Business Rules - RAG Systems

This document lists the fundamental business rules of the RAG Systems domain (Retrieval-Augmented Generation systems), identified by a unique BR-NNN code. Each rule is technology-independent and describes an invariant constraint on the indexing and querying pipeline.

---

BR-001 - Embedding-space consistency between indexing and query

Code: BR-001 Name: Same embedding model for documents and queries Description: The Embeddings of the indexed Chunks and those of the Querys must be produced by the same model and have the same dimension. Changing the embedding model requires a full re-indexing of the VectorIndex. Rationale: Similarity is meaningful only within the same vector space; different models make comparisons meaningless. Priority: CRITICAL Actors: Indexing Pipeline, Retriever Violation: searching with a different model's embedding → forced re-indexing. Traceability: INV-01; indexing-pipeline.md.

---

BR-002 - Every Chunk preserves its provenance

Code: BR-002 Name: Source traceability for every fragment Description: Every indexed Chunk must keep the reference to its source Document, its position and the indexing date. Grounded answers cite the source through these references. Rationale: Enable citation, verification and attribution of claims (verifiable grounding). Priority: HIGH Actors: Indexing Pipeline, Generator Traceability: BR-006; INV-02; query-answering.md.

---

BR-003 - The answer is grounded in the retrieved context

Code: BR-003 Name: Mandatory grounding on the context Description: The Generator must base the answer exclusively on the relevant retrieved Chunks. When the context does not contain the information, the answer must declare uncertainty ("not available in the documents") rather than resorting to unverified parametric knowledge. Rationale: Reduce hallucinations and ensure faithfulness (the central promise of RAG). Priority: CRITICAL Actors: Generator Violation: claims unsupported by the context → the faithfulness evaluation fails. Traceability: INV-03; query-answering.md.

---

BR-004 - Relevance threshold on retrieved chunks

Code: BR-004 Name: Filtering of low-relevance results Description: Only Chunks with a relevance score above a threshold (after any re-ranking) are passed to the Generator. Irrelevant chunks must be discarded so as not to pollute the context and induce hallucinations. Rationale: Noisy context degrades answer quality more than its absence. Priority: HIGH Actors: Retriever, Reranker Traceability: query-answering.md; INV-04.

---

BR-005 - The context respects the window and the budget

Code: BR-005 Name: Top-k and token-budget management Description: The number of Chunks sent to the Generator (top-k) and their overall length must fit within the context window and the token budget. Beyond the limit, explicit selection/compaction is applied, never a silent truncation of the most relevant chunk. Rationale: Avoid overflow and loss of critical context ("lost in the middle"). Priority: HIGH Actors: Retriever, Generator Traceability: INV-05; query-answering.md.

---

BR-006 - Re-indexing on source changes

Code: BR-006 Name: Index freshness Description: When a Document is added, modified or removed, the Indexing Pipeline must update the VectorIndex so the answers reflect the current state of the sources. Chunks of removed documents must no longer be retrievable. Rationale: Avoid answers based on stale information or on documents no longer valid. Priority: HIGH Actors: Indexing Pipeline Traceability: indexing-pipeline.md; INV-06.

---

BR-007 - Data isolation in multi-tenancy

Code: BR-007 Name: Metadata filtering and access boundaries Description: When the index serves multiple tenants or authorisation levels, every Query must retrieve only the Chunks the requester is entitled to access, via metadata filtering applied before or during the search. Rationale: Prevent data leakage across tenants or beyond the user's permissions. Priority: CRITICAL Actors: Retriever, Guardrail Traceability: INV-07; query-answering.md.

---

BR-008 - Defence against prompt injection in documents

Code: BR-008 Name: Retrieved content is untrusted data Description: Retrieved Chunks (coming from external documents) are untrusted content: they must be delimited in the prompt and inspected by the Guardrails, because they may contain injection instructions aimed at the model. Rationale: Defend against prompt injection carried by indexed documents. Priority: CRITICAL Actors: Generator, Guardrail Traceability: OWASP LLM Top 10 (LLM01); INV-08.