showed that the LangGraph agent matched the CSR of the single-shot RAG, while the OpenAI Agents SDK agent had a significantly lower CSR, losing about half of its citations. The experiment underscores the importance of ensuring citations are preserved when AI agents interact with RAG systems over MCP boundaries.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
