Skip to content
Multigrid

Search, ranking & recommendation

How a ranked list is built, measured and defended — the evaluation mathematics, the tuning loop and the product decisions that sit on top of whatever retrieval engine you chose.

Search looks like a solved problem from the outside and is not, because the hard part was never retrieval. Retrieval is a well-understood engineering job with good open-source answers. The hard part is deciding which of two orderings is better, proving it with enough confidence to ship, and doing that repeatedly without the metric you optimise quietly drifting away from the thing you care about.

So these pages are mostly about measurement and judgement. Where an equation decides something — the logarithmic discount in nDCG, the saturation term in BM25, the examination probability that makes raw click-through rate uninterpretable, the variance reduction that makes a paired experiment sensitive — the equation is written out, its terms defined, and one example worked by hand with the inputs labelled as assumptions. Substitute your own and the conclusion moves with them.

Embeddings, vector indexes and cross-encoder rerankers are covered in the embeddings cluster and the RAG cluster. This one assumes you have them and asks whether the list they produce is any good.

How Modern Search Works: Retrieve, Rank, Rerank

The three-stage cascade every large search system converges on, why the arithmetic forces it, and the recall ceiling that stage one silently imposes on everything downstream.

6 min read

Relevance Tuning: Making Search Feel Right

The judgement list, the BM25 parameters worth touching, and the sample-size arithmetic that tells you whether a tuning change is real or noise.

6 min read

Measuring Search Quality: nDCG, MRR and Click Metrics

nDCG, MRR and MAP computed by hand on one ranking, and the examination model that explains why a raw click-through rate cannot be compared between positions.

7 min read

Query Understanding: Intent, Entities and Spelling

The stages that run before a single document is retrieved — normalisation, correction, segmentation, entity extraction, intent — and the noisy-channel arithmetic behind spelling correction.

6 min read

Faceted Search and Filters That Don’t Return Zero Results

How facet counts are computed, the multiplicative arithmetic that makes empty result sets inevitable, and the recovery ladder that stops a filter combination becoming a dead end.

6 min read

Autocomplete and Typeahead That People Trust

The latency budget a suggestion has to fit inside, the index structure that makes it possible, and the two safety rules that keep query-log suggestions from embarrassing you.

6 min read

Personalisation vs Relevance: The Trade-off

Why blending a personal signal into a ranking can only help on queries that are genuinely ambiguous, how click entropy tells you which those are, and how the feedback loop closes on itself.

6 min read

Recommendation Systems: Collaborative vs Content-Based

The two families, an item-item similarity computed by hand, and the popularity bias that quietly turns any collaborative system into a bestseller list.

6 min read

The Cold Start Problem

Three different cold starts, the shrinkage estimator that stops one click looking like a hit, and the bandit arithmetic that decides how much traffic exploration deserves.

6 min read

Embeddings for Recommendations: The Two-Tower Model

Why retrieval-scale recommendation converged on two independent encoders, the sampled-softmax correction that stops popular items being over-penalised, and what the architecture structurally cannot do.

6 min read

LLMs in Search: Where They Help and Where They Don’t

The four places a language model can sit in a search pipeline, the published work on LLM rerankers, and the token arithmetic that decides which of those places is affordable.

6 min read

Generative Answers vs Ten Blue Links

What a generated answer costs per query compared with a list of links, which query types each format suits, and the attribution failure that makes the generated version harder to trust.

6 min read

Search Result Diversity and Filter Bubbles

Why ranking purely by relevance produces near-duplicate lists, how maximal marginal relevance fixes it, and what the diversity parameter costs when the query is unambiguous.

6 min read

A/B Testing Search Changes

Why a conventional A/B test is a blunt instrument for a ranking change, the variance arithmetic that makes a paired design sharper, and how team-draft interleaving builds the pairing.

6 min read

Handling Long-Tail and Zero-Result Queries

The shape of a query distribution derived from a Zipf assumption, the arithmetic that turns a zero-result rate into a bound on what it costs you, and the failure that is worse than zero results.

6 min read

Multilingual Search Without Separate Indexes

Why mixing languages in one index corrupts BM25's own statistics, what per-language fields fix, and the analyzer traps that only appear in production.

6 min read

E-commerce Search: Attributes, Stock and Intent

Why product search is a different problem from document search: the ranking objective is expected value, the text is too short to score, and near-miss substitution is a real failure with a real cost.

6 min read

Enterprise Search: Permissions and Access Control

Why filtering results by permission after retrieval destroys recall, the arithmetic that shows how badly, and the index design and staleness handling that make pre-filtering safe.

6 min read

Search Latency: Where the Milliseconds Go

A latency budget allocated across the whole pipeline from a stated p95 target, and the fan-out arithmetic that explains why a system of fast components has a slow tail.

6 min read

Building Search on a Small Budget

A working full-text stack on Postgres you already run, the ranking limitation in its built-in functions that nobody mentions, and an evaluation script so you can measure what adding a reranker buys.

6 min read

Other topics

Search, ranking & recommendation · Multigrid