Solr Search Tuning — Field Weights, Minimum Match and Relevance, Without Editing Configs

1. Field weights (qf) — where a match matters most

A match in a title almost always signals more relevance than the same match buried in body text. The edismax query parser expresses this with the qf parameter — a list of fields with multipliers. A battle-tested starting point for content sites:

title^5  tags^3  summary^2  body^1

Symptoms and fixes: “the page that is literally titled what I searched is not first” → raise the title weight. “tag/keyword pages dominate over real content” → lower the tag field weight. Keep ratios modest (1–10 range); a title^50 does not make results 10× better, it makes every other signal irrelevant.

2. Minimum match (mm) — precision vs. recall

mm decides how many of the query's words must match. mm=100% means every word must be present — precise, but a single typo or extra word returns zero results. mm=1 matches any word — never empty, often noisy. The sweet spot for most sites:

2<75%   (all words required up to 2; above that, 75% of them)

Symptoms: “too many zero-result searches” → loosen mm. “searching for solr replication returns everything that mentions solr” → tighten it. Your Query Analytics page shows the zero-result rate, so this is measurable, not a guess.

3. Freshness boost — when newer should win

For news, changelogs, documentation and anything versioned, a relevance tie should break toward the newer document. The tuning panel's freshness boost applies a date-decay function to your date field — gentle by default, so it re-orders near-ties instead of overriding topical relevance. If year-old articles outrank this month's update on the same topic, raise it one step; if evergreen cornerstone pages get buried by minor news, lower it.

4. Hybrid balance — lexical vs. vector

With hybrid search enabled, every query runs both a lexical (keyword) leg and a vector (semantic) leg, blended by a weight you control. The single most common mistake we see in real tuning sessions: setting the lexical weight so low that vector similarity drowns out exact matches — users type an exact phrase that exists word-for-word in a document, and semantically-similar-but-wrong documents outrank it. Rule of thumb: start balanced. Move toward vector only when your users ask conceptual questions (“how do I make search faster”), move toward lexical when they search exact names, codes, or phrases. Test both query styles after every change.

The tuning method that actually works

1. Collect 10 real queries from your Query Analytics — the actual searches your users run, not the ones you imagine.
2. For each, write down which document SHOULD be first. This is your test set.
3. Change one dial, re-run all 10, count improvements vs. regressions.
4. Keep the change only if it wins on net. Repeat.