pg_recall_guard

CI

Watch vector indexes for recall drift against a baseline you approved.

A degraded ANN index does not fail. It returns k plausible neighbours and never mentions that four of the ten good ones were left out. No error, no log entry, no alert — you find out when a user tells you the search “feels worse”, or when you happen to run a benchmark months later.

This extension measures the recall of any vector index against exact ground truth, records the number you accepted as good, and tells you when it drifts.

CREATE EXTENSION pg_recall_guard;

-- What can be measured? Discovered from the catalog, no configuration.
SELECT index_name, table_name, access_method, operator
  FROM recall_guard.vector_indexes;

--     index_name     | table_name | access_method | operator
-- -------------------+------------+---------------+----------
--  public.items_hnsw | items      | hnsw          | <=>
--  public.docs_dann  | docs       | diskann       | <=>

-- Record what you consider good.
SELECT recall_guard.approve('items_hnsw', p_k => 10, p_sample_size => 100);

-- Later — from pg_cron, from your monitoring, from a shell:
SELECT * FROM recall_guard.check();

--     index_name     | baseline | current |  drift  | verdict
-- -------------------+----------+---------+---------+---------
--  public.items_hnsw |   0.9800 |  0.8100 | -0.1700 | critical

Works with any vector index, because it never names one

Indexes are found through pg_amop.amoppurpose = 'o' — access methods that can answer an ORDER BY on a distance operator. That is a PostgreSQL property, not an extension’s, so it covers hnsw and ivfflat (pgvector), diskann (pgvectorscale), gist and spgist, and whatever ships next, with no change here. The tests exercise hnsw, ivfflat and gist; diskann and spgist are covered by the same catalog property but are not in the test suite.

Measured with bench/ef_search.sh on PostgreSQL 13.23 and pgvector 0.8.6, over a generated dataset of 20,000 vector(64) rows (uniform, fixed seed; hnsw with the default m and ef_construction; k = 10, 100 sampled queries, three runs each; details in bench/results.md):

hnsw.ef_search measured recall
1 0.002 – 0.003
10 0.825 – 0.847
40 (default) 0.910 – 0.914
100 0.949 – 0.954

It is built so it cannot flatter itself

Two failure modes would produce a reassuring 1.0000 that means nothing. Both are checked rather than assumed:

  • The indexed side must actually use the index – this one. If the planner ignores it, the measurement compares the index against itself and always scores perfect; if it picks a twin index on the same column, the number is the twin’s. Since 0.2.6 every relation the plan reads must be read through the index asked about (or, for a partitioned index, one of its partitions), or the function raises and says which index the planner chose. Until 0.2.5 any Index Scan passed, and measure(X) reported the recall of whichever index the planner preferred. An index the planner never chooses for these queries cannot be measured – and is not the one your queries are getting either.
  • The ground truth must not use it. Same reasoning, other direction.

There is a third trap that is easy to miss. Sample queries are drawn from rows of the table itself, and a stored vector always finds itself at distance zero — a free hit in every single query, which with k = 10 means recall can never drop below 0.1 no matter how broken the index is. pg_recall_guard asks for k+1 neighbours and discards the originating ctid from both sides. That is the difference between measuring the index and measuring that a vector equals itself: at ef_search = 1 the naive version cannot report less than 0.1000; this one reports what the index returns, 0.002 in the table above, and test/sql/vector.sql asserts it falls below 0.10.

Recall is judged by distance, ties included (0.2.7). A returned neighbour is a hit if it is no farther than the k-th exact one: with duplicate vectors – a chunk stored twice – an index cannot be asked which of two identical rows to return. Until 0.2.6 rows were matched by identity, and an index exact by distance over 50 vectors stored 40 times each measured 0.3878.

A row is its table and its ctid (0.2.6): a ctid alone repeats across the partitions of a partitioned table, which inflated the recall of a partitioned index. A plain table is read with ONLY: an inheritance child is not covered by the parent’s index, and until 0.2.5 one diluted the measurement of a broken index until it read as healthy. test/audit.sh pins both against a control.

Measuring an index runs nothing its table’s owner wrote (0.2.6). The sampled vector is cast to the column’s base type, not to the column’s type: on a domain column that cast would evaluate the domain’s CHECK constraints – functions their owner may add after the index was approved – as whoever runs check().

It measures the index you approved, from any session

A baseline names its index with the schema (public.items_hnsw), and every function runs with search_path = pg_catalog, pg_temp. So check() measures the same index whether it runs from your session, from pg_cron, or from a session whose search_path lacks the index’s schema. And a temporary table in the checking session cannot take its place: PostgreSQL looks in pg_temp first for any relation whose path does not name it, so until 0.2.3 a temporary index with the approved name was the one measured, and a degraded real index came back ok. test/pg_temp.sh shows both, red on 0.2.3 and green from 0.2.4.

Verdicts

check() returns one row per baseline, and its verdict is one of:

  • ok – the recall is within 0.02 of the baseline;
  • degraded – 0.02 to 0.10 below it;
  • critical – more than 0.10 below it;
  • COULD NOT MEASURE: <error> – the measurement raised (the index was dropped, the planner chose another one, …); current and drift are NULL.

Changed in 0.2.9: up to 0.2.8 the values were Spanish. A monitor that filters on them must be updated:

up to 0.2.8 from 0.2.9
ok ok
degradado degraded
critico critical
NO SE PUDO MEDIR: <error> COULD NOT MEASURE: <error>

What a verdict compares, since 0.2.8

  • The index, under the settings it was approved with. approve() records the search settings (hnsw.*, ivfflat.*, diskann.*) and check() measures under them, so no role can make a degraded index pass with a SET, and every measurement records what it ran under. The other side of that choice: check() measures the index, not the application’s configuration – if someone lowers ivfflat.probes for the whole application, this does not see it.
  • A random sample of the rows that have a vector, not contiguous rows.
  • Only what the caller may read. baselines and measurements show a role only the rows of indexes on tables it may read; the owner (and so pg_dump) sees all. Writing a baseline is still approving one: grant UPDATE on baselines to whom you would let approve.
  • By name. A baseline names its index; an index dropped and created again with the same name, or another schema renamed into its place, inherits it. Re-approve after such a change. A temporary index cannot be approved.
  • check() holds a lock on each table while it measures it, so a schema change waits behind it, and the queries behind that change: run it with lock_timeout set.

Privileges

Out of the box only the owner (and superusers) can use it: the schema’s tables and view are granted to nobody. A role that should run check() from your monitoring needs, besides SELECT on the tables it measures:

GRANT USAGE ON SCHEMA recall_guard TO monitor;
GRANT SELECT ON recall_guard.vector_indexes, recall_guard.baselines TO monitor;
GRANT SELECT, INSERT ON recall_guard.measurements TO monitor;
GRANT USAGE ON SEQUENCE recall_guard.measurements_id_seq TO monitor;

It then sees and measures only the baselines of tables it may read. Approving stays with whom you grant INSERT/UPDATE on baselines. Up to 0.2.9 even these grants were not enough: a helper measure() calls was revoked from PUBLIC (fixed in 0.2.10; make check-audit RG-S1).

Cost

check() runs one exact sequential scan per sampled query, so it is not free and it is not meant to run on every request. It is meant to run on a schedule, on a sample, the way you would run ANALYZE — the degradation it looks for is gradual and silent, so the only moment it gets caught is when something bothers to look.

Start with p_sample_size => 30 and raise it if the numbers move around between runs.

Tested on

0.2.10 was measured on 2026-11-12, not assumed, on local builds of PostgreSQL 10.23, 11.22, 12.22, 13.23, 14.24, 15.19, 16.15, 17.11, 18.6 and 19rc1: make installcheck and ci/upgrade_check.sh from every earlier release (0.1.0 to 0.2.9) on each; make installcheck-vector (pgvector 0.8.6) on 13 to 19, where pgvector was built; and make check-suites (check-pgtemp, check-recall, check-audit) on 18 and 19. All green. The CI in .github/workflows runs installcheck and the upgrade check on 10 to 19, and the suites on 18 and 19.

10 11 12 13 14 15 16 17 18 19
✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

10 to 13 no longer receive updates from the PostgreSQL project.

Install

pgxn install pg_recall_guard

Or from source:

make install    # or: make PG_CONFIG=/path/to/pg_config install

Then, in the database you want to watch:

CREATE EXTENSION pg_recall_guard;

No other extension is required. Up to 0.2.9 it required tsm_system_rows, which measure() stopped using in 0.2.8; an installation created then keeps that dependency until you drop it. Measuring needs a vector index, of course: pgvector, pgvectorscale or anything else whose access method can order by a distance.

Prior art, and where this differs

VectorChord 0.5 added vchordrq_evaluate_query_recall and query sampling — a good design, and this extension borrows the shape of it. But by their own documentation it is “not supported by vchordg”: it works for their IVF+RaBitQ index only. That is a rational choice for a vendor and a gap for everyone else, since each vector extension has an incentive to measure its own index and none has an incentive to measure the others'.

pg_recall_guard is the neutral version: it belongs to no index, so it can watch all of them.

Tests

make installcheck            # no vector extension required
make installcheck-vector     # end-to-end, requires pgvector
make check-suites            # PostgreSQL 18+, requires pgvector: a throwaway cluster
                             # on port 5496 (PGPORT overrides), stopped afterwards

They are split on purpose. installcheck covers what must hold on any PostgreSQL — the extension loads, discovery invents nothing, and pointing it at a plain btree fails with a message that says where to look. installcheck-vector builds a real HNSW index and asserts the thing that matters: that recall collapses when the index is throttled. If it did not, the tool would not be measuring anything and every other number it prints would be decorative.

Assertions are ranges, not exact values, because measure() samples queries at random — a test demanding 0.9000 would fail every few runs, and a test that fails for reasons that are not failures gets ignored.

installcheck has the same expected output on every version in Tested on.

Maturity

Released as stable since 0.2.0 — the test suite is what earned the label, not time in the field. What that means concretely: the mechanism is tested, including the assertion that recall collapses when an index is throttled, and it is green on PostgreSQL 10 to 19 with identical expected output.

What it does not mean, and is worth knowing before you rely on it:

  • The recall figures above come from one machine and one generated dataset (20,000 vectors, 64 dimensions, hnsw). Real embeddings, other dimensions, m values and access methods will give other numbers: run bench/ef_search.sh, or measure() itself, on yours.
  • Nobody has run this against a production workload. The 0.1.0 release was testing, which meant pgxn install skipped it by default and nobody could find it — so field reports had no way of arriving. That is the main reason this release is stable.
  • The sampling strategy — drawing query vectors from rows of the table — is a reasonable proxy for real traffic, not real traffic. Capturing actual queries needs an executor hook in C, which this does not have.

Treat the recall numbers it reports as trustworthy and its coverage as narrow. Reports from other datasets, dimensions and index parameters are the single most useful thing anyone could send.

License

Apache License 2.0 – see LICENSE. Copyright 2026 Manuel Reyes Bravo.

The name is not licensed with the code: see TRADEMARK.md.