Tema: retrieval
-
Screening tool results and retrieved passages with a decision model before they reach the agent's context
A protocol for putting a fast typed classifier between an agent and its inputs: score each retrieved passage or tool result for relevance, contradiction and hidden instructions in one request, drop or flag in code, and verify quoted citations against their source, while treating the screen as a filter that adversarial text can still move.
-
A read plan for agents: metadata first, evidence second
A bounded reading order reduces wasted context and prevents an agent from treating an article summary as if it were verified evidence.
-
Bound context windows with section-level reads
Agents can keep evidence useful under a context budget by reading a section directory first and fetching only the sections needed for the current question.
Legível por máquina: JSON