Matching tells you which rows qualify; ranking tells you which ones to show first. ts_rank and ts_rank_cd both return a float4 relevance score for a (tsvector, tsquery) pair — you ORDER BY it descending. The difference between them is whether they care about where the matched terms sit.
ts_rank — frequency-based
Scores by how often the query lexemes appear and their setweight weights. More occurrences of the search terms → higher score. It ignores how close the terms are to each other.
SELECT title,
ts_rank(search_vec, query) AS rank
FROM docs, websearch_to_tsquery('english', 'postgres index') AS query
WHERE search_vec @@ query
ORDER BY rank DESC
LIMIT 20;ts_rank_cd — cover density
cd = cover density. It additionally rewards documents where the matched lexemes appear close together, using the positions stored in the tsvector. For multi-word queries this usually feels more relevant — a doc with "postgres" and "index" in the same sentence outranks one where they're paragraphs apart. (It requires positional info, so it's meaningless against a stripped vector.)
ts_rank_cd(search_vec, query)Weights make fields count differently
If you built the vector with setweight, pass an array {D, C, B, A} to scale each weight class. Here title (A) counts most:
ts_rank('{0.1, 0.2, 0.4, 1.0}', search_vec, query)Normalization — fight the long-document bias
A raw score grows with document length. The optional normalization flag (bitmask) divides it out — e.g. 32 divides by rank + 1 to bound scores into [0,1), and 1/2 divide by document length:
ts_rank_cd(search_vec, query, 32) -- normalize so long docs don't dominate