SCENARIOS = [ { "name": "Employee handbook FAQ", "risk": "low", "corpus_complexity": "simple", "workflow": "answer", "source_formats": ["markdown"], }, { "name": "Security runbook assistant", "risk": "high", "corpus_complexity": "medium", "workflow": "answer_or_escalate", "source_formats": ["markdown"], }, { "name": "Company relationship research assistant", "risk": "medium", "corpus_complexity": "connected", "workflow": "answer", "source_formats": ["html", "pdf"], }, { "name": "Support ticket assistant", "risk": "medium", "corpus_complexity": "medium", "workflow": "action", "source_formats": ["markdown", "api"], }, { "name": "Product knowledge assistant", "risk": "medium", "corpus_complexity": "multimodal", "workflow": "answer", "source_formats": ["pdf", "slides", "images", "tables"], }, ] def recommend_rag_type(item): if item["corpus_complexity"] == "multimodal": return "multimodal RAG" if item["corpus_complexity"] == "connected": return "graph RAG" if item["workflow"] == "action": return "agentic RAG with approval gates" if item["risk"] == "high": return "corrective RAG" if item["corpus_complexity"] == "medium": return "advanced RAG" return "naive RAG" def risks_for(rag_type): risks = { "naive RAG": ["stale source", "missing citation"], "advanced RAG": ["metadata errors", "reranking latency"], "corrective RAG": ["false confidence signal", "over-refusal"], "graph RAG": ["wrong entity resolution", "missing relationship"], "agentic RAG with approval gates": ["tool misuse", "approval bypass"], "multimodal RAG": ["format extraction error", "table/image grounding failure"], } return risks.get(rag_type, ["unknown architecture risk"]) def advantages_for(rag_type): advantages = { "naive RAG": ["fast to build", "low cost", "easy to debug"], "advanced RAG": ["better precision", "metadata-aware", "stronger grounding"], "corrective RAG": ["safer under weak evidence", "supports refusal and escalation"], "graph RAG": ["relationship-aware", "good for multi-hop entity questions"], "agentic RAG with approval gates": ["handles workflows", "can use tools with control"], "multimodal RAG": ["works across images, tables, slides, and text"], } return advantages.get(rag_type, ["depends on implementation"]) def disadvantages_for(rag_type): disadvantages = { "naive RAG": ["weak for multi-hop", "sensitive to chunk quality"], "advanced RAG": ["more tuning", "more latency"], "corrective RAG": ["can over-refuse", "thresholds are hard to tune"], "graph RAG": ["entity resolution complexity", "graph freshness burden"], "agentic RAG with approval gates": ["tool risk", "harder to test", "higher cost"], "multimodal RAG": ["harder extraction", "harder citation and evaluation"], } return disadvantages.get(rag_type, ["depends on implementation"]) def engineering_profile_for(rag_type): profiles = { "naive RAG": { "scale": "low to medium", "accuracy": "good when the right chunk is retrieved", "consistency": "medium; drops with conflicting chunks", "latency_cost": "low", }, "advanced RAG": { "scale": "medium to high", "accuracy": "high with strong metadata and reranking", "consistency": "high when filters and index versions are stable", "latency_cost": "medium", }, "corrective RAG": { "scale": "medium", "accuracy": "high safety; lower unsupported-answer rate", "consistency": "depends on confidence thresholds", "latency_cost": "medium to high", }, "graph RAG": { "scale": "high if graph pipeline is maintained", "accuracy": "high for connected knowledge when entity links are correct", "consistency": "depends on graph freshness and entity resolution", "latency_cost": "high", }, "agentic RAG with approval gates": { "scale": "medium; requires workflow controls", "accuracy": "task-oriented, not just answer-oriented", "consistency": "harder because plans can vary", "latency_cost": "high", }, "multimodal RAG": { "scale": "high pipeline complexity", "accuracy": "depends on extraction and multimodal grounding", "consistency": "harder across formats", "latency_cost": "high", }, } return profiles.get( rag_type, {"scale": "unknown", "accuracy": "unknown", "consistency": "unknown", "latency_cost": "unknown"}, ) def client_fit_for(rag_type): fit = { "naive RAG": ["internal FAQ", "small docs portal", "training assistant"], "advanced RAG": ["customer support", "HR policy", "developer docs", "finance policy"], "corrective RAG": ["security", "legal triage", "compliance", "healthcare policy"], "graph RAG": ["research intelligence", "fraud", "supply-chain risk", "legal discovery"], "agentic RAG with approval gates": ["support operations", "DevOps", "ticketing", "workflow copilots"], "multimodal RAG": ["product support", "healthcare", "manufacturing", "education"], } return fit.get(rag_type, ["custom client fit"]) def evals_for(rag_type): evals = { "naive RAG": ["Recall@5", "citation precision"], "advanced RAG": ["filter accuracy", "reranked source recall"], "corrective RAG": ["unsafe-answer rejection rate", "escalation correctness"], "graph RAG": ["entity resolution accuracy", "relationship recall"], "agentic RAG with approval gates": ["tool-call precision", "approval-gate pass rate"], "multimodal RAG": ["format-specific grounding", "cross-modal citation accuracy"], } return evals.get(rag_type, ["manual review"]) def decision_record(item): rag_type = recommend_rag_type(item) return { "scenario": item["name"], "selected_rag_type": rag_type, "why": f"Risk={item['risk']}, corpus={item['corpus_complexity']}, workflow={item['workflow']}.", "what_it_is": f"{rag_type} selected for this scenario shape.", "client_fit": client_fit_for(rag_type), "advantages": advantages_for(rag_type), "disadvantages": disadvantages_for(rag_type), "production_risks": risks_for(rag_type), "engineering_profile": engineering_profile_for(rag_type), "evals": evals_for(rag_type), "observability": ["selected_path", "retrieved_sources", "answer_status", "latency_ms", "cost_estimate"], } if __name__ == "__main__": for scenario in SCENARIOS: print(decision_record(scenario))