Hybrid Search Engineering — Part 1: BM25, ANN Recall & Rank Fusion — 42 Templates & Checklists for BM25 & the Keyword Arm, ANN Recall Contracts, Rank Fusion, Query Understanding & Reranking Cascades

Your vector migration fixed the paraphrases and broke the part numbers — this is the working half of hybrid search that fixes both. 42 production templates & checklists distilled from Chapters 1–6 of Hybrid Search Engineering: the three-question query decomposition and four-class taxonomy, BM25 taken apart (index, dials, analyzer), the ANN recall contract fusion demands, rank fusion without the raw-score error, the governed query-understanding layer, and the reranking cascade. PDF + Markdown source.

The bulletin whose title was the query ranked fourteenth.

Nothing had malfunctioned. A field technician searched "seal kit SK-2204 exploded view" on a service portal that had just migrated from keywords to vectors — migration metrics all green — and got three beautifully written articles about seal maintenance philosophy. The embedding model had never been told that "SK-2204" was a name rather than a topic, and embeddings are paid to find topics. Meanwhile the queries the old keyword engine used to fail — "pump sounds like gravel" — now sailed through. The portal had not replaced a bad search with a good one; it had traded one half-answer for the other. Hybrid Search Engineering — Part 1 is the working discipline for refusing that trade: 42 copy-paste templates and checklists distilled from Chapters 1–6 of the dedicated hybrid-search book — decompose every query into its three questions, classify your real log into four classes and price the errors per class, take the keyword arm apart down to its analyzer, hold the vector arm to a measured recall contract, fuse the two arms' evidence without adding scores that were never on the same scale, interrogate the query before retrieval without ever substituting it, and spend expensive reading only on the finalists.

Every template carries the book's tested positions built in: a one-week audit of 200 real queries before any architecture is chosen, BM25 understood as three moves — rarity priced by IDF, repetition bent by the k1 saturation curve, acreage taxed by b — with the defaults (k1 near 1.2 to 1.5, b near 0.75) treated as other people's corpora until your judged set says otherwise, an analyzer that protects identifier-shaped tokens instead of splitting them, recall measured against brute-force ground truth at the depth fusion consumes (the book's contract shape: ≥0.95 at depth 100, no class below 0.90), reciprocal rank fusion at the book's conventional k=60 with its blindnesses named, the absolute spelling rule — never correct a token that exists in your dictionary — rewrites and HyDE running as logged siblings beside the untouched original, and a reranking seam sized from the worst class, defended as cost per correct top result. Where the book gives an example — its contract floors, its worked score ranges, its fictional cases — the template labels it as the book's example for you to tune. Bracketed fields are yours to fill; nothing here invents a threshold for your corpus.

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

1. Two Half-Answers — Why Neither Arm Alone Survives Production (6) — Three-Question Query Decomposition Sheet · Exact-Token Blind-Spot Audit · Vocabulary-Mismatch & Intent Audit · Four-Class Query Taxonomy Worksheet · Per-Class Error-Pricing Worksheet · One-Week Hybrid Audit Protocol

2. The Old Engine, Taken Apart — BM25, Indexes & Analyzers (6) — Inverted-Index Anatomy & Audit Sheet · BM25 Three-Moves Hand-Recomputation Worksheet · k1 & b Tuning Record · Analyzer Chain Decision Record · BM25+ / BM25L Repair & Failure-Mode Diagnostic Sheet · Fields, Field Boosts & Proximity Worksheet

3. Recall Is a Contract — ANN, Candidate Depth & the Vector Arm (6) — Recall@Depth & Ground-Truth Harness Sheet · HNSW vs IVF Dial Map · Vector-Arm Recall Contract · Candidate-Depth, Overlap & Amplification Worksheet · Quantisation Tax Audit · Embedding-Model & Fresh-Vectors Contract Sheet

4. Fused Properly — Rank Fusion, Blending & Learned Combination (6) — Currency-Problem Audit · Reciprocal Rank Fusion Worksheet · Weighted-Blending Worksheet · Voting-Family Selector · Learned-Fusion Readiness Sheet · Fusion List-Hygiene & Provenance Record

5. Interrogate the Query First — Classification, Expansion & Rewriting (6) — Shape Classifier Specification · Governed Expansion & Synonym File Sheet · Do-No-Harm Spelling Rule Sheet · Rewrite, HyDE & Decomposition Sibling-Run Log · Per-Class Routing Table · Session, Freshness & Personalisation Guardrail Sheet

6. The Expensive Second Look — Reranking Cascades (6) — Two-Stage Cascade Design Sheet · Reranker Family Selector · Seam-Sizing Worksheet · Reader Authority & Interface Evidence Sheet · Cascade Economics Sheet · Reader Failure-Mode, Training & Cold-Start Sheet

7. Part 1 Gates & Working Artefacts (6) — Failure Autopsy & Demo-Trap Evidence Sheet · Fusion Ablation & Arm-Health Bench Record · Query-Layer Transformation Audit Log · Old-Engine Audit Consolidation Sheet · Configuration Version Stamp · Part 1 Production Hybrid Gate — the book's six chapter checklists consolidated into one evidence-backed gate (48 lines)

Who this is for

Search and retrieval engineers whose vector migration fixed semantics and broke exact match, RAG builders whose assistant aces paraphrases and fails on part numbers, error codes, and names, and platform architects deciding how keyword, vector, and structured evidence should actually be fused. The source book assumes you have shipped, or seriously studied, a retrieval system; no machine-learning research background is required. If nobody on your team can say what your vector arm's recall is at the depth your fusion layer reads — start at Template 13.

What this pack is not

Not Part 2. Structured search and engine-native hybrid, the filtering layer (facets, permissions, pushdown), per-class evaluation with the golden set and regression gate, latency budgets, the production reference architecture, and operations are the book's Chapters 7–12 and Appendices A–B. They live in the Part 2 companion pack; nothing here restates them, and this pack is complete and useful without them.

Not the RAG pipeline, and not a vector-database manual. Chunking, generation, and grounding belong to the RAG Architecture book (Product #2); vector-index construction and storage internals belong to the Vector Databases book (Product #9). This pack is the dedicated hybrid-search book's method — BM25 at derivation depth, recall as a contract, fusion mathematics, the governed query layer, cascade economics — and restates neither earlier pack.

No runnable code or tooling. The book's chapters demonstrate each method in standard-library Python sketches; the pack ships the templates that specify, audit, and operate these components — not a search engine, fusion library, classifier, or reranker.

No vendor or engine guidance. No search engine or vector product is selected, installed, configured, or benchmarked here. Engine-native hybrid (Postgres, OpenSearch, Elasticsearch) is Chapter 7 — Part 2 territory.

No guarantee of relevance. No ranking improvement, conversion, or cost saving is promised for your corpus. The book's own discipline — and every template's — is that each method proves itself on your judged queries, per class, or it does not ship.

Companion products

Hybrid Search Engineering — Part 2: Filtering, Evaluation, Latency Budgets & Operations Templates is this pack's sibling, built from Chapters 7–12 and Appendices A–B of the same book: a companion for when you are ready for the second half — never a requirement for this pack to work. The RAG Architecture Templates pack (Product #2) and the Vector Databases Templates pack (Product #9) are neighbours from different books — the pipeline this pack's retrievers plug into, and the substrate underneath the vector arm. 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 Hybrid Search Engineering: Vector + Keyword + Structured Retrieval, Fused Properly, from the hybrid-search discipline behind production retrieval 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, Hybrid Search Engineering: Vector + Keyword + Structured Retrieval, Fused Properly, from which every template in this pack is distilled.

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.