GraphRAG in Production — Part 2: Hybrid Retrieval, Query Planning, Evaluation & Scale Templates — 42 Templates & Checklists for the Reference Pipeline, Hybrid Retrieval, Query Planning, Honest Evaluation, Enterprise Scale, and Graph Governance

The graph is built — now it has to answer, prove itself, scale, and be governed. 42 production templates & checklists distilled from Chapters 7–12 and Appendices A–B of GraphRAG in Production: one reference pipeline from corpus to cited answer, hybrid graph + vector + keyword retrieval fused by reciprocal rank fusion, query planning with templates, budgets, and verification, honest evaluation by ablation and cost per correct answer, enterprise scale through incremental indexing and partial refresh, and governance for the facts a graph reveals. PDF + Markdown source.

## One harmless question. 84,000 nodes. Two suppliers who had nothing to do with it.

The question looked harmless in the log: *which suppliers are affected if the Rotterdam parts depot stays closed for another week, and which customer orders depend on those suppliers?* The system did exactly what it had been told to do — matched the word "suppliers" to thousands of nodes, expanded two hops, then three, then four, and returned 84,000 nodes for a question whose correct answer involved eleven suppliers and nineteen orders. The model, handed a mountain of disconnected context, answered confidently — naming two suppliers with no connection to Rotterdam at all. The graph was sound. The data was sound. What was missing was everything in this pack. **GraphRAG in Production — Part 2: Hybrid Retrieval, Query Planning, Evaluation & Scale Templates** is the working second half of the discipline: 42 copy-paste templates and checklists distilled from Chapters 7–12 and Appendices A–B of the book — run one reference pipeline from corpus to cited answer, fuse three retrievers into one honest evidence stream, plan every question before you traverse, prove by ablation whether the graph actually helps, scale by paying for change instead of existence, and govern the facts before a reviewer governs them for you.

Every template carries the book's tested positions built in: stages that are separately testable and separately blameworthy, with the corpus as the source of truth and retrieval returning evidence objects — never prose; routing that separates local, global, and drift questions and logs the chosen mode with every answer; reciprocal rank fusion with its constant of 60 as the robust default, blending only under normalization discipline, and reranking that respects its own ceiling; question classification, identifier-first entity linking with explicit ambiguous and not-found outcomes, a reviewed template registry instead of free-form text-to-Cypher, and hop, fan-out, and node budgets enforced in code; evaluation on a versioned question set with correctness, faithfulness, comprehensiveness, and recall reported separately by class, a calibrated judge, and the three-arm ablation priced by cost per correct answer; incremental indexing from change events, dirty-community refresh, and caches judged by their invalidation path; and access control at node and edge level, filtered at retrieval time so the model never sees a fact the requester may not. Where the book gives a number — its fusion constant, its planner budget defaults, its illustrative ablation table — the template carries the book's number and labels it as the book's, illustrative where the book labels it illustrative. Bracketed fields are yours to fill; nothing here invents a threshold for your system.

### What's inside — 42 templates & checklists in 7 sections

**1. The Reference Pipeline — One Architecture from Corpus to Answer (6)** — Stage Contract & Failure-Ownership Sheet · Search-Mode Routing Sheet · Local Search Specification · Global Search Map-Reduce Specification · Evidence Object & Generator Contract · Latency Budget, Deployment Shape & Replay Sheet

**2. Hybrid Retrieval — Three Retrievers, One Honest Evidence Stream (6)** — Three-Folder Failure Audit & Solo Lane Audit · Reciprocal Rank Fusion Specification · Weighted Blending & Cascade / Escalation Decision Sheet · Reranking Specification · Chunk-Level Deduplication & Merged-Provenance Sheet · Hybrid NOT-Case & Cost-Ledger Decision Sheet

**3. Query Planning — From Question to Traversal (6)** — Question Classification Sheet · Entity Linking Specification · Template Registry & Text-to-Cypher Sheet · Validation & Dry-Run Gate Sheet · Traversal Budget Sheet · Decomposition & Answer-Verification Sheet

**4. Honest Evaluation — Did the Graph Actually Help? (6)** — Question-Set Design Sheet · Four-Lens Metrics Sheet · Judge Calibration Sheet · Ablation Protocol Sheet · Regression Gates in CI Sheet · Worked Verdict & Decision Record

**5. Scale — Fresh, Fast, and Affordable at Enterprise Size (6)** — Three-Breakpoint Measurement Sheet · Incremental Indexing from Change Data Capture Specification · Partial Community Refresh Sheet · Caching Layers & Invalidation Sheet · Latency, Timeouts & Hot-Node Sheet · Multi-Tenancy, Partitioning & Full Cost Model Sheet

**6. Governance — Trust, but Govern (6)** — Sensitive-Join Register · Retrieval-Time Access Control Specification · Redaction & Personal Data Sheet · Lineage, Audit & Retraction Sheet · Poisoning Defence & Intake Sheet · Governance Roles, Review Workflows & Erasure Sheet

**7. Part 2 Gates & Working Artefacts (6)** — Reference Pipeline Gate (Chapter 7 checklist) · Hybrid Retrieval Gate (Chapter 8 checklist) · Planning Gates (Chapter 9 checklist) · Honest Evaluation Gate (Chapter 10 checklist) · Scaling Gates (Chapter 11 checklist) · Part 2 Production Gate — governance checklist, production readiness & quick reference consolidated (the book's six chapter checklists, Appendix A's 18 readiness lines, and Appendix B's 8 rules consolidated across Items 37–42, 74 gate lines)

### Who this is for

AI and platform engineers running GraphRAG in production, search and retrieval engineers who own a hybrid stack, and technical leads who must prove — with arithmetic — what the graph earns, what it costs, and who may see what it reveals. The source book assumes the graph is built (that is Part 1) and that you build production systems; it teaches the query, evaluation, scale, and governance disciplines, not programming. If your system can return a confident answer it cannot replay from a planner log, start at Template 1.

### What this pack is not

- **Not Part 1.** Question classification as an architecture decision, graph fundamentals, extraction, ontology design, store selection, and graph indexing are the book's Chapters 1–6. They live in the Part 1 companion pack; nothing here restates them, and this pack is complete and useful without them.

- **No runnable code, service, or tooling.** The book's Python sketches (a pipeline skeleton, a fusion function, a planner module, an evaluation harness, an incremental indexer, a governed fact filter) are simplified, structurally honest teaching shapes. This pack ships the templates that specify, govern, and operate these components — not a working retrieval service, planner, evaluation harness, or indexer.

- **No vendor selection for you.** Stores, models, and judge approaches are discussed as the book discusses them, and every decision sheet runs on YOUR workload evidence. Nothing is installed, configured, licensed, or endorsed here.

- **No benchmark or results claims.** The book's worked figures — the ablation table, traversal times, cache and cost shapes in its fictional cases — are the book's illustrative examples, labelled as such in every template that carries them. This pack promises no accuracy, latency, cost, or freshness outcome for your system.

- **No legal advice or compliance determination.** The governance chapter's regulatory material is orientation for engineers, in the book's own framing — your counsel owns the final word for your jurisdictions. No compliance crosswalk ships in this pack.

- **No guarantee the graph earns its cost.** The book's own position, built into Templates 12 and 22: hybrid is sometimes overkill, and an honest ablation may route whole question classes away from the graph. These templates include the refusal cases, not just the build.

### Companion products

**GraphRAG in Production — Part 1: Knowledge Graph Build Templates** is this pack's sibling, built from Chapters 1–6 of the same book (question classification, graph fundamentals, extraction, ontology design, store selection, graph indexing): a companion for the build half — never a requirement for this pack to work. This pack stands alone.

*Note: the $99 Architecture Templates Bundle currently contains Products #1–#3 only. This pack launches standalone; no bundle inclusion is promised for this product.*

Built by the author of **GraphRAG in Production**, from the production discipline behind enterprise knowledge-graph systems.

### About the author

Naveen Kumar Josyula is the author of 35 AI books on Amazon KDP and the CEO and founder of AISoltrics Technologies Inc. He is the author of the source book, *GraphRAG in Production*, from which every template in this pack is distilled. A separate KDP book line also carries a GraphRAG title; this pack is distilled from the manuscript named above.

### Support promise

If anything in this file is wrong or unclear, email support@aisoltrics.com before you judge it — we fix verified issues within one business day and send you the corrected file free. (You can also reach the founder's office at aadhya@aisoltrics.com.) Updates to this product are free to past buyers via your library. Sold by AISoltrics Technologies Inc.