Limitations

• One column per index — Create a separate TIN index for each searchable column (or expression). Combine columns in a query with multiple ==> predicates. tin.score(ctid) combines scores across those fields.
• Partitioned tables — TIN indexes on partitioned tables are supported, and querying the parent works when every partition has a TIN index. BM25 scores are computed per partition (each partition has its own corpus statistics), so ordering by score across partitions mixes scores from different subsets. Some query shapes that pass tin.score through aggregates or window functions against the parent are refused with an error that names the workaround: compute the score in a MATERIALIZED CTE (or query a single partition) and consume it above.
• Index relation size — The index relation grows to a high-water mark and does not shrink on its own. Space freed by deleted and updated rows is reused internally, and REINDEX is the only way to shrink it. Measure it with pg_relation_size('<index name>'). CREATE INDEX CONCURRENTLY, REINDEX, and REINDEX CONCURRENTLY are supported.
• Tokenizer — Tokenization is configured per index via WITH (…) (see Index options). After changing analysis options on a populated index, REINDEX so existing rows are re-tokenized. tin.tokenize accepts the same arguments.
• Read replicas — On physical standbys, set hot_standby_feedback = on. Standby results are exact. Under heavy replay a query can be cancelled with SQLSTATE 40001 (serialization failure) and should be retried.

​

Need help?