Align values [skip ci]

This reverts commit 3f674c9994.

CHANGELOG.md

@@ -1,9 +1,6 @@
 ## 0.7.0 (unreleased)
 
-- Added support for binary vectors to HNSW
+- Added `sparsevec` type
-- Added `hamming_distance` function
-- Added `jaccard_distance` function
-- Added `quantize_binary` function
 
 ## 0.6.2 (2024-03-18)
 

Makefile

@@ -3,8 +3,8 @@ EXTVERSION = 0.6.2
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/bitvector.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
+OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
-HEADERS = src/vector.h
+HEADERS = src/sparsevec.h src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))

Makefile.win

@@ -1,8 +1,8 @@
 EXTENSION = vector
 EXTVERSION = 0.6.2
 
-OBJS = src\bitvector.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
+OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj
-HEADERS = src\vector.h
+HEADERS = src\sparsevec.h src\vector.h
 
 REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
 REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)

README.md

@@ -5,7 +5,7 @@ Open-source vector similarity search for Postgres
 Store your vectors with the rest of your data. Supports:
 
 - exact and approximate nearest neighbor search
-- L2 distance, inner product, cosine distance, and more
+- L2 distance, inner product, and cosine distance
 - any [language](#languages) with a Postgres client
 
 Plus [ACID](https://en.wikipedia.org/wiki/ACID) compliance, point-in-time recovery, JOINs, and all of the other [great features](https://www.postgresql.org/about/) of Postgres
@@ -221,19 +221,7 @@ Cosine distance
 CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
 ```
 
-Hamming distance - added in 0.7.0
+Vectors with up to 2,000 dimensions can be indexed.
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding bit_hamming_ops);
-```
-
-Jaccard distance - added in 0.7.0
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding bit_jaccard_ops);
-```
-
-Vectors with up to 2,000 dimensions can be indexed, or bit vectors with up to 64,000 dimensions.
 
 ### Index Options
 
@@ -427,7 +415,19 @@ You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python
 
 ### Tuning
 
-Use a tool like [PgTune](https://pgtune.leopard.in.ua/) to set initial values for Postgres server parameters.
+Use a tool like [PgTune](https://pgtune.leopard.in.ua/) to set initial values for Postgres server parameters. For instance, `shared_buffers` should typically be 25% of the server’s memory. You can find the config file with:
+
+```sql
+SHOW config_file;
+```
+
+And check individual settings with:
+
+```sql
+SHOW shared_buffers;
+```
+
+Be sure to restart Postgres for changes to take effect.
 
 ### Loading
 
@@ -523,6 +523,26 @@ Scale vertically by increasing memory, CPU, and storage on a single instance. Us
 
 Scale horizontally with [replicas](https://www.postgresql.org/docs/current/hot-standby.html), or use [Citus](https://github.com/citusdata/citus) or another approach for sharding ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/citus.py)).
 
+## Sparse Vectors
+
+Create a sparse vector column with 10 dimensions
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding sparsevec(10));
+```
+
+Insert vectors
+
+```sql
+INSERT INTO items (embedding) VALUES ('{0:1,1:2,2:3}/10'), ('{0:4,1:5,2:6}/10');
+```
+
+Get the nearest neighbors by L2 distance
+
+```sql
+SELECT * FROM items ORDER BY embedding <-> '{0:3,1:1,2:2}/10' LIMIT 5;
+```
+
 ## Languages
 
 Use pgvector from any language with a Postgres client. You can even generate and store vectors in one language and query them in another.
@@ -683,6 +703,8 @@ ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
 
 Results are limited by the size of the dynamic candidate list (`hnsw.ef_search`). There may be even less results due to dead tuples or filtering conditions in the query. We recommend setting `hnsw.ef_search` to at least twice the `LIMIT` of the query. If you need more than 500 results, use an IVFFlat index instead.
 
+Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
+
 #### Why are there less results for a query after adding an IVFFlat index?
 
 The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
@@ -693,11 +715,13 @@ DROP INDEX index_name;
 
 Results can also be limited by the number of probes (`ivfflat.probes`).
 
+Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
+
 ## Reference
 
 ### Vector Type
 
-Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a single precision floating-point number (like the `real` type in Postgres), and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 16,000 dimensions.
+Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a single-precision floating-point number (like the `real` type in Postgres), and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 16,000 dimensions.
 
 ### Vector Operators
 
@@ -718,31 +742,16 @@ cosine_distance(vector, vector) → double precision | cosine distance |
 inner_product(vector, vector) → double precision | inner product |
 l2_distance(vector, vector) → double precision | Euclidean distance |
 l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
-quantize_binary(vector) → bit | quantize | 0.7.0
 vector_dims(vector) → integer | number of dimensions |
 vector_norm(vector) → double precision | Euclidean norm |
 
-### Aggregate Functions
+### Vector Aggregate Functions
 
 Function | Description | Added
 --- | --- | ---
 avg(vector) → vector | average |
 sum(vector) → vector | sum | 0.5.0
 
-### Bit Operators
-
-Operator | Description | Added
---- | --- | ---
-<~> | Hamming distance | 0.7.0
-<%> | Jaccard distance | 0.7.0
-
-### Bit Functions
-
-Function | Description | Added
---- | --- | ---
-hamming_distance(bit, bit) → double precision | Hamming distance | 0.7.0
-jaccard_distance(bit, bit) → double precision | Jaccard distance | 0.7.0
-
 ## Installation Notes - Linux and Mac
 
 ### Postgres Location

sql/vector--0.6.2--0.7.0.sql

@@ -1,31 +1,95 @@
 -- complain if script is sourced in psql, rather than via CREATE EXTENSION
 \echo Use "ALTER EXTENSION vector UPDATE TO '0.7.0'" to load this file. \quit
 
-CREATE FUNCTION quantize_binary(vector) RETURNS bit
+CREATE TYPE sparsevec;
+
+CREATE FUNCTION sparsevec_in(cstring, oid, integer) RETURNS sparsevec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
+CREATE FUNCTION sparsevec_out(sparsevec) RETURNS cstring
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
+CREATE FUNCTION sparsevec_typmod_in(cstring[]) RETURNS integer
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE OPERATOR <~> (
+CREATE FUNCTION sparsevec_recv(internal, oid, integer) RETURNS sparsevec
-	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-	COMMUTATOR = '<~>'
+
+CREATE FUNCTION sparsevec_send(sparsevec) RETURNS bytea
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE TYPE sparsevec (
+	INPUT     = sparsevec_in,
+	OUTPUT    = sparsevec_out,
+	TYPMOD_IN = sparsevec_typmod_in,
+	RECEIVE   = sparsevec_recv,
+	SEND      = sparsevec_send,
+	STORAGE   = external
 );
 
-CREATE OPERATOR <%> (
+CREATE FUNCTION l2_distance(sparsevec, sparsevec) RETURNS float8
-	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-	COMMUTATOR = '<%>'
+
+CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_negative_inner_product(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec(sparsevec, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE CAST (sparsevec AS sparsevec)
+	WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (sparsevec AS vector)
+	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (vector AS sparsevec)
+	WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
+
+CREATE OPERATOR <-> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l2_distance,
+	COMMUTATOR = '<->'
 );
 
-CREATE OPERATOR CLASS bit_hamming_ops
+CREATE OPERATOR <#> (
-	FOR TYPE bit USING hnsw AS
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_negative_inner_product,
-	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
+	COMMUTATOR = '<#>'
-	FUNCTION 1 hamming_distance(bit, bit);
+);
 
-CREATE OPERATOR CLASS bit_jaccard_ops
+CREATE OPERATOR <=> (
-	FOR TYPE bit USING hnsw AS
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = cosine_distance,
-	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
+	COMMUTATOR = '<=>'
-	FUNCTION 1 jaccard_distance(bit, bit);
+);
+
+CREATE OPERATOR CLASS sparsevec_l2_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <-> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
+
+CREATE OPERATOR CLASS sparsevec_ip_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+
+CREATE OPERATOR CLASS sparsevec_cosine_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 2 sparsevec_norm(sparsevec);

sql/vector.sql

@@ -1,7 +1,7 @@
 -- complain if script is sourced in psql, rather than via CREATE EXTENSION
 \echo Use "CREATE EXTENSION vector" to load this file. \quit
 
--- type
+-- vector type
 
 CREATE TYPE vector;
 
@@ -29,7 +29,7 @@ CREATE TYPE vector (
 	STORAGE   = external
 );
 
--- functions
+-- vector functions
 
 CREATE FUNCTION l2_distance(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -58,10 +58,7 @@ CREATE FUNCTION vector_sub(vector, vector) RETURNS vector
 CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION quantize_binary(vector) RETURNS bit
+-- vector private functions
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
--- private functions
 
 CREATE FUNCTION vector_lt(vector, vector) RETURNS bool
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -102,7 +99,7 @@ CREATE FUNCTION vector_avg(double precision[]) RETURNS vector
 CREATE FUNCTION vector_combine(double precision[], double precision[]) RETURNS double precision[]
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
--- aggregates
+-- vector aggregates
 
 CREATE AGGREGATE avg(vector) (
 	SFUNC = vector_accum,
@@ -120,7 +117,7 @@ CREATE AGGREGATE sum(vector) (
 	PARALLEL = SAFE
 );
 
--- cast functions
+-- vector cast functions
 
 CREATE FUNCTION vector(vector, integer, boolean) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -140,7 +137,7 @@ CREATE FUNCTION array_to_vector(numeric[], integer, boolean) RETURNS vector
 CREATE FUNCTION vector_to_float4(vector, integer, boolean) RETURNS real[]
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
--- casts
+-- vector casts
 
 CREATE CAST (vector AS vector)
 	WITH FUNCTION vector(vector, integer, boolean) AS IMPLICIT;
@@ -160,7 +157,7 @@ CREATE CAST (double precision[] AS vector)
 CREATE CAST (numeric[] AS vector)
 	WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;
 
--- operators
+-- vector operators
 
 CREATE OPERATOR <-> (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l2_distance,
@@ -243,7 +240,7 @@ CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnswhandler;
 
 COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
 
--- opclasses
+-- vector opclasses
 
 CREATE OPERATOR CLASS vector_ops
 	DEFAULT FOR TYPE vector USING btree AS
@@ -291,30 +288,109 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
 
--- bit functions
+--- sparsevec type
 
-CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
+CREATE TYPE sparsevec;
+
+CREATE FUNCTION sparsevec_in(cstring, oid, integer) RETURNS sparsevec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
+CREATE FUNCTION sparsevec_out(sparsevec) RETURNS cstring
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE OPERATOR <~> (
+CREATE FUNCTION sparsevec_typmod_in(cstring[]) RETURNS integer
-	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-	COMMUTATOR = '<~>'
+
+CREATE FUNCTION sparsevec_recv(internal, oid, integer) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_send(sparsevec) RETURNS bytea
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE TYPE sparsevec (
+	INPUT     = sparsevec_in,
+	OUTPUT    = sparsevec_out,
+	TYPMOD_IN = sparsevec_typmod_in,
+	RECEIVE   = sparsevec_recv,
+	SEND      = sparsevec_send,
+	STORAGE   = external
 );
 
-CREATE OPERATOR <%> (
+-- sparsevec functions
-	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
+
-	COMMUTATOR = '<%>'
+CREATE FUNCTION l2_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- sparsevec private functions
+
+CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_negative_inner_product(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- sparsevec cast functions
+
+CREATE FUNCTION sparsevec(sparsevec, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- sparsevec casts
+
+CREATE CAST (sparsevec AS sparsevec)
+	WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (sparsevec AS vector)
+	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (vector AS sparsevec)
+	WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
+
+-- sparsevec operators
+
+CREATE OPERATOR <-> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l2_distance,
+	COMMUTATOR = '<->'
 );
 
-CREATE OPERATOR CLASS bit_hamming_ops
+CREATE OPERATOR <#> (
-	FOR TYPE bit USING hnsw AS
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_negative_inner_product,
-	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
+	COMMUTATOR = '<#>'
-	FUNCTION 1 hamming_distance(bit, bit);
+);
 
-CREATE OPERATOR CLASS bit_jaccard_ops
+CREATE OPERATOR <=> (
-	FOR TYPE bit USING hnsw AS
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = cosine_distance,
-	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
+	COMMUTATOR = '<=>'
-	FUNCTION 1 jaccard_distance(bit, bit);
+);
+
+-- sparsevec opclasses
+
+CREATE OPERATOR CLASS sparsevec_l2_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <-> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
+
+CREATE OPERATOR CLASS sparsevec_ip_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+
+CREATE OPERATOR CLASS sparsevec_cosine_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 2 sparsevec_norm(sparsevec);

src/bitvector.c

@@ -1,90 +0,0 @@
-#include "postgres.h"
-
-#include "bitvector.h"
-#include "port/pg_bitutils.h"
-#include "utils/varbit.h"
-
-#if PG_VERSION_NUM >= 160000
-#include "varatt.h"
-#endif
-
-/*
- * Allocate and initialize a new bit vector
- */
-VarBit *
-InitBitVector(int dim)
-{
-	VarBit	   *result;
-	int			size;
-
-	size = VARBITTOTALLEN(dim);
-	result = (VarBit *) palloc0(size);
-	SET_VARSIZE(result, size);
-	VARBITLEN(result) = dim;
-
-	return result;
-}
-
-/*
- * Ensure same number of bits
- */
-static inline void
-CheckBitLengths(uint32 aLen, uint32 bLen)
-{
-	if (aLen != bLen)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("different bit lengths %u and %u", aLen, bLen)));
-}
-
-/*
- * Get the Hamming distance between two bit strings
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(hamming_distance);
-Datum
-hamming_distance(PG_FUNCTION_ARGS)
-{
-	VarBit	   *a = PG_GETARG_VARBIT_P(0);
-	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		distance = 0;
-
-	CheckBitLengths(VARBITLEN(a), VARBITLEN(b));
-
-	/* TODO Improve performance */
-	for (uint32 i = 0; i < VARBITBYTES(a); i++)
-		distance += pg_number_of_ones[ax[i] ^ bx[i]];
-
-	PG_RETURN_FLOAT8((double) distance);
-}
-
-/*
- * Get the Jaccard distance between two bit strings
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(jaccard_distance);
-Datum
-jaccard_distance(PG_FUNCTION_ARGS)
-{
-	VarBit	   *a = PG_GETARG_VARBIT_P(0);
-	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		ab = 0;
-	uint64		aa;
-	uint64		bb;
-
-	CheckBitLengths(VARBITLEN(a), VARBITLEN(b));
-
-	/* TODO Improve performance */
-	for (uint32 i = 0; i < VARBITBYTES(a); i++)
-		ab += pg_number_of_ones[ax[i] & bx[i]];
-
-	if (ab == 0)
-		PG_RETURN_FLOAT8(1);
-
-	aa = pg_popcount((char *) ax, VARBITBYTES(a));
-	bb = pg_popcount((char *) bx, VARBITBYTES(b));
-
-	PG_RETURN_FLOAT8(1 - (ab / ((double) (aa + bb - ab))));
-}

src/bitvector.h

@@ -1,8 +0,0 @@
-#ifndef BITVECTOR_H
-#define BITVECTOR_H
-
-#include "utils/varbit.h"
-
-VarBit	   *InitBitVector(int dim);
-
-#endif

src/hnsw.h

@@ -17,6 +17,7 @@
 #endif
 
 #define HNSW_MAX_DIM 2000
+#define HNSW_MAX_NNZ 1000
 
 /* Support functions */
 #define HNSW_DISTANCE_PROC 1
@@ -55,6 +56,12 @@
 #define HNSW_UPDATE_ENTRY_GREATER 1
 #define HNSW_UPDATE_ENTRY_ALWAYS 2
 
+typedef enum HnswType
+{
+	HNSW_TYPE_VECTOR,
+	HNSW_TYPE_SPARSEVEC
+}			HnswType;
+
 /* Build phases */
 /* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
 #define PROGRESS_HNSW_PHASE_LOAD		2
@@ -242,6 +249,7 @@ typedef struct HnswBuildState
 	Relation	index;
 	IndexInfo  *indexInfo;
 	ForkNumber	forkNum;
+	HnswType	type;
 
 	/* Settings */
 	int			dimensions;
@@ -262,7 +270,6 @@ typedef struct HnswBuildState
 	HnswGraph  *graph;
 	double		ml;
 	int			maxLevel;
-	Vector	   *normvec;
 
 	/* Memory */
 	MemoryContext graphCtx;
@@ -367,7 +374,9 @@ typedef struct HnswVacuumState
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
-bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+HnswType	HnswGetType(Relation index);
+bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type);
+void		HnswCheckValue(Datum value, HnswType type);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);

src/hnswbuild.c

@@ -44,7 +44,6 @@
 #include "access/xact.h"
 #include "access/xloginsert.h"
 #include "catalog/index.h"
-#include "catalog/pg_type_d.h"
 #include "commands/progress.h"
 #include "hnsw.h"
 #include "miscadmin.h"
@@ -487,10 +486,13 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
+	/* Check value */
+	HnswCheckValue(value, buildstate->type);
+
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
+		if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
 			return false;
 	}
 
@@ -673,20 +675,21 @@ static void
 InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo, ForkNumber forkNum)
 {
 	int			maxDimensions = HNSW_MAX_DIM;
-	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
-
-	if (typid == BITOID || typid == VARBITOID)
-		maxDimensions *= 32;
 
 	buildstate->heap = heap;
 	buildstate->index = index;
 	buildstate->indexInfo = indexInfo;
 	buildstate->forkNum = forkNum;
+	buildstate->type = HnswGetType(index);
 
 	buildstate->m = HnswGetM(index);
 	buildstate->efConstruction = HnswGetEfConstruction(index);
 	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
 
+	/* No limit on sparse vector dimensions */
+	if (buildstate->type == HNSW_TYPE_SPARSEVEC)
+		maxDimensions = INT_MAX;
+
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)
 		elog(ERROR, "column does not have dimensions");
@@ -710,9 +713,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
 
-	/* Reuse for each tuple */
-	buildstate->normvec = InitVector(buildstate->dimensions);
-
 	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
 												   "Hnsw build graph context",
 #if PG_VERSION_NUM >= 150000
@@ -736,7 +736,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 static void
 FreeBuildState(HnswBuildState * buildstate)
 {
-	pfree(buildstate->normvec);
 	MemoryContextDelete(buildstate->graphCtx);
 	MemoryContextDelete(buildstate->tmpCtx);
 }

src/hnswinsert.c

@@ -614,15 +614,19 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	Datum		value;
 	FmgrInfo   *normprocinfo;
 	Oid			collation = index->rd_indcollation[0];
+	HnswType	type = HnswGetType(index);
 
 	/* Detoast once for all calls */
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
+	/* Check value */
+	HnswCheckValue(value, type);
+
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
+		if (!HnswNormValue(normprocinfo, collation, &value, type))
 			return;
 	}
 

src/hnswscan.c

@@ -1,8 +1,6 @@
 #include "postgres.h"
 
 #include "access/relscan.h"
-#include "bitvector.h"
-#include "catalog/pg_type_d.h"
 #include "hnsw.h"
 #include "pgstat.h"
 #include "storage/bufmgr.h"
@@ -42,29 +40,6 @@ GetScanItems(IndexScanDesc scan, Datum q)
 	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
 }
 
-/*
- * Get dimensions from metapage
- */
-static int
-GetDimensions(Relation index)
-{
-	Buffer		buf;
-	Page		page;
-	HnswMetaPage metap;
-	int			dimensions;
-
-	buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
-	LockBuffer(buf, BUFFER_LOCK_SHARE);
-	page = BufferGetPage(buf);
-	metap = HnswPageGetMeta(page);
-
-	dimensions = metap->dimensions;
-
-	UnlockReleaseBuffer(buf);
-
-	return dimensions;
-}
-
 /*
  * Get scan value
  */
@@ -75,15 +50,7 @@ GetScanValue(IndexScanDesc scan)
 	Datum		value;
 
 	if (scan->orderByData->sk_flags & SK_ISNULL)
-	{
+		value = PointerGetDatum(NULL);
-		Oid			typid = TupleDescAttr(scan->indexRelation->rd_att, 0)->atttypid;
-		int			dimensions = GetDimensions(scan->indexRelation);
-
-		if (typid == BITOID || typid == VARBITOID)
-			value = PointerGetDatum(InitBitVector(dimensions));
-		else
-			value = PointerGetDatum(InitVector(dimensions));
-	}
 	else
 	{
 		value = scan->orderByData->sk_argument;
@@ -94,7 +61,7 @@ GetScanValue(IndexScanDesc scan)
 
 		/* Fine if normalization fails */
 		if (so->normprocinfo != NULL)
-			HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
+			HnswNormValue(so->normprocinfo, so->collation, &value, HnswGetType(scan->indexRelation));
 	}
 
 	return value;

src/hnswutils.c

@@ -3,12 +3,15 @@
 #include <math.h>
 
 #include "access/generic_xlog.h"
+#include "catalog/pg_type.h"
 #include "hnsw.h"
 #include "lib/pairingheap.h"
+#include "sparsevec.h"
 #include "storage/bufmgr.h"
 #include "utils/datum.h"
 #include "utils/memdebug.h"
 #include "utils/rel.h"
+#include "utils/syscache.h"
 #include "vector.h"
 
 #if PG_VERSION_NUM >= 130000
@@ -149,6 +152,32 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 	return index_getprocinfo(index, 1, procnum);
 }
 
+/*
+ * Get vector type
+ */
+HnswType
+HnswGetType(Relation index)
+{
+	Oid			typeOid = TupleDescAttr(index->rd_att, 0)->atttypid;
+	HeapTuple	tuple;
+	Form_pg_type type;
+	int			result;
+
+	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typeOid));
+	if (!HeapTupleIsValid(tuple))
+		elog(ERROR, "cache lookup failed for type %u", typeOid);
+
+	type = (Form_pg_type) GETSTRUCT(tuple);
+	if (strcmp(NameStr(type->typname), "sparsevec") == 0)
+		result = HNSW_TYPE_SPARSEVEC;
+	else
+		result = HNSW_TYPE_VECTOR;
+
+	ReleaseSysCache(tuple);
+
+	return result;
+}
+
 /*
  * Divide by the norm
  *
@@ -158,21 +187,40 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
  * if it's different than the original value
  */
 bool
-HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
+HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
 	{
-		Vector	   *v = DatumGetVector(*value);
+		/* TODO Remove vector-specific code */
+		if (type == HNSW_TYPE_VECTOR)
+		{
+			Vector	   *v = DatumGetVector(*value);
+			Vector	   *result = InitVector(v->dim);
 
-		if (result == NULL)
+			for (int i = 0; i < v->dim; i++)
-			result = InitVector(v->dim);
+				result->x[i] = v->x[i] / norm;
 
-		for (int i = 0; i < v->dim; i++)
+			*value = PointerGetDatum(result);
-			result->x[i] = v->x[i] / norm;
+		}
+		else if (type == HNSW_TYPE_SPARSEVEC)
+		{
+			SparseVector *v = DatumGetSparseVector(*value);
+			SparseVector *result = InitSparseVector(v->dim, v->nnz);
+			float	   *vx = SPARSEVEC_VALUES(v);
+			float	   *rx = SPARSEVEC_VALUES(result);
 
-		*value = PointerGetDatum(result);
+			for (int i = 0; i < v->nnz; i++)
+			{
+				result->indices[i] = v->indices[i];
+				rx[i] = vx[i] / norm;
+			}
+
+			*value = PointerGetDatum(result);
+		}
+		else
+			elog(ERROR, "Unsupported type");
 
 		return true;
 	}
@@ -180,6 +228,21 @@ HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 	return false;
 }
 
+/*
+ * Check if a value can be indexed
+ */
+void
+HnswCheckValue(Datum value, HnswType type)
+{
+	if (type == HNSW_TYPE_SPARSEVEC)
+	{
+		SparseVector *vec = DatumGetSparseVector(value);
+
+		if (vec->nnz > HNSW_MAX_NNZ)
+			elog(ERROR, "sparsevec cannot have more than %d non-zero elements for hnsw index", HNSW_MAX_NNZ);
+	}
+}
+
 /*
  * New buffer
  */
@@ -575,7 +638,12 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
+	{
+		if (DatumGetPointer(*q) == NULL)
+			*distance = 0;
+		else
+			*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
+	}
 
 	UnlockReleaseBuffer(buf);
 }

src/ivfbuild.c

@@ -57,7 +57,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->normvec))
+		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value))
 			return;
 	}
 
@@ -105,7 +105,7 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Add sample */
-	AddSample(values, state);
+	AddSample(values, buildstate);
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -153,7 +153,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
+		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value))
 			return;
 	}
 
@@ -356,9 +356,6 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions);
 	buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
 
-	/* Reuse for each tuple */
-	buildstate->normvec = InitVector(buildstate->dimensions);
-
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Ivfflat build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
@@ -380,7 +377,6 @@ FreeBuildState(IvfflatBuildState * buildstate)
 {
 	VectorArrayFree(buildstate->centers);
 	pfree(buildstate->listInfo);
-	pfree(buildstate->normvec);
 
 #ifdef IVFFLAT_KMEANS_DEBUG
 	pfree(buildstate->listSums);

src/ivfflat.h

@@ -172,7 +172,6 @@ typedef struct IvfflatBuildState
 	VectorArray samples;
 	VectorArray centers;
 	ListInfo   *listInfo;
-	Vector	   *normvec;
 
 #ifdef IVFFLAT_KMEANS_DEBUG
 	double		inertia;
@@ -267,7 +266,7 @@ void		VectorArrayFree(VectorArray arr);
 void		PrintVectorArray(char *msg, VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value);
 int			IvfflatGetLists(Relation index);
 void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

src/ivfinsert.c

@@ -85,7 +85,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
+		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value))
 			return;
 	}
 

src/ivfscan.c

@@ -293,7 +293,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 
 			/* Fine if normalization fails */
 			if (so->normprocinfo != NULL)
-				IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL);
+				IvfflatNormValue(so->normprocinfo, so->collation, &value);
 		}
 
 		IvfflatBench("GetScanLists", GetScanLists(scan, value));

src/ivfutils.c

@@ -75,16 +75,14 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
  * if it's different than the original value
  */
 bool
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
 	{
 		Vector	   *v = DatumGetVector(*value);
-
+		Vector	   *result = InitVector(v->dim);
-		if (result == NULL)
-			result = InitVector(v->dim);
 
 		for (int i = 0; i < v->dim; i++)
 			result->x[i] = v->x[i] / norm;

src/sparsevec.c

@@ -0,0 +1,779 @@
+#include "postgres.h"
+
+#include <limits.h>
+#include <math.h>
+
+#include "fmgr.h"
+#include "libpq/pqformat.h"
+#include "sparsevec.h"
+#include "utils/array.h"
+#include "utils/builtins.h"
+#include "vector.h"
+
+#if PG_VERSION_NUM >= 120000
+#include "common/shortest_dec.h"
+#include "utils/float.h"
+#else
+#include <float.h>
+#include "utils/builtins.h"
+#endif
+
+/*
+ * Ensure same dimensions
+ */
+static inline void
+CheckDims(SparseVector * a, SparseVector * b)
+{
+	if (a->dim != b->dim)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("different sparsevec dimensions %d and %d", a->dim, b->dim)));
+}
+
+/*
+ * Ensure expected dimensions
+ */
+static inline void
+CheckExpectedDim(int32 typmod, int dim)
+{
+	if (typmod != -1 && typmod != dim)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("expected %d dimensions, not %d", typmod, dim)));
+}
+
+/*
+ * Ensure valid dimensions
+ */
+static inline void
+CheckDim(int dim)
+{
+	if (dim < 1)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("sparsevec must have at least 1 dimension")));
+
+	if (dim > SPARSEVEC_MAX_DIM)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("sparsevec cannot have more than %d dimensions", SPARSEVEC_MAX_DIM)));
+}
+
+/*
+ * Ensure valid nnz
+ */
+static inline void
+CheckNnz(int nnz, int dim)
+{
+	if (nnz < 0)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("sparsevec must have at least one element")));
+
+	if (nnz > dim)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("sparsevec cannot have more elements than dimensions")));
+}
+
+/*
+ * Ensure valid index
+ */
+static inline void
+CheckIndex(int32 *indices, int i, int dim)
+{
+	int32		index = indices[i];
+
+	if (index < 0)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("index must not be negative")));
+
+	if (index >= dim)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("index must be less than dimensions")));
+
+	if (i > 0)
+	{
+		if (index < indices[i - 1])
+			ereport(ERROR,
+					(errcode(ERRCODE_DATA_EXCEPTION),
+					 errmsg("indexes must be in ascending order")));
+
+		if (index == indices[i - 1])
+			ereport(ERROR,
+					(errcode(ERRCODE_DATA_EXCEPTION),
+					 errmsg("indexes must not contain duplicates")));
+	}
+}
+
+/*
+ * Ensure finite element
+ */
+static inline void
+CheckElement(float value)
+{
+	if (isnan(value))
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("NaN not allowed in sparsevec")));
+
+	if (isinf(value))
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("infinite value not allowed in sparsevec")));
+}
+
+/*
+ * Allocate and initialize a new sparse vector
+ */
+SparseVector *
+InitSparseVector(int dim, int nnz)
+{
+	SparseVector *result;
+	int			size;
+
+	size = SPARSEVEC_SIZE(nnz);
+	result = (SparseVector *) palloc0(size);
+	SET_VARSIZE(result, size);
+	result->dim = dim;
+	result->nnz = nnz;
+
+	return result;
+}
+
+/*
+ * Check for whitespace, since array_isspace() is static
+ */
+static inline bool
+sparsevec_isspace(char ch)
+{
+	if (ch == ' ' ||
+		ch == '\t' ||
+		ch == '\n' ||
+		ch == '\r' ||
+		ch == '\v' ||
+		ch == '\f')
+		return true;
+	return false;
+}
+
+/*
+ * Convert textual representation to internal representation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_in);
+Datum
+sparsevec_in(PG_FUNCTION_ARGS)
+{
+	char	   *lit = PG_GETARG_CSTRING(0);
+	int32		typmod = PG_GETARG_INT32(2);
+	int			dim;
+	char	   *pt;
+	char	   *stringEnd;
+	SparseVector *result;
+	float	   *rvalues;
+	char	   *litcopy = pstrdup(lit);
+	char	   *str = litcopy;
+	int32	   *indices;
+	float	   *values;
+	int			maxNnz;
+	int			nnz = 0;
+
+	maxNnz = 1;
+	pt = str;
+	while (*pt != '\0')
+	{
+		if (*pt == ',')
+			maxNnz++;
+
+		pt++;
+	}
+
+	indices = palloc(maxNnz * sizeof(int32));
+	values = palloc(maxNnz * sizeof(float));
+
+	while (sparsevec_isspace(*str))
+		str++;
+
+	if (*str != '{')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
+				 errdetail("Vector contents must start with \"{\".")));
+
+	str++;
+	pt = strtok(str, ",");
+	stringEnd = pt;
+
+	while (pt != NULL && *stringEnd != '}')
+	{
+		long		index;
+		float		value;
+
+		/* TODO Better error */
+		if (nnz == maxNnz)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("ran out of buffer: \"%s\"", lit)));
+
+		while (sparsevec_isspace(*pt))
+			pt++;
+
+		/* Check for empty string like float4in */
+		if (*pt == '\0')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		/* Use similar logic as int2vectorin */
+		errno = 0;
+		index = strtol(pt, &stringEnd, 10);
+
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		if (errno == ERANGE || index < 0 || index > INT_MAX)
+			ereport(ERROR,
+					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+					 errmsg("index \"%ld\" is out of range for type sparsevec", index)));
+
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		while (sparsevec_isspace(*stringEnd))
+			stringEnd++;
+
+		if (*stringEnd != ':')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		stringEnd++;
+
+		while (sparsevec_isspace(*stringEnd))
+			stringEnd++;
+
+		errno = 0;
+		pt = stringEnd;
+		value = strtof(pt, &stringEnd);
+
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		/* Check for range error like float4in */
+		if (errno == ERANGE && (value == 0 || isinf(value)))
+			ereport(ERROR,
+					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+					 errmsg("\"%s\" is out of range for type sparsevec", pt)));
+
+		/* TODO Decide whether to store zero values */
+		if (value != 0)
+		{
+			indices[nnz] = index;
+			values[nnz] = value;
+			nnz++;
+		}
+
+		if (*stringEnd != '\0' && *stringEnd != '}')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		pt = strtok(NULL, ",");
+	}
+
+	if (stringEnd == NULL || *stringEnd != '}')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
+				 errdetail("Unexpected end of input.")));
+
+	stringEnd++;
+
+	if (*stringEnd != '/')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
+				 errdetail("Unexpected end of input.")));
+
+	stringEnd++;
+
+	/* Use similar logic as int2vectorin */
+	errno = 0;
+	pt = stringEnd;
+	dim = strtol(pt, &stringEnd, 10);
+
+	if (stringEnd == pt)
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+	/* Only whitespace is allowed after the closing brace */
+	while (sparsevec_isspace(*stringEnd))
+		stringEnd++;
+
+	if (*stringEnd != '\0')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
+				 errdetail("Junk after closing.")));
+
+	pfree(litcopy);
+
+	CheckDim(dim);
+	CheckExpectedDim(typmod, dim);
+
+	result = InitSparseVector(dim, nnz);
+	rvalues = SPARSEVEC_VALUES(result);
+	for (int i = 0; i < nnz; i++)
+	{
+		result->indices[i] = indices[i];
+		rvalues[i] = values[i];
+
+		CheckIndex(result->indices, i, dim);
+		CheckElement(rvalues[i]);
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+#define AppendChar(ptr, c) (*(ptr)++ = (c))
+#define AppendFloat(ptr, f) ((ptr) += float_to_shortest_decimal_bufn((f), (ptr)))
+
+#if PG_VERSION_NUM >= 140000
+#define AppendInt(ptr, i) ((ptr) += pg_ltoa((i), (ptr)))
+#else
+#define AppendInt(ptr, i) \
+	do { \
+		pg_ltoa(i, ptr); \
+		while (*ptr != '\0') \
+			ptr++; \
+	} while (0)
+#endif
+
+/*
+ * Convert internal representation to textual representation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_out);
+Datum
+sparsevec_out(PG_FUNCTION_ARGS)
+{
+	SparseVector *sparsevec = PG_GETARG_SPARSEVEC_P(0);
+	float	   *values = SPARSEVEC_VALUES(sparsevec);
+	char	   *buf;
+	char	   *ptr;
+
+	/*
+	 * Need:
+	 *
+	 * nnz * 10 bytes for index (positive integer)
+	 *
+	 * nnz bytes for :
+	 *
+	 * nnz * (FLOAT_SHORTEST_DECIMAL_LEN - 1) bytes for
+	 * float_to_shortest_decimal_bufn
+	 *
+	 * nnz - 1 bytes for ,
+	 *
+	 * 10 bytes for dimensions
+	 *
+	 * 4 bytes for {, }, /, and \0
+	 */
+	buf = (char *) palloc((11 + FLOAT_SHORTEST_DECIMAL_LEN) * sparsevec->nnz + 13);
+	ptr = buf;
+
+	AppendChar(ptr, '{');
+
+	for (int i = 0; i < sparsevec->nnz; i++)
+	{
+		if (i > 0)
+			AppendChar(ptr, ',');
+
+		AppendInt(ptr, sparsevec->indices[i]);
+		AppendChar(ptr, ':');
+		AppendFloat(ptr, values[i]);
+	}
+
+	AppendChar(ptr, '}');
+	AppendChar(ptr, '/');
+	AppendInt(ptr, sparsevec->dim);
+
+	*ptr = '\0';
+
+	PG_FREE_IF_COPY(sparsevec, 0);
+	PG_RETURN_CSTRING(buf);
+}
+
+/*
+ * Convert type modifier
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_typmod_in);
+Datum
+sparsevec_typmod_in(PG_FUNCTION_ARGS)
+{
+	ArrayType  *ta = PG_GETARG_ARRAYTYPE_P(0);
+	int32	   *tl;
+	int			n;
+
+	tl = ArrayGetIntegerTypmods(ta, &n);
+
+	if (n != 1)
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
+				 errmsg("invalid type modifier")));
+
+	if (*tl < 1)
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
+				 errmsg("dimensions for type sparsevec must be at least 1")));
+
+	if (*tl > SPARSEVEC_MAX_DIM)
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
+				 errmsg("dimensions for type sparsevec cannot exceed %d", SPARSEVEC_MAX_DIM)));
+
+	PG_RETURN_INT32(*tl);
+}
+
+/*
+ * Convert external binary representation to internal representation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_recv);
+Datum
+sparsevec_recv(PG_FUNCTION_ARGS)
+{
+	StringInfo	buf = (StringInfo) PG_GETARG_POINTER(0);
+	int32		typmod = PG_GETARG_INT32(2);
+	SparseVector *result;
+	int32		dim;
+	int32		nnz;
+	int32		unused;
+	float	   *values;
+
+	dim = pq_getmsgint(buf, sizeof(int32));
+	nnz = pq_getmsgint(buf, sizeof(int32));
+	unused = pq_getmsgint(buf, sizeof(int32));
+
+	CheckDim(dim);
+	CheckNnz(nnz, dim);
+	CheckExpectedDim(typmod, dim);
+
+	if (unused != 0)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("expected unused to be 0, not %d", unused)));
+
+	result = InitSparseVector(dim, nnz);
+	values = SPARSEVEC_VALUES(result);
+
+	for (int i = 0; i < nnz; i++)
+	{
+		result->indices[i] = pq_getmsgint(buf, sizeof(int32));
+		CheckIndex(result->indices, i, dim);
+	}
+
+	for (int i = 0; i < nnz; i++)
+	{
+		values[i] = pq_getmsgfloat4(buf);
+		CheckElement(values[i]);
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Convert internal representation to the external binary representation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_send);
+Datum
+sparsevec_send(PG_FUNCTION_ARGS)
+{
+	SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
+	float	   *values = SPARSEVEC_VALUES(svec);
+	StringInfoData buf;
+
+	pq_begintypsend(&buf);
+	pq_sendint(&buf, svec->dim, sizeof(int32));
+	pq_sendint(&buf, svec->nnz, sizeof(int32));
+	pq_sendint(&buf, svec->unused, sizeof(int32));
+	for (int i = 0; i < svec->nnz; i++)
+		pq_sendint(&buf, svec->indices[i], sizeof(int32));
+	for (int i = 0; i < svec->nnz; i++)
+		pq_sendfloat4(&buf, values[i]);
+
+	PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
+}
+
+/*
+ * Convert sparse vector to sparse vector
+ * This is needed to check the type modifier
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec);
+Datum
+sparsevec(PG_FUNCTION_ARGS)
+{
+	SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
+	int32		typmod = PG_GETARG_INT32(1);
+
+	CheckExpectedDim(typmod, svec->dim);
+
+	PG_RETURN_POINTER(svec);
+}
+
+/*
+ * Convert dense vector to sparse vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_to_sparsevec);
+Datum
+vector_to_sparsevec(PG_FUNCTION_ARGS)
+{
+	Vector	   *vec = PG_GETARG_VECTOR_P(0);
+	int32		typmod = PG_GETARG_INT32(1);
+	SparseVector *result;
+	int			dim = vec->dim;
+	int			nnz = 0;
+	float	   *values;
+	int			j = 0;
+
+	CheckDim(dim);
+	CheckExpectedDim(typmod, dim);
+
+	for (int i = 0; i < dim; i++)
+	{
+		if (vec->x[i] != 0)
+			nnz++;
+	}
+
+	result = InitSparseVector(dim, nnz);
+	values = SPARSEVEC_VALUES(result);
+	for (int i = 0; i < dim; i++)
+	{
+		if (vec->x[i] != 0)
+		{
+			/* Safety check */
+			if (j == nnz)
+				elog(ERROR, "safety check failed");
+
+			result->indices[j] = i;
+			values[j] = vec->x[i];
+			j++;
+		}
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ *  Get the L2 squared distance between sparse vectors
+ */
+static double
+l2_distance_squared_internal(SparseVector * a, SparseVector * b)
+{
+	float	   *ax = SPARSEVEC_VALUES(a);
+	float	   *bx = SPARSEVEC_VALUES(b);
+	double		distance = 0.0;
+	int			bpos = 0;
+
+	for (int i = 0; i < a->nnz; i++)
+	{
+		int			ai = a->indices[i];
+		int			bi = -1;
+
+		for (int j = bpos; j < b->nnz; j++)
+		{
+			bi = b->indices[j];
+
+			if (ai == bi)
+			{
+				double		diff = ax[i] - bx[j];
+
+				distance += diff * diff;
+			}
+			else if (ai > bi)
+				distance += bx[j] * bx[j];
+
+			/* Update start for next iteration */
+			if (ai >= bi)
+				bpos = j + 1;
+
+			/* Found or passed it */
+			if (bi >= ai)
+				break;
+		}
+
+		if (ai != bi)
+			distance += ax[i] * ax[i];
+	}
+
+	for (int j = bpos; j < b->nnz; j++)
+		distance += bx[j] * bx[j];
+
+	return distance;
+}
+
+/*
+ * Get the L2 distance between sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_distance);
+Datum
+sparsevec_l2_distance(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	CheckDims(a, b);
+
+	PG_RETURN_FLOAT8(sqrt(l2_distance_squared_internal(a, b)));
+}
+
+/*
+ * Get the L2 squared distance between sparse vectors
+ * This saves a sqrt calculation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_squared_distance);
+Datum
+sparsevec_l2_squared_distance(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	CheckDims(a, b);
+
+	PG_RETURN_FLOAT8(l2_distance_squared_internal(a, b));
+}
+
+/*
+ * Get the inner product of two sparse vectors
+ */
+static double
+inner_product_internal(SparseVector * a, SparseVector * b)
+{
+	float	   *ax = SPARSEVEC_VALUES(a);
+	float	   *bx = SPARSEVEC_VALUES(b);
+	double		distance = 0.0;
+	int			bpos = 0;
+
+	for (int i = 0; i < a->nnz; i++)
+	{
+		int			ai = a->indices[i];
+
+		for (int j = bpos; j < b->nnz; j++)
+		{
+			int			bi = b->indices[j];
+
+			/* Only update when the same index */
+			if (ai == bi)
+				distance += ax[i] * bx[j];
+
+			/* Update start for next iteration */
+			if (ai >= bi)
+				bpos = j + 1;
+
+			/* Found or passed it */
+			if (bi >= ai)
+				break;
+		}
+	}
+
+	return distance;
+}
+
+/*
+ * Get the inner product of two sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_inner_product);
+Datum
+sparsevec_inner_product(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	CheckDims(a, b);
+
+	PG_RETURN_FLOAT8(inner_product_internal(a, b));
+}
+
+/*
+ * Get the negative inner product of two sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_negative_inner_product);
+Datum
+sparsevec_negative_inner_product(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	CheckDims(a, b);
+
+	PG_RETURN_FLOAT8(-inner_product_internal(a, b));
+}
+
+/*
+ * Get the cosine distance between two sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_cosine_distance);
+Datum
+sparsevec_cosine_distance(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+	float	   *ax = SPARSEVEC_VALUES(a);
+	float	   *bx = SPARSEVEC_VALUES(b);
+	float		norma = 0.0;
+	float		normb = 0.0;
+	double		similarity;
+
+	CheckDims(a, b);
+
+	similarity = inner_product_internal(a, b);
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->nnz; i++)
+		norma += ax[i] * ax[i];
+
+	/* Auto-vectorized */
+	for (int i = 0; i < b->nnz; i++)
+		normb += bx[i] * bx[i];
+
+	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
+	similarity /= sqrt((double) norma * (double) normb);
+
+#ifdef _MSC_VER
+	/* /fp:fast may not propagate NaN */
+	if (isnan(similarity))
+		PG_RETURN_FLOAT8(NAN);
+#endif
+
+	/* Keep in range */
+	if (similarity > 1)
+		similarity = 1.0;
+	else if (similarity < -1)
+		similarity = -1.0;
+
+	PG_RETURN_FLOAT8(1.0 - similarity);
+}
+
+/*
+ * Get the L2 norm of a sparse vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_norm);
+Datum
+sparsevec_norm(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	float	   *ax = SPARSEVEC_VALUES(a);
+	double		norm = 0.0;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->nnz; i++)
+		norm += (double) ax[i] * (double) ax[i];
+
+	PG_RETURN_FLOAT8(sqrt(norm));
+}

src/sparsevec.h

@@ -0,0 +1,23 @@
+#ifndef SPARSEVEC_H
+#define SPARSEVEC_H
+
+#define SPARSEVEC_MAX_DIM 100000
+
+#define SPARSEVEC_SIZE(_nnz)		(offsetof(SparseVector, indices) + MAXALIGN((_nnz) * sizeof(int32)) + (_nnz * sizeof(float)))
+#define SPARSEVEC_VALUES(x)		((float *) (((char *) (x)) + offsetof(SparseVector, indices) + MAXALIGN((x)->nnz * sizeof(int32))))
+#define DatumGetSparseVector(x)		((SparseVector *) PG_DETOAST_DATUM(x))
+#define PG_GETARG_SPARSEVEC_P(x)	DatumGetSparseVector(PG_GETARG_DATUM(x))
+#define PG_RETURN_SPARSEVEC_P(x)	PG_RETURN_POINTER(x)
+
+typedef struct SparseVector
+{
+	int32		vl_len_;		/* varlena header (do not touch directly!) */
+	int32		dim;			/* number of dimensions */
+	int32		nnz;
+	int32		unused;
+	int32		indices[FLEXIBLE_ARRAY_MEMBER];
+}			SparseVector;
+
+SparseVector *InitSparseVector(int dim, int nnz);
+
+#endif

src/vector.c

@@ -2,7 +2,6 @@
 
 #include <math.h>
 
-#include "bitvector.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -11,6 +10,7 @@
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
 #include "port.h"				/* for strtof() */
+#include "sparsevec.h"
 #include "utils/array.h"
 #include "utils/builtins.h"
 #include "utils/float.h"
@@ -861,26 +861,6 @@ vector_mul(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
-/*
- * Quantize a vector
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(quantize_binary);
-Datum
-quantize_binary(PG_FUNCTION_ARGS)
-{
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	float	   *ax = a->x;
-	VarBit	   *result = InitBitVector(a->dim);
-	unsigned char *rx = VARBITS(result);
-
-	/* TODO Improve */
-	for (int i = 0; i < a->dim; i++)
-		rx[i / 8] |= (ax[i] > 0) << (7 - (i % 8));
-
-	PG_RETURN_VARBIT_P(result);
-}
-
-
 /*
  * Internal helper to compare vectors
  */
@@ -1181,3 +1161,26 @@ vector_avg(PG_FUNCTION_ARGS)
 
 	PG_RETURN_POINTER(result);
 }
+
+/*
+ * Convert sparse vector to dense vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_to_vector);
+Datum
+sparsevec_to_vector(PG_FUNCTION_ARGS)
+{
+	SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
+	int32		typmod = PG_GETARG_INT32(1);
+	Vector	   *result;
+	int			dim = svec->dim;
+	float	   *values = SPARSEVEC_VALUES(svec);
+
+	CheckDim(dim);
+	CheckExpectedDim(typmod, dim);
+
+	result = InitVector(dim);
+	for (int i = 0; i < svec->nnz; i++)
+		result->x[svec->indices[i]] = values[i];
+
+	PG_RETURN_POINTER(result);
+}

test/expected/bit_functions.out

@@ -1,64 +0,0 @@
-SELECT hamming_distance(B'111', B'111');
- hamming_distance 
-------------------
-                0
-(1 row)
-
-SELECT hamming_distance(B'111', B'110');
- hamming_distance 
-------------------
-                1
-(1 row)
-
-SELECT hamming_distance(B'111', B'100');
- hamming_distance 
-------------------
-                2
-(1 row)
-
-SELECT hamming_distance(B'111', B'000');
- hamming_distance 
-------------------
-                3
-(1 row)
-
-SELECT hamming_distance(B'111', B'00');
-ERROR:  different bit lengths 3 and 2
-SELECT jaccard_distance(B'1111', B'1111');
- jaccard_distance 
-------------------
-                0
-(1 row)
-
-SELECT jaccard_distance(B'1111', B'1110');
- jaccard_distance 
-------------------
-             0.25
-(1 row)
-
-SELECT jaccard_distance(B'1111', B'1100');
- jaccard_distance 
-------------------
-              0.5
-(1 row)
-
-SELECT jaccard_distance(B'1111', B'1000');
- jaccard_distance 
-------------------
-             0.75
-(1 row)
-
-SELECT jaccard_distance(B'1111', B'0000');
- jaccard_distance 
-------------------
-                1
-(1 row)
-
-SELECT jaccard_distance(B'1100', B'1000');
- jaccard_distance 
-------------------
-              0.5
-(1 row)
-
-SELECT jaccard_distance(B'1111', B'000');
-ERROR:  different bit lengths 4 and 3

test/expected/functions.out

@@ -104,122 +104,110 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]'::vector, '[3,4]');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('[0,0]', '[0,1]');
+SELECT l2_distance('[0,0]'::vector, '[0,1]');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT l2_distance('[1,2]', '[3]');
+SELECT l2_distance('[1,2]'::vector, '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l2_distance('[3e38]', '[-3e38]');
+SELECT l2_distance('[3e38]'::vector, '[-3e38]');
  l2_distance 
 -------------
     Infinity
 (1 row)
 
-SELECT inner_product('[1,2]', '[3,4]');
+SELECT inner_product('[1,2]'::vector, '[3,4]');
  inner_product 
 ---------------
             11
 (1 row)
 
-SELECT inner_product('[1,2]', '[3]');
+SELECT inner_product('[1,2]'::vector, '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT inner_product('[3e38]', '[3e38]');
+SELECT inner_product('[3e38]'::vector, '[3e38]');
  inner_product 
 ---------------
       Infinity
 (1 row)
 
-SELECT cosine_distance('[1,2]', '[2,4]');
+SELECT cosine_distance('[1,2]'::vector, '[2,4]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,2]', '[0,0]');
+SELECT cosine_distance('[1,2]'::vector, '[0,0]');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,1]'::vector, '[1,1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,0]', '[0,2]');
+SELECT cosine_distance('[1,0]'::vector, '[0,2]');
  cosine_distance 
 -----------------
                1
 (1 row)
 
-SELECT cosine_distance('[1,1]', '[-1,-1]');
+SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[1,2]', '[3]');
+SELECT cosine_distance('[1,2]'::vector, '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT cosine_distance('[1,1]', '[1.1,1.1]');
+SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[3e38]', '[3e38]');
+SELECT cosine_distance('[3e38]'::vector, '[3e38]');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT l1_distance('[0,0]', '[3,4]');
+SELECT l1_distance('[0,0]'::vector, '[3,4]');
  l1_distance 
 -------------
            7
 (1 row)
 
-SELECT l1_distance('[0,0]', '[0,1]');
+SELECT l1_distance('[0,0]'::vector, '[0,1]');
  l1_distance 
 -------------
            1
 (1 row)
 
-SELECT l1_distance('[1,2]', '[3]');
+SELECT l1_distance('[1,2]'::vector, '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l1_distance('[3e38]', '[-3e38]');
+SELECT l1_distance('[3e38]'::vector, '[-3e38]');
  l1_distance 
 -------------
     Infinity
 (1 row)
 
-SELECT quantize_binary('[1,0,-1]');
- quantize_binary 
------------------
- 100
-(1 row)
-
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]');
- quantize_binary 
------------------
- 01001110101
-(1 row)
-
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
     avg    
 -----------

test/expected/hnsw_hamming.out

@@ -1,21 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val bit(3));
-INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
-CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
-INSERT INTO t (val) VALUES (B'110');
-SELECT * FROM t ORDER BY val <~> B'111';
- val 
------
- 111
- 110
- 100
- 000
-(4 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
- count 
--------
-     4
-(1 row)
-
-DROP TABLE t;

test/expected/hnsw_jaccard.out

@@ -1,21 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val bit(4));
-INSERT INTO t (val) VALUES (B'0000'), (B'1100'), (B'1111'), (NULL);
-CREATE INDEX ON t USING hnsw (val bit_jaccard_ops);
-INSERT INTO t (val) VALUES (B'1110');
-SELECT * FROM t ORDER BY val <%> B'1111';
- val  
-------
- 1111
- 1110
- 1100
- 0000
-(4 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <%> (SELECT NULL::bit)) t2;
- count 
--------
-     4
-(1 row)
-
-DROP TABLE t;

test/expected/hnsw_l2.out

@@ -12,14 +12,11 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
-   val   
+ count 
----------
+-------
- [0,0,0]
+     4
- [1,1,1]
+(1 row)
- [1,2,3]
- [1,2,4]
-(4 rows)
 
 SELECT COUNT(*) FROM t;
  count 

test/expected/hnsw_sparsevec_cosine.out

@@ -0,0 +1,26 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+SELECT * FROM t ORDER BY val <=> '{0:3,1:3,2:3}/3';
+       val       
+-----------------
+ {0:1,1:1,2:1}/3
+ {0:1,1:2,2:3}/3
+ {0:1,1:2,2:4}/3
+(3 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;
+ count 
+-------
+     3
+(1 row)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::sparsevec)) t2;
+ count 
+-------
+     3
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_sparsevec_ip.out

@@ -0,0 +1,21 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+SELECT * FROM t ORDER BY val <#> '{0:3,1:3,2:3}/3';
+       val       
+-----------------
+ {0:1,1:2,2:4}/3
+ {0:1,1:2,2:3}/3
+ {0:1,1:1,2:1}/3
+ {}/3
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::sparsevec)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_sparsevec_l2.out

@@ -0,0 +1,43 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+       val       
+-----------------
+ {0:1,1:2,2:3}/3
+ {0:1,1:2,2:4}/3
+ {0:1,1:1,2:1}/3
+ {}/3
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::sparsevec)) t2;
+ count 
+-------
+     4
+(1 row)
+
+SELECT COUNT(*) FROM t;
+ count 
+-------
+     5
+(1 row)
+
+TRUNCATE t;
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+ val 
+-----
+(0 rows)
+
+DROP TABLE t;
+-- TODO move
+CREATE TABLE t (val sparsevec(1001));
+INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+ERROR:  sparsevec cannot have more than 1000 non-zero elements for hnsw index
+TRUNCATE t;
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
+ERROR:  sparsevec cannot have more than 1000 non-zero elements for hnsw index
+DROP TABLE t;

test/expected/input.out

@@ -66,6 +66,22 @@ SELECT '[4e38,1]'::vector;
 ERROR:  infinite value not allowed in vector
 LINE 1: SELECT '[4e38,1]'::vector;
                ^
+SELECT '[-4e38,1]'::vector;
+ERROR:  infinite value not allowed in vector
+LINE 1: SELECT '[-4e38,1]'::vector;
+               ^
+SELECT '[1e-46,1]'::vector;
+ vector 
+--------
+ [0,1]
+(1 row)
+
+SELECT '[-1e-46,1]'::vector;
+ vector 
+--------
+ [-0,1]
+(1 row)
+
 SELECT '[1,2,3'::vector;
 ERROR:  malformed vector literal: "[1,2,3"
 LINE 1: SELECT '[1,2,3'::vector;

test/expected/sparsevec_functions.out

@@ -0,0 +1,62 @@
+SELECT l2_distance('{}/2'::sparsevec, '{0:3,1:4}/2');
+ l2_distance 
+-------------
+           5
+(1 row)
+
+SELECT l2_distance('{}/2'::sparsevec, '{1:1}/2');
+ l2_distance 
+-------------
+           1
+(1 row)
+
+SELECT '{}/2'::sparsevec <-> '{0:3,1:4}/2';
+ ?column? 
+----------
+        5
+(1 row)
+
+SELECT inner_product('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+ inner_product 
+---------------
+            10
+(1 row)
+
+SELECT sparsevec_negative_inner_product('{0:1,1:2}/2', '{0:2,1:4}/2');
+ sparsevec_negative_inner_product 
+----------------------------------
+                              -10
+(1 row)
+
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{}/2');
+ cosine_distance 
+-----------------
+             NaN
+(1 row)
+
+SELECT cosine_distance('{0:1,1:1}/2'::sparsevec, '{0:-1,1:-1}/2');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{1:2}/2');
+ cosine_distance 
+-----------------
+               1
+(1 row)
+
+SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
+ cosine_distance 
+-----------------
+             NaN
+(1 row)
+
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{0:1}/3');
+ERROR:  different sparsevec dimensions 2 and 3

test/expected/sparsevec_input.out

@@ -0,0 +1,62 @@
+SELECT '{0:1.5,2:3.5}/5'::sparsevec;
+    sparsevec    
+-----------------
+ {0:1.5,2:3.5}/5
+(1 row)
+
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector;
+     vector      
+-----------------
+ [1.5,0,3.5,0,0]
+(1 row)
+
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(5);
+     vector      
+-----------------
+ [1.5,0,3.5,0,0]
+(1 row)
+
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(4);
+ERROR:  expected 4 dimensions, not 5
+SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
+    sparsevec    
+-----------------
+ {1:1.5,3:3.5}/5
+(1 row)
+
+SELECT '{0:0,1:1,2:0}/3'::sparsevec;
+ sparsevec 
+-----------
+ {1:1}/3
+(1 row)
+
+SELECT '{1:1,0:1}/2'::sparsevec;
+ERROR:  indexes must be in ascending order
+LINE 1: SELECT '{1:1,0:1}/2'::sparsevec;
+               ^
+SELECT '{}/5'::sparsevec;
+ sparsevec 
+-----------
+ {}/5
+(1 row)
+
+SELECT '{}/-1'::sparsevec;
+ERROR:  sparsevec must have at least 1 dimension
+LINE 1: SELECT '{}/-1'::sparsevec;
+               ^
+SELECT '{}/100001'::sparsevec;
+ERROR:  sparsevec cannot have more than 100000 dimensions
+LINE 1: SELECT '{}/100001'::sparsevec;
+               ^
+SELECT '{}/16001'::sparsevec::vector;
+ERROR:  vector cannot have more than 16000 dimensions
+SELECT '{-1:1}/1'::sparsevec;
+ERROR:  index "-1" is out of range for type sparsevec
+LINE 1: SELECT '{-1:1}/1'::sparsevec;
+               ^
+SELECT '{1:1}/1'::sparsevec;
+ERROR:  index must be less than dimensions
+LINE 1: SELECT '{1:1}/1'::sparsevec;
+               ^
+SELECT '{}/1'::sparsevec(2);
+ERROR:  expected 2 dimensions, not 1

test/sql/bit_functions.sql

@@ -1,13 +0,0 @@
-SELECT hamming_distance(B'111', B'111');
-SELECT hamming_distance(B'111', B'110');
-SELECT hamming_distance(B'111', B'100');
-SELECT hamming_distance(B'111', B'000');
-SELECT hamming_distance(B'111', B'00');
-
-SELECT jaccard_distance(B'1111', B'1111');
-SELECT jaccard_distance(B'1111', B'1110');
-SELECT jaccard_distance(B'1111', B'1100');
-SELECT jaccard_distance(B'1111', B'1000');
-SELECT jaccard_distance(B'1111', B'0000');
-SELECT jaccard_distance(B'1100', B'1000');
-SELECT jaccard_distance(B'1111', B'000');

test/sql/functions.sql

@@ -24,32 +24,29 @@ SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
 
-SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]'::vector, '[3,4]');
-SELECT l2_distance('[0,0]', '[0,1]');
+SELECT l2_distance('[0,0]'::vector, '[0,1]');
-SELECT l2_distance('[1,2]', '[3]');
+SELECT l2_distance('[1,2]'::vector, '[3]');
-SELECT l2_distance('[3e38]', '[-3e38]');
+SELECT l2_distance('[3e38]'::vector, '[-3e38]');
 
-SELECT inner_product('[1,2]', '[3,4]');
+SELECT inner_product('[1,2]'::vector, '[3,4]');
-SELECT inner_product('[1,2]', '[3]');
+SELECT inner_product('[1,2]'::vector, '[3]');
-SELECT inner_product('[3e38]', '[3e38]');
+SELECT inner_product('[3e38]'::vector, '[3e38]');
 
-SELECT cosine_distance('[1,2]', '[2,4]');
+SELECT cosine_distance('[1,2]'::vector, '[2,4]');
-SELECT cosine_distance('[1,2]', '[0,0]');
+SELECT cosine_distance('[1,2]'::vector, '[0,0]');
-SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,1]'::vector, '[1,1]');
-SELECT cosine_distance('[1,0]', '[0,2]');
+SELECT cosine_distance('[1,0]'::vector, '[0,2]');
-SELECT cosine_distance('[1,1]', '[-1,-1]');
+SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
-SELECT cosine_distance('[1,2]', '[3]');
+SELECT cosine_distance('[1,2]'::vector, '[3]');
-SELECT cosine_distance('[1,1]', '[1.1,1.1]');
+SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
-SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
-SELECT cosine_distance('[3e38]', '[3e38]');
+SELECT cosine_distance('[3e38]'::vector, '[3e38]');
 
-SELECT l1_distance('[0,0]', '[3,4]');
+SELECT l1_distance('[0,0]'::vector, '[3,4]');
-SELECT l1_distance('[0,0]', '[0,1]');
+SELECT l1_distance('[0,0]'::vector, '[0,1]');
-SELECT l1_distance('[1,2]', '[3]');
+SELECT l1_distance('[1,2]'::vector, '[3]');
-SELECT l1_distance('[3e38]', '[-3e38]');
+SELECT l1_distance('[3e38]'::vector, '[-3e38]');
-
-SELECT quantize_binary('[1,0,-1]');
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]');
 
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;

test/sql/hnsw_hamming.sql

@@ -1,12 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val bit(3));
-INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
-CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
-
-INSERT INTO t (val) VALUES (B'110');
-
-SELECT * FROM t ORDER BY val <~> B'111';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
-
-DROP TABLE t;

test/sql/hnsw_jaccard.sql

@@ -1,12 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val bit(4));
-INSERT INTO t (val) VALUES (B'0000'), (B'1100'), (B'1111'), (NULL);
-CREATE INDEX ON t USING hnsw (val bit_jaccard_ops);
-
-INSERT INTO t (val) VALUES (B'1110');
-
-SELECT * FROM t ORDER BY val <%> B'1111';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <%> (SELECT NULL::bit)) t2;
-
-DROP TABLE t;

test/sql/hnsw_l2.sql

@@ -7,7 +7,7 @@ CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
 SELECT COUNT(*) FROM t;
 
 TRUNCATE t;

test/sql/hnsw_sparsevec_cosine.sql

@@ -0,0 +1,13 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
+
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+
+SELECT * FROM t ORDER BY val <=> '{0:3,1:3,2:3}/3';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::sparsevec)) t2;
+
+DROP TABLE t;

test/sql/hnsw_sparsevec_ip.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
+
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+
+SELECT * FROM t ORDER BY val <#> '{0:3,1:3,2:3}/3';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::sparsevec)) t2;
+
+DROP TABLE t;

test/sql/hnsw_sparsevec_l2.sql

@@ -0,0 +1,25 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::sparsevec)) t2;
+SELECT COUNT(*) FROM t;
+
+TRUNCATE t;
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+
+DROP TABLE t;
+
+-- TODO move
+CREATE TABLE t (val sparsevec(1001));
+INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+TRUNCATE t;
+CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
+INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
+DROP TABLE t;

test/sql/input.sql

@@ -11,6 +11,9 @@ SELECT '[1.5e38,-1.5e38]'::vector;
 SELECT '[1.5e+38,-1.5e+38]'::vector;
 SELECT '[1.5e-38,-1.5e-38]'::vector;
 SELECT '[4e38,1]'::vector;
+SELECT '[-4e38,1]'::vector;
+SELECT '[1e-46,1]'::vector;
+SELECT '[-1e-46,1]'::vector;
 SELECT '[1,2,3'::vector;
 SELECT '[1,2,3]9'::vector;
 SELECT '1,2,3'::vector;

test/sql/sparsevec_functions.sql

@@ -0,0 +1,13 @@
+SELECT l2_distance('{}/2'::sparsevec, '{0:3,1:4}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{1:1}/2');
+SELECT '{}/2'::sparsevec <-> '{0:3,1:4}/2';
+
+SELECT inner_product('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+SELECT sparsevec_negative_inner_product('{0:1,1:2}/2', '{0:2,1:4}/2');
+
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{}/2');
+SELECT cosine_distance('{0:1,1:1}/2'::sparsevec, '{0:-1,1:-1}/2');
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{1:2}/2');
+SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{0:1}/3');

test/sql/sparsevec_input.sql

@@ -0,0 +1,19 @@
+SELECT '{0:1.5,2:3.5}/5'::sparsevec;
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector;
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(5);
+SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(4);
+SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
+
+SELECT '{0:0,1:1,2:0}/3'::sparsevec;
+
+SELECT '{1:1,0:1}/2'::sparsevec;
+
+SELECT '{}/5'::sparsevec;
+SELECT '{}/-1'::sparsevec;
+SELECT '{}/100001'::sparsevec;
+SELECT '{}/16001'::sparsevec::vector;
+
+SELECT '{-1:1}/1'::sparsevec;
+SELECT '{1:1}/1'::sparsevec;
+
+SELECT '{}/1'::sparsevec(2);

test/t/016_hnsw_vacuum_recall.pl

@@ -86,7 +86,7 @@ foreach (@queries)
 	push(@expected, $res);
 }
 
-test_recall(0.20, $limit, "before vacuum");
+test_recall(0.19, $limit, "before vacuum");
 test_recall(0.95, 100, "before vacuum");
 
 # TODO Test concurrent inserts with vacuum

test/t/020_hnsw_bit_build_recall.pl

@@ -1,137 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-my $dim = 52;
-my $max = 2**$dim;
-
-sub test_recall
-{
-	my ($min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 100;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator $queries[0] LIMIT $limit;
-	));
-	like($explain, qr/Index Scan/);
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SET hnsw.ef_search = 100;
-			SELECT i FROM tst ORDER BY v $operator $queries[$i] LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-
-		my @expected_ids = split("\n", $expected[$i]);
-		my %expected_set = map { $_ => 1 } @expected_ids;
-
-		foreach (@actual_ids)
-		{
-			if (exists($expected_set{$_}))
-			{
-				$correct++;
-			}
-		}
-
-		$total += $limit;
-	}
-
-	cmp_ok($correct / $total, ">=", $min, $operator);
-}
-
-# Initialize node
-$node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v bit($dim));");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, (random() * $max)::bigint::bit($dim) FROM generate_series(1, 10000) i;"
-);
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r = int(rand() * $max);
-	push(@queries, "${r}::bigint::bit($dim)");
-}
-
-# Check each index type
-my @operators = ("<~>", "<\%>");
-my @opclasses = ("bit_hamming_ops", "bit_jaccard_ops");
-
-for my $i (0 .. $#operators)
-{
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
-
-	# Get exact results
-	@expected = ();
-	foreach (@queries)
-	{
-		# Handle ties
-		my $res = $node->safe_psql("postgres", qq(
-			WITH top AS (
-				SELECT v $operator $_ AS distance FROM tst ORDER BY v $operator $_ LIMIT $limit
-			)
-			SELECT i FROM tst WHERE (v $operator $_) <= (SELECT MAX(distance) FROM top)
-		));
-		push(@expected, $res);
-	}
-
-	# Build index serially
-	$node->safe_psql("postgres", qq(
-		SET max_parallel_maintenance_workers = 0;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
-
-	# Test approximate results
-	my $min = $operator eq "<\%>" ? 0.96 : 0.99;
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-
-	# Build index in parallel in memory
-	my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
-		SET client_min_messages = DEBUG;
-		SET min_parallel_table_scan_size = 1;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
-	is($ret, 0, $stderr);
-	like($stderr, qr/using \d+ parallel workers/);
-
-	# Test approximate results
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-
-	# Build index in parallel on disk
-	# Set parallel_workers on table to use workers with low maintenance_work_mem
-	($ret, $stdout, $stderr) = $node->psql("postgres", qq(
-		ALTER TABLE tst SET (parallel_workers = 2);
-		SET client_min_messages = DEBUG;
-		SET maintenance_work_mem = '4MB';
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-		ALTER TABLE tst RESET (parallel_workers);
-	));
-	is($ret, 0, $stderr);
-	like($stderr, qr/using \d+ parallel workers/);
-	like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem/);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-}
-
-done_testing();