Baseline RAG
Version: 1.0.0 | Last updated: 2026-07-16
Purpose
Define the minimum retrieval-generation baseline used before adding rewriting, fusion, graphs, or agents.
Why
A simple, observable baseline exposes whether complexity produces real value and localizes retrieval versus generation failures.
How
Authenticate the principal, derive tenant/ACL scope, retrieve from one evaluated index with predicates applied in the datastore, assemble bounded source-labelled context, require answer/abstain behavior, and emit citations to immutable source versions. Log query, policy, index, retriever, candidates, scores, prompt template, model, answer, and citation verification under privacy policy.
Tradeoffs
The baseline is easy to test and operate but can miss mixed lexical/semantic or multi-hop queries. Keep it as the control even after advanced paths ship.
Anti-patterns
- Calling post-retrieval filtering authorization.
- Returning an answer when authorized evidence is absent.
- Evaluating only end-answer quality.
Enterprise Considerations
Apply tenant isolation, data minimization, retention, deletion, source licensing, and provider data-processing controls.
Checklist
- ACL is enforced before retrieval.
- Every result and citation resolves to a source version.
- Retrieval and generation traces are reproducible.
- Unanswerable and cross-tenant tests pass.
References
Changelog
- 1.0.0 — 2026-07-16: Initial production baseline.