After the Edit, Freshness and Access Control of Tool-Calling Retrieval and Vector RAG for a Community-Edited Knowledge Base

Community knowledge bases change, as editors correct records, unpublish drafts and hide prices. An assistant grounded through tool calls to a live, gated data interface should give the new value at the next question and never reveal gated content, whereas vector retrieval-augmented generation (RAG) serves what its index copied until the index is updated. We test both expectations on the knowledge base of a Greek–English agricultural platform on Kythera, Greece, with Claude Haiku 4.5 as router and answer model. Forty-eight counterfactual edits were probed within the conversation that had given the old value and in new conversations before and after a scheduled re-index, and 48 synthetic canary records with unique markers were probed while gated, while visible and after revocation. In new conversations after an edit, a reconstruction of the platform's tool agent gave the new value in 91.7% of cases, RAG with immediate index updates in 95.8% and RAG with batch re-indexing in none before the re-index. Challenged in the conversation that had given the old value, the tool agent gave the new value in 88.4% of conversations, but RAG with immediate index updates, which retrieves with the latest message alone, only in 19.6%, usually withdrawing its earlier answer even when it had been right. The tool agent disclosed gated content in none of 240 probes, whereas batch-indexed RAG disclosed every revoked canary until the re-index and an indexer that ignored the gating flags disclosed gated content in 93 of 96 probes. Freshness depends on the conversation as well as the index, and gating must be enforced wherever content is copied.