Declarative Agent-Filtered Corpus Architecture for Grounded Retrieval-Augmented Generation

We present a system architecture for grounded retrieval-augmented generation (RAG) in which multiple specialized agents share a single document corpus but each agent operates over a distinct, automatically maintained subset of the corpus. The architecture introduces three mechanisms not found in prior work: (1) a declarative agent configuration format that co-defines agent persona parameters and corpus filter specifications in a single YAML file, where the persona drives communication behavior and the corpus filter drives knowledge scope; (2) an ingestion-triggered selective embedding update system that automatically detects which agents are affected by a newly ingested document based on collection-tag matching and triggers only those agents' index updates; and (3) an incremental vector index management scheme with tombstone-based soft deletion, over-fetch compensation during similarity search, and staleness detection against a corpus manifest. The system is implemented in the open-source Grounding-AI library (PyPI: grounding-ai) and is released under a permissive license to establish prior art and prevent future patent encumbrance of the described techniques.