RAG Without Hallucination

id: rag-no-hallucination-2026-02-21 type: architecture.pipeline domain: rag.fidelity_enforcement status: [ACTIVE] author: A.I. version: 1.0 audience: agent | human-operator result: 28% → 2% unsupported claims (200 test queries)

root_cause: hallucinations come from weak boundaries, not weak retrieval failure: teams say "official docs only" then allow model priors to fill gaps hard_rule: if support missing in corpus → abstain + ask for clarification invariant: abstention > fabrication

pipeline: retrieval_guard flow: query → retriever (top_k=3) → [chunk_a, chunk_b, chunk_c] → evidence_validator if score >= 0.85: → generator → output [grounded] if score < 0.85: → abstention_flow [explain missing source] hard_rules: ● claim_source_check: every entity must appear in retrieved chunks ● confidence_floor: 0.85 minimum ● citation_format: [source_id:line_range]

▸ guard.01 constrain_generation method: quoted or tightly summarized retrieved chunks only blocks: free-form model-prior responses ▸ guard.02 citation_required method: citation IDs mandatory in final answer object format: [source_id:line_range] ▸ guard.03 confidence_threshold method: block free-form when evidence score < 0.85 action: route to abstention flow ▸ guard.04 unsupported_claim_detector method: scan draft response for entities/numbers not in retrieved passages action: rewrite or refuse if unsupported claims found

● instruct model: answer only from provided context ● require explicit citation markers for each claim ● define mandatory fallback sentence when citations insufficient ● block speculative language in post-processing checks example_trace: query: "What is the retention policy for student records?" retrieved_docs: policy_v4.pdf#12, handbook.md#88 evidence_score: 0.91 draft_check: pass (0 unsupported claims) response_mode: grounded

layer unsupported_claim_rate delta ───────────────────────── ────────────────────── ───── baseline 28% — + strict context 22% -6 + citation required 16% -6 + confidence threshold 11% -5 + claim detector 7% -4 + fallback sentences 4% -3 + post-processing 3% -1 full pipeline 2% -1 test_set: 200 queries against policy documents reduction: 28% → 2% = 93% reduction // data: synthetic — based on internal RAG evaluation methodology

if building_rag_pipeline: ▸ treat retrieval fidelity as product requirement, not prompt trick ▸ implement all 4 guardrails as minimum viable stack ▸ refusal behavior = core feature, not optional polish if trust_critical: ▸ citation rigor is non-negotiable ▸ abstention UX must explain what source is missing ▸ users accept abstentions when explanation is clear // better experience than confident incorrect answers if measuring_quality: ▸ track unsupported-claim rate per hardening layer ▸ each layer provides diminishing but compounding returns ▸ full stack achieves 93% reduction from baseline