1️⃣ Query Classification Type of Knowledge Required Regulatory + procedural knowledge Requires current legal rules and practical implementation steps Internal vs External Knowledge ❌ Internal model knowledge is insufficient ✅ External, authoritative references required Ambiguity Level Medium “Latest” is time-sensitive “Small e-commerce sellers” may have turnover thresholds Still answerable with scoped assumptions 2️⃣ Retrieval Necessity Decision Decision: ✅ Retrieval Required Why Tax laws change frequently Incorrect information creates legal and financial risk Hallucination risk is high without sources User explicitly expects current compliance guidance 3️⃣ Retrieval Scope Control Sources to Query Government GST portals / notifications (PDFs) Trusted tax advisory websites Official GST APIs (if available) Documents / Chunks 5–8 chunks maximum Priority order: Most recent notifications Summary circulars Implementation guides Freshness Requirement Must be within last 3–6 months Older than 1 year = discard unless referenced by newer updates 4️⃣ No-Retrieval Response Rules (If Retrieval Failed) If retrieval fails or data is stale: Safe Answer Strategy Provide high-level framework only Avoid exact thresholds, dates, or penalties Use language like: “As of recent general guidance…” “Sellers should verify with official GST notifications…” Clarifying Question Trigger Ask user: “Are you below or above ₹20/40 lakh turnover?” “Are you selling inter-state or intra-state?” 5️⃣ Fallback & Escalation Logic Retry Retrieval When Sources conflict No post-dated (recent) document found Broaden Search When Only summaries found → include primary notifications Narrow Search When Too many generic articles → restrict to government PDFs Human Review Trigger Conflicting compliance rules Ambiguous legal interpretation High-risk financial advice 6️⃣ Cost vs Accuracy Trade-Off Notes Cost Impact Government PDFs + parsing = higher compute Web scraping increases latency When Higher Cost Is Justified Legal / tax / compliance topics External users making business decisions Lean Retrieval Safer When Explaining concepts (e.g., “What is GST?”) Non-actionable educational content 7️⃣ Hallucination Risk Signals High-Risk Query Types “Latest rules” “Legal compliance” “Penalties / fines” “Exact thresholds or dates” Risk-Increasing Patterns Answering from memory Mixing old and new rules Over-confident tone without citations Guardrails Mandatory retrieval for “latest” Source date validation Explicit uncertainty flags if data conflicts 8️⃣ Decision Traceability Log These Signals Retrieval decision (required / optional / skipped) Sources used + timestamps Freshness validation result Confidence level of final answer Audit Explanation “Retrieval triggered due to fast-changing regulatory domain and high accuracy sensitivity.” Monitor Over Time Retrieval frequency User corrections Regulatory update lag 9️⃣ Example Decision Walkthrough (Condensed) Step 1: Detect “latest” + “GST” → high volatility Step 2: Accuracy sensitivity = High → retrieval mandatory Step 3: Query government PDFs + tax portals Step 4: Filter documents newer than 6 months Step 5: Generate answer with cited, dated guidance Step 6: Add disclaimer for jurisdictional differences Final Response Strategy Bullet-point compliance changes Clear invoicing update checklist Suggest official verification link 🔟 Retrieval Decision Summary ✅ Decision Rules “Latest + Legal + High Risk” → Always Retrieve Procedural steps without volatility → Optional Conceptual explanations → No Retrieval 📌 Best Practices Prefer authoritative sources Limit chunks to reduce noise Validate freshness before answering ❌ Top Mistakes to Avoid Answering regulatory questions from memory Over-retrieving low-quality blogs Hiding uncertainty in confident language 🎯 Outcome This decision layer maximizes trust, reduces hallucinations, and controls cost by retrieving information only when it materially improves correctness.