**Rules of Thumb for Modern Information Retrieval — v2** **Retrieval Architecture** 1. Retrieval is the bottleneck — a hallucinating LLM with perfect retrieval beats a perfect LLM with bad retrieval. 2. Classify query type (navigational, informational, transactional) before you touch an index. It determines your entire retrieval strategy. 3. Always: bi-encoder for retrieval, cross-encoder for reranking top-k. Cross-encoder reranking is the single biggest quality lever in RAG. 4. Use RRF over linear score combination unless you've tuned α on held-out queries — RRF is score-distribution-agnostic. 5. For semi-structured corpora, hybrid metadata filtering + semantic search outperforms pure semantic. Encode metadata as filter pre-conditions, not just embedding dimensions. **Chunking & Indexing** 6. Chunk at retrieval granularity (paragraphs/passages), index at document granularity (parent-child). Retrieve the child, pass the parent to the LLM. 7. Use semantic chunking (split where inter-sentence cosine similarity drops), not fixed-size windows. Overlap is a band-aid for bad boundaries. 8. Scale breakpoints for vector indices: flat <1M vectors, IVF-PQ >10M, binary quantization >100M. **Query Understanding** 9. Use HyDE (generate a hypothetical answer, retrieve against it) to close the query-document distribution gap. 10. SPLADE gives you learned sparse expansion — bridges vocabulary mismatch while staying compatible with inverted index infrastructure. Use it where BM25 fails on synonyms and paraphrases. 11. Wormhole vectors inject categorical/metadata features directly into the dense embedding space, letting exact-match filtering happen inside ANN search instead of as a slow pre/post-filter. 12. MUVERA decomposes complex queries into aspect embeddings and retrieves against each. ~15-20% improvement over single-vector on complex queries (BEIR). Use when queries are multi-faceted. **Retrieval Quality** 13. Set a cosine similarity floor on retrieved chunks. Don't pass junk context to the LLM. 14. ANN, quantization, and representation learning are composable layers, not alternatives. Quantize your HNSW graph for memory-efficient ANN over learned representations. 15. Representation learning has the highest quality ceiling. But chunking, query preprocessing, and behavioral signals have the highest ROI per effort — because most teams haven't done them at all. Do those first. **Behavioral Signals & Click Models** 16. A click on result #8 is more informative than a click on result #1. Always weight by position-corrected propensity. 17. Use Bayesian smoothing (Beta priors on CTR) for query-document pairs with sparse click data. Don't trust signals below a minimum click threshold. 18. SDBN debiased satisfaction scores are your gold-standard pseudo-labels for LTR training. 19. SDBN residuals (where predictions diverge from corrected clicks) reveal missing features. Cluster them to discover what your feature set isn't capturing. 20. Build knowledge graph edges from co-clicked documents and co-queried terms. Embed with TransE/RESCAL. Enables "users who searched X also needed Y" signals. **Learning to Rank** 21. LambdaMART: no GPU, interpretable, fast inference — still the production workhorse. 22. The features that matter most: per-field BM25 scores, bi-encoder similarity, position-corrected CTR, document freshness decay, user behavioral embedding dot product, and query-document term overlap statistics. 23. Don't label randomly — uncertainty sampling gives ~3-5× label efficiency over random. **Production Optimization** 24. Semantic caching: serve cached responses for queries within ε embedding distance. Use an ANN index over recent queries as the cache lookup, not just hashing. Cuts LLM latency 40-60% on repetitive enterprise patterns. 25. For latency-sensitive reranking: distill cross-encoders to <100M params, quantize to INT8, serve with ONNX Runtime. Sub-10ms reranking of top-50 is achievable. --- [[A Practitioner's Reference to Neural IR End-to-End (Landscape)]]