Added strict ordering [skip ci]

Added todo [skip ci]
Updated changelog [skip ci]
2026-07-23 04:20:56 +08:00 · 2024-09-28 11:31:03 -07:00 · 2024-09-28 10:07:09 -07:00 · 2024-09-28 09:57:05 -07:00 · 2024-09-26 08:34:07 -07:00 · 2024-09-25 17:29:20 -07:00
25 changed files with 1024 additions and 853 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,9 +1,8 @@
 ## 0.8.0 (unreleased)

- Added support for inline filtering with IVFFlat
+- Added support for iterative index scans
 - Added casts for arrays to `sparsevec`
 - Improved cost estimation
- Improved performance of HNSW inserts and on-disk index builds
 - Reduced memory usage for HNSW index scans
 - Dropped support for Postgres 12

--- a/2
+++ b/2
@@ -1,4 +1,4 @@
-ARG PG_MAJOR=17
+ARG PG_MAJOR=16
 FROM postgres:$PG_MAJOR
 ARG PG_MAJOR

--- a/2
+++ b/2
@@ -66,7 +66,7 @@ dist:
 	git archive --format zip --prefix=$(EXTENSION)-$(EXTVERSION)/ --output dist/$(EXTENSION)-$(EXTVERSION).zip master

 # for Docker
-PG_MAJOR ?= 17
+PG_MAJOR ?= 16

 .PHONY: docker

--- a/README.md
+++ b/README.md
@@ -52,8 +52,6 @@ nmake /F Makefile.win
 nmake /F Makefile.win install
 ```

-Note: Postgres 17 is not supported yet due to an upstream issue
-
 See the [installation notes](#installation-notes---windows) if you run into issues

 You can also install it with [Docker](#docker) or [conda-forge](#conda-forge).
@@ -102,8 +100,6 @@ Or add a vector column to an existing table
 ALTER TABLE items ADD COLUMN embedding vector(3);
 ```

-Also supports [half-precision](#half-precision-vectors), [binary](#binary-vectors), and [sparse](#sparse-vectors) vectors
-
 Insert vectors

 ```sql
@@ -149,8 +145,6 @@ Supported distance functions are:
 - `<#>` - (negative) inner product
 - `<=>` - cosine distance
 - `<+>` - L1 distance (added in 0.7.0)
- `<~>` - Hamming distance (binary vectors, added in 0.7.0)
- `<%>` - Jaccard distance (binary vectors, added in 0.7.0)

 Get the nearest neighbors to a row

@@ -439,12 +433,6 @@ Create an index on one [or more](https://www.postgresql.org/docs/current/indexes
 CREATE INDEX ON items (category_id);
 ```

-Or a composite IVFFlat index for approximate search (added in 0.8.0)
-
-```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops, category_id) WITH (lists = 100);
-```
-
 Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search

 ```sql
@@ -457,6 +445,63 @@ Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html
 CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
 ```

+## Streaming Queries [unreleased]
+
+*Added in 0.8.0*
+
+With approximate indexes, you can end up with less results than expected due to filtering conditions in the query.
+
+Starting with 0.8.0, you can enable streaming queries. If too few results from the initial index scan match the query filters, it will resume scanning until enough results are found. This can significantly improve recall (at the cost of speed).
+
+```tsql
+SET hnsw.streaming = on;
+-- or
+SET ivfflat.streaming = on;
+```
+
+### Streaming Options
+
+Since scanning a large portion of the index is expensive, there are options to control when the scan ends.
+
+#### HNSW
+
+Specify the max number of additional tuples visited
+
+```sql
+SET hnsw.ef_stream = 10000;
+```
+
+The scan will also end if reaches `work_mem`, at which point a notice is shown
+
+```text
+NOTICE:  hnsw index scan exceeded work_mem after 50000 tuples
+HINT:  Increase work_mem to scan more tuples.
+```
+
+Adjust this with:
+
+```sql
+SET work_mem = '8MB';
+```
+
+#### IVFFlat
+
+Specify the max number of probes
+
+```sql
+SET ivfflat.max_probes = 100;
+```
+
+### Streaming Order
+
+With streaming queries, it’s possible for rows to be slightly out of order by distance. For strict ordering, use:
+
+```sql
+WITH approx_order AS MATERIALIZED (
+    SELECT *, embedding <-> '[1,2,3]' AS distance FROM items WHERE ... ORDER BY distance LIMIT 5
+) SELECT * FROM approx_order ORDER BY distance;
+```
+
 ## Half-Precision Vectors

 *Added in 0.7.0*
@@ -995,7 +1040,7 @@ l2_normalize(sparsevec) → sparsevec | Normalize with Euclidean norm | 0.7.0
 If your machine has multiple Postgres installations, specify the path to [pg_config](https://www.postgresql.org/docs/current/app-pgconfig.html) with:

 ```sh
-export PG_CONFIG=/Library/PostgreSQL/17/bin/pg_config
+export PG_CONFIG=/Library/PostgreSQL/16/bin/pg_config
 ```

 Then re-run the installation instructions (run `make clean` before `make` if needed). If `sudo` is needed for `make install`, use:
@@ -1006,11 +1051,11 @@ sudo --preserve-env=PG_CONFIG make install

 A few common paths on Mac are:

- EDB installer - `/Library/PostgreSQL/17/bin/pg_config`
- Homebrew (arm64) - `/opt/homebrew/opt/postgresql@17/bin/pg_config`
- Homebrew (x86-64) - `/usr/local/opt/postgresql@17/bin/pg_config`
+- EDB installer - `/Library/PostgreSQL/16/bin/pg_config`
+- Homebrew (arm64) - `/opt/homebrew/opt/postgresql@16/bin/pg_config`
+- Homebrew (x86-64) - `/usr/local/opt/postgresql@16/bin/pg_config`

-Note: Replace `17` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### Missing Header

@@ -1019,10 +1064,10 @@ If compilation fails with `fatal error: postgres.h: No such file or directory`,
 For Ubuntu and Debian, use:

 ```sh
-sudo apt install postgresql-server-dev-17
+sudo apt install postgresql-server-dev-16
 ```

-Note: Replace `17` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### Missing SDK

@@ -1055,17 +1100,17 @@ If installation fails with `Access is denied`, re-run the installation instructi
 Get the [Docker image](https://hub.docker.com/r/pgvector/pgvector) with:

 ```sh
-docker pull pgvector/pgvector:pg17
+docker pull pgvector/pgvector:pg16
 ```

-This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (replace `17` with your Postgres server version, and run it the same way).
+This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (replace `16` with your Postgres server version, and run it the same way).

 You can also build the image manually:

 ```sh
 git clone --branch v0.7.4 https://github.com/pgvector/pgvector.git
 cd pgvector
-docker build --pull --build-arg PG_MAJOR=17 -t myuser/pgvector .
+docker build --pull --build-arg PG_MAJOR=16 -t myuser/pgvector .
 ```

 ### Homebrew
@@ -1076,7 +1121,7 @@ With Homebrew Postgres, you can use:
 brew install pgvector
 ```

-Note: This only adds it to the `postgresql@17` and `postgresql@14` formulas
+Note: This only adds it to the `postgresql@14` formula

 ### PGXN

@@ -1091,22 +1136,22 @@ pgxn install vector
 Debian and Ubuntu packages are available from the [PostgreSQL APT Repository](https://wiki.postgresql.org/wiki/Apt). Follow the [setup instructions](https://wiki.postgresql.org/wiki/Apt#Quickstart) and run:

 ```sh
-sudo apt install postgresql-17-pgvector
+sudo apt install postgresql-16-pgvector
 ```

-Note: Replace `17` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### Yum

 RPM packages are available from the [PostgreSQL Yum Repository](https://yum.postgresql.org/). Follow the [setup instructions](https://www.postgresql.org/download/linux/redhat/) for your distribution and run:

 ```sh
-sudo yum install pgvector_17
+sudo yum install pgvector_16
 # or
-sudo dnf install pgvector_17
+sudo dnf install pgvector_16
 ```

-Note: Replace `17` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### pkg

--- a/sql/vector--0.7.4--0.8.0.sql
+++ b/sql/vector--0.7.4--0.8.0.sql
@@ -24,5 +24,3 @@ CREATE CAST (double precision[] AS sparsevec)

 CREATE CAST (numeric[] AS sparsevec)
 	WITH FUNCTION array_to_sparsevec(numeric[], integer, boolean) AS ASSIGNMENT;
-
-- TODO add ivfflat attributes
--- a/sql/vector.sql
+++ b/sql/vector.sql
@@ -916,13 +916,3 @@ CREATE OPERATOR CLASS sparsevec_l1_ops
 	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 l1_distance(sparsevec, sparsevec),
 	FUNCTION 3 hnsw_sparsevec_support(internal);
-
-- ivfflat attributes
-
-CREATE OPERATOR CLASS vector_integer_ops
-	DEFAULT FOR TYPE integer USING ivfflat AS
-	OPERATOR 2 < ,
-	OPERATOR 3 <= ,
-	OPERATOR 4 = ,
-	OPERATOR 5 >= ,
-	OPERATOR 6 > ;
--- a/src/hnsw.c
+++ b/src/hnsw.c
@@ -19,6 +19,8 @@
 #endif

 int			hnsw_ef_search;
+int			hnsw_ef_stream;
+bool		hnsw_streaming;
 int			hnsw_lock_tranche_id;
 static relopt_kind hnsw_relopt_kind;

@@ -69,6 +71,17 @@ HnswInit(void)
 							"Valid range is 1..1000.", &hnsw_ef_search,
 							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);

+	/* TODO Figure out name */
+	DefineCustomBoolVariable("hnsw.streaming", "Use streaming mode",
+							 NULL, &hnsw_streaming,
+							 HNSW_DEFAULT_STREAMING, PGC_USERSET, 0, NULL, NULL, NULL);
+
+	/* TODO Figure out name */
+	/* TODO Use same value as ivfflat.max_probes for "all" */
+	DefineCustomIntVariable("hnsw.ef_stream", "Sets the max number of additional candidates to visit for streaming search",
+							"-1 means all", &hnsw_ef_stream,
+							HNSW_DEFAULT_EF_STREAM, HNSW_MIN_EF_STREAM, HNSW_MAX_EF_STREAM, PGC_USERSET, 0, NULL, NULL, NULL);
+
 	MarkGUCPrefixReserved("hnsw");
 }

@@ -89,6 +102,33 @@ hnswbuildphasename(int64 phasenum)
 	}
 }

+/*
+ * Estimate ef needed for iterative scans
+ */
+static int
+EstimateEf(PlannerInfo *root, IndexPath *path)
+{
+	double		selectivity = 1;
+	ListCell   *lc;
+
+	/* Cannot estimate without limit */
+	/* limit_tuples includes offset */
+	if (root->limit_tuples < 0)
+		return 0;
+
+	/* Get the selectivity of non-index conditions */
+	foreach(lc, path->indexinfo->indrestrictinfo)
+	{
+		RestrictInfo *rinfo = lfirst(lc);
+
+		/* Skip DEFAULT_INEQ_SEL since it may be a distance filter */
+		if (rinfo->norm_selec >= 0 && rinfo->norm_selec <= 1 && rinfo->norm_selec != (Selectivity) DEFAULT_INEQ_SEL)
+			selectivity *= rinfo->norm_selec;
+	}
+
+	return root->limit_tuples / Max(selectivity, 0.00001);
+}
+
 /*
 * Estimate the cost of an index scan
 */
@@ -100,8 +140,11 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 {
 	GenericCosts costs;
 	int			m;
-	double		ratio;
-	double		startupPages;
+	int			ef;
+	int			entryLevel;
+	int			layer0TuplesMax;
+	double		layer0Selectivity;
+	double		scalingFactor = 0.55;
 	double		spc_seq_page_cost;
 	Relation	index;

@@ -118,12 +161,12 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,

 	MemSet(&costs, 0, sizeof(costs));

-	genericcostestimate(root, path, loop_count, &costs);
-
 	index = index_open(path->indexinfo->indexoid, NoLock);
 	HnswGetMetaPageInfo(index, &m, NULL);
 	index_close(index, NoLock);

+	ef = hnsw_streaming ? Max(hnsw_ef_search, EstimateEf(root, path)) : hnsw_ef_search;
+
 	/*
 	 * HNSW cost estimation follows a formula that accounts for the total
 	 * number of tuples indexed combined with the parameters that most
@@ -151,38 +194,30 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	 * at L0, accounting for previously visited tuples, multiplied by the
 	 * "scalingFactor" (currently hardcoded).
 	 */
-	if (path->indexinfo->tuples > 0)
-	{
-		double		scalingFactor = 0.55;
-		int			entryLevel = (int) (log(path->indexinfo->tuples) * HnswGetMl(m));
-		int			layer0TuplesMax = HnswGetLayerM(m, 0) * hnsw_ef_search;
-		double		layer0Selectivity = scalingFactor * log(path->indexinfo->tuples) / (log(m) * (1 + log(hnsw_ef_search)));
+	entryLevel = (int) (log(path->indexinfo->tuples + 1) * HnswGetMl(m));
+	layer0TuplesMax = HnswGetLayerM(m, 0) * ef;
+	layer0Selectivity = (scalingFactor * log(path->indexinfo->tuples + 1)) /
+		(log(m) * (1 + log(ef)));

-		ratio = (entryLevel * m + layer0TuplesMax * layer0Selectivity) / path->indexinfo->tuples;
+	costs.numIndexTuples = (entryLevel * m) +
+		(layer0TuplesMax * layer0Selectivity);

-		if (ratio > 1)
-			ratio = 1;
-	}
-	else
-		ratio = 1;
+	genericcostestimate(root, path, loop_count, &costs);

 	get_tablespace_page_costs(path->indexinfo->reltablespace, NULL, &spc_seq_page_cost);

-	/* Startup cost is cost before returning the first row */
-	costs.indexStartupCost = costs.indexTotalCost * ratio;
-
 	/* Adjust cost if needed since TOAST not included in seq scan cost */
-	startupPages = costs.numIndexPages * ratio;
-	if (startupPages > path->indexinfo->rel->pages && ratio < 0.5)
+	if (costs.numIndexPages > path->indexinfo->rel->pages)
 	{
 		/* Change all page cost from random to sequential */
-		costs.indexStartupCost -= startupPages * (costs.spc_random_page_cost - spc_seq_page_cost);
+		costs.indexTotalCost -= costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);

 		/* Remove cost of extra pages */
-		costs.indexStartupCost -= (startupPages - path->indexinfo->rel->pages) * spc_seq_page_cost;
+		costs.indexTotalCost -= (costs.numIndexPages - path->indexinfo->rel->pages) * spc_seq_page_cost;
 	}

-	*indexStartupCost = costs.indexStartupCost;
+	/* Use total cost since most work happens before first tuple is returned */
+	*indexStartupCost = costs.indexTotalCost;
 	*indexTotalCost = costs.indexTotalCost;
 	*indexSelectivity = costs.indexSelectivity;
 	*indexCorrelation = costs.indexCorrelation;
--- a/src/hnsw.h
+++ b/src/hnsw.h
@@ -12,6 +12,10 @@
 #include "utils/sampling.h"
 #include "vector.h"

+#ifdef HNSW_BENCH
+#include "portability/instr_time.h"
+#endif
+
 #define HNSW_MAX_DIM 2000
 #define HNSW_MAX_NNZ 1000

@@ -42,6 +46,10 @@
 #define HNSW_DEFAULT_EF_SEARCH	40
 #define HNSW_MIN_EF_SEARCH		1
 #define HNSW_MAX_EF_SEARCH		1000
+#define HNSW_DEFAULT_STREAMING	false
+#define HNSW_DEFAULT_EF_STREAM	-1
+#define HNSW_MIN_EF_STREAM		-1
+#define HNSW_MAX_EF_STREAM		INT_MAX

 /* Tuple types */
 #define HNSW_ELEMENT_TUPLE_TYPE  1
@@ -68,6 +76,21 @@
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
 #define HnswPageGetMeta(page)	((HnswMetaPageData *) PageGetContents(page))

+#ifdef HNSW_BENCH
+#define HnswBench(name, code) \
+	do { \
+		instr_time	start; \
+		instr_time	duration; \
+		INSTR_TIME_SET_CURRENT(start); \
+		(code); \
+		INSTR_TIME_SET_CURRENT(duration); \
+		INSTR_TIME_SUBTRACT(duration, start); \
+		elog(INFO, "%s: %.3f ms", name, INSTR_TIME_GET_MILLISEC(duration)); \
+	} while (0)
+#else
+#define HnswBench(name, code) (code)
+#endif
+
 #if PG_VERSION_NUM >= 150000
 #define RandomDouble() pg_prng_double(&pg_global_prng_state)
 #define SeedRandom(seed) pg_prng_seed(&pg_global_prng_state, seed)
@@ -88,9 +111,6 @@
 /* Ensure fits on page and in uint8 */
 #define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / (m)) - 2, 255)

-#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
-#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
-
 #define HnswGetValue(base, element) PointerGetDatum(HnswPtrAccess(base, (element)->value))

 #if PG_VERSION_NUM < 140005
@@ -109,6 +129,8 @@

 /* Variables */
 extern int	hnsw_ef_search;
+extern int	hnsw_ef_stream;
+extern bool hnsw_streaming;
 extern int	hnsw_lock_tranche_id;

 typedef struct HnswElementData HnswElementData;
@@ -132,6 +154,7 @@ struct HnswElementData
 	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
+	uint8		version;
 	uint32		hash;
 	HnswNeighborsPtr neighbors;
 	BlockNumber blkno;
@@ -163,9 +186,12 @@ typedef struct HnswSearchCandidate
 	pairingheap_node c_node;
 	pairingheap_node w_node;
 	HnswElementPtr element;
-	double		distance;
+	float		distance;
 }			HnswSearchCandidate;

+#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
+#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
+
 /* HNSW index options */
 typedef struct HnswOptions
 {
@@ -188,8 +214,8 @@ typedef struct HnswGraph

 	/* Allocations state */
 	LWLock		allocatorLock;
-	Size		memoryUsed;
-	Size		memoryTotal;
+	long		memoryUsed;
+	long		memoryTotal;

 	/* Flushed state */
 	LWLock		flushLock;
@@ -240,18 +266,6 @@ typedef struct HnswTypeInfo
 	void		(*checkValue) (Pointer v);
 }			HnswTypeInfo;

-typedef struct HnswSupport
-{
-	FmgrInfo   *procinfo;
-	FmgrInfo   *normprocinfo;
-	Oid			collation;
-}			HnswSupport;
-
-typedef struct HnswQuery
-{
-	Datum		value;
-}			HnswQuery;
-
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -271,7 +285,9 @@ typedef struct HnswBuildState
 	double		reltuples;

 	/* Support functions */
-	HnswSupport support;
+	FmgrInfo   *procinfo;
+	FmgrInfo   *normprocinfo;
+	Oid			collation;

 	/* Variables */
 	HnswGraph	graphData;
@@ -319,10 +335,10 @@ typedef struct HnswElementTupleData
 	uint8		type;
 	uint8		level;
 	uint8		deleted;
-	uint8		unused;
+	uint8		version;
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
-	uint16		unused2;
+	uint16		unused;
 	Vector		data;
 }			HnswElementTupleData;

@@ -331,7 +347,7 @@ typedef HnswElementTupleData * HnswElementTuple;
 typedef struct HnswNeighborTupleData
 {
 	uint8		type;
-	uint8		unused;
+	uint8		version;
 	uint16		count;
 	ItemPointerData indextids[FLEXIBLE_ARRAY_MEMBER];
 }			HnswNeighborTupleData;
@@ -345,21 +361,23 @@ typedef union
 	struct tidhash_hash *tids;
 }			visited_hash;

-typedef union
-{
-	HnswElement element;
-	ItemPointerData indextid;
-}			HnswUnvisited;
-
 typedef struct HnswScanOpaqueData
 {
 	const		HnswTypeInfo *typeInfo;
 	bool		first;
 	List	   *w;
+	visited_hash v;
+	pairingheap *discarded;
+	Datum		q;
+	int			m;
+	int64		tuples;
+	double		previousDistance;
 	MemoryContext tmpCtx;

 	/* Support functions */
-	HnswSupport support;
+	FmgrInfo   *procinfo;
+	FmgrInfo   *normprocinfo;
+	Oid			collation;
 }			HnswScanOpaqueData;

 typedef HnswScanOpaqueData * HnswScanOpaque;
@@ -377,7 +395,8 @@ typedef struct HnswVacuumState
 	int			efConstruction;

 	/* Support functions */
-	HnswSupport support;
+	FmgrInfo   *procinfo;
+	Oid			collation;

 	/* Variables */
 	struct tidhash_hash *deleted;
@@ -393,33 +412,30 @@ typedef struct HnswVacuumState
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
-void		HnswInitSupport(HnswSupport * support, Relation index);
 Datum		HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value);
-bool		HnswCheckNorm(HnswSupport * support, Datum value);
+bool		HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);
-List	   *HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement);
+List	   *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples);
 HnswElement HnswGetEntryPoint(Relation index);
 void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 void	   *HnswAlloc(HnswAllocator * allocator, Size size);
 HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
 HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
-void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing);
-HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, HnswQuery * q, Relation rel, HnswSupport * support, bool loadVec);
+void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
+HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
 void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
 void		HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
 void		HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
-HnswNeighborArray *HnswInitNeighborArray(int lm, HnswAllocator * allocator);
 void		HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * alloc);
-bool		HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building);
-void		HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e, int m, bool checkExisting, bool building);
+bool		HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building);
+void		HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building);
 void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
-void		HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance);
-bool		HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support);
+void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, float *maxDistance);
 void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
-void		HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support);
-bool		HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation index, int m, int lm, int lc);
+void		HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int lm, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
+void		HnswLoadNeighbors(HnswElement element, Relation index, int m);
 void		HnswInitLockTranche(void);
 const		HnswTypeInfo *HnswGetTypeInfo(Relation index);
 PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
--- a/src/hnswbuild.c
+++ b/src/hnswbuild.c
@@ -366,7 +366,7 @@ AddElementInMemory(char *base, HnswGraph * graph, HnswElement element)
 * Update neighbors
 */
 static void
-UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
+UpdateNeighborsInMemory(char *base, FmgrInfo *procinfo, Oid collation, HnswElement e, int m)
 {
 	for (int lc = e->level; lc >= 0; lc--)
 	{
@@ -388,7 +388,7 @@ UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
 			Assert(neighborElement);

 			LWLockAcquire(&neighborElement->lock, LW_EXCLUSIVE);
-			HnswUpdateConnection(base, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, NULL, support);
+			HnswUpdateConnection(base, e, hc, lm, lc, NULL, NULL, procinfo, collation);
 			LWLockRelease(&neighborElement->lock);
 		}
 	}
@@ -398,7 +398,7 @@ UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
 * Update graph in memory
 */
 static void
-UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, int efConstruction, HnswElement entryPoint, HnswBuildState * buildstate)
+UpdateGraphInMemory(FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, HnswBuildState * buildstate)
 {
 	HnswGraph  *graph = buildstate->graph;
 	char	   *base = buildstate->hnswarea;
@@ -411,7 +411,7 @@ UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, int efCon
 	AddElementInMemory(base, graph, element);

 	/* Update neighbors */
-	UpdateNeighborsInMemory(base, support, element, m);
+	UpdateNeighborsInMemory(base, procinfo, collation, element, m);

 	/* Update entry point if needed (already have lock) */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -424,8 +424,9 @@ UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, int efCon
 static void
 InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 {
+	FmgrInfo   *procinfo = buildstate->procinfo;
+	Oid			collation = buildstate->collation;
 	HnswGraph  *graph = buildstate->graph;
-	HnswSupport *support = &buildstate->support;
 	HnswElement entryPoint;
 	LWLock	   *entryLock = &graph->entryLock;
 	LWLock	   *entryWaitLock = &graph->entryWaitLock;
@@ -457,10 +458,10 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 	}

 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false);
+	HnswFindElementNeighbors(base, element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);

 	/* Update graph in memory */
-	UpdateGraphInMemory(support, element, m, efConstruction, entryPoint, buildstate);
+	UpdateGraphInMemory(procinfo, collation, element, m, efConstruction, entryPoint, buildstate);

 	/* Release entry lock */
 	LWLockRelease(entryLock);
@@ -472,19 +473,30 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 static bool
 InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, HnswBuildState * buildstate)
 {
+	const		HnswTypeInfo *typeInfo = buildstate->typeInfo;
 	HnswGraph  *graph = buildstate->graph;
 	HnswElement element;
 	HnswAllocator *allocator = &buildstate->allocator;
-	HnswSupport *support = &buildstate->support;
 	Size		valueSize;
 	Pointer		valuePtr;
 	LWLock	   *flushLock = &graph->flushLock;
 	char	   *base = buildstate->hnswarea;
-	Datum		value;

-	/* Form index value */
-	if (!HnswFormIndexValue(&value, values, isnull, buildstate->typeInfo, support))
-		return false;
+	/* Detoast once for all calls */
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Check value */
+	if (typeInfo->checkValue != NULL)
+		typeInfo->checkValue(DatumGetPointer(value));
+
+	/* Normalize if needed */
+	if (buildstate->normprocinfo != NULL)
+	{
+		if (!HnswCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
+			return false;
+
+		value = HnswNormValue(typeInfo, buildstate->collation, value);
+	}

 	/* Get datum size */
 	valueSize = VARSIZE_ANY(DatumGetPointer(value));
@@ -497,7 +509,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	{
 		LWLockRelease(flushLock);

-		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
+		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, true);
 	}

 	/*
@@ -529,7 +541,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn

 		LWLockRelease(flushLock);

-		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
+		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, true);
 	}

 	/* Ok, we can proceed to allocate the element */
@@ -595,7 +607,7 @@ BuildCallback(Relation index, ItemPointer tid, Datum *values,
 * Initialize the graph
 */
 static void
-InitGraph(HnswGraph * graph, char *base, Size memoryTotal)
+InitGraph(HnswGraph * graph, char *base, long memoryTotal)
 {
 	/* Initialize the lock tranche if needed */
 	HnswInitLockTranche();
@@ -692,9 +704,11 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->indtuples = 0;

 	/* Get support functions */
-	HnswInitSupport(&buildstate->support, index);
+	buildstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	buildstate->collation = index->rd_indcollation[0];

-	InitGraph(&buildstate->graphData, NULL, (Size) maintenance_work_mem * 1024L);
+	InitGraph(&buildstate->graphData, NULL, maintenance_work_mem * 1024L);
 	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
--- a/src/hnswinsert.c
+++ b/src/hnswinsert.c
@@ -36,7 +36,7 @@ GetInsertPage(Relation index)
 * Check for a free offset
 */
 static bool
-HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size etupSize, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage)
+HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size etupSize, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage, uint8 *tupleVersion)
 {
 	OffsetNumber offno;
 	OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
@@ -98,6 +98,7 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
 			{
 				*freeOffno = offno;
 				*freeNeighborOffno = neighborOffno;
+				*tupleVersion = etup->version;
 				return true;
 			}
 			else if (*nbuf != buf)
@@ -153,6 +154,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 	OffsetNumber freeOffno = InvalidOffsetNumber;
 	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
 	BlockNumber newInsertPage = InvalidBlockNumber;
+	uint8		tupleVersion;
 	char	   *base = NULL;

 	/* Calculate sizes */
@@ -202,7 +204,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 		}

 		/* Next, try space from a deleted element */
-		if (HnswFreeOffset(index, buf, page, e, etupSize, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
+		if (HnswFreeOffset(index, buf, page, e, etupSize, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage, &tupleVersion))
 		{
 			if (nbuf != buf)
 			{
@@ -212,6 +214,10 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 					npage = GenericXLogRegisterBuffer(state, nbuf, 0);
 			}

+			/* Set tuple version */
+			etup->version = tupleVersion;
+			ntup->version = tupleVersion;
+
 			break;
 		}

@@ -334,107 +340,6 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 		*updatedInsertPage = newInsertPage;
 }

-/*
- * Load neighbors
- */
-static HnswNeighborArray *
-HnswLoadNeighbors(HnswElement element, Relation index, int m, int lm, int lc)
-{
-	char	   *base = NULL;
-	HnswNeighborArray *neighbors = HnswInitNeighborArray(lm, NULL);
-	ItemPointerData indextids[HNSW_MAX_M * 2];
-
-	if (!HnswLoadNeighborTids(element, indextids, index, m, lm, lc))
-		return neighbors;
-
-	for (int i = 0; i < lm; i++)
-	{
-		ItemPointer indextid = &indextids[i];
-		HnswElement e;
-		HnswCandidate *hc;
-
-		if (!ItemPointerIsValid(indextid))
-			break;
-
-		e = HnswInitElementFromBlock(ItemPointerGetBlockNumber(indextid), ItemPointerGetOffsetNumber(indextid));
-		hc = &neighbors->items[neighbors->length++];
-		HnswPtrStore(base, hc->element, e);
-	}
-
-	return neighbors;
-}
-
-/*
- * Load elements for insert
- */
-static void
-LoadElementsForInsert(HnswNeighborArray * neighbors, HnswQuery * q, int *idx, Relation index, HnswSupport * support)
-{
-	char	   *base = NULL;
-
-	for (int i = 0; i < neighbors->length; i++)
-	{
-		HnswCandidate *hc = &neighbors->items[i];
-		HnswElement element = HnswPtrAccess(base, hc->element);
-		double		distance;
-
-		HnswLoadElement(element, &distance, q, index, support, true, NULL);
-		hc->distance = distance;
-
-		/* Prune element if being deleted */
-		if (element->heaptidsLength == 0)
-		{
-			*idx = i;
-			break;
-		}
-	}
-}
-
-/*
- * Get update index
- */
-static int
-GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int m, int lm, int lc, Relation index, HnswSupport * support, MemoryContext updateCtx)
-{
-	char	   *base = NULL;
-	int			idx = -1;
-	HnswNeighborArray *neighbors;
-	MemoryContext oldCtx = MemoryContextSwitchTo(updateCtx);
-
-	/*
-	 * Get latest neighbors since they may have changed. Do not lock yet since
-	 * selecting neighbors can take time. Could use optimistic locking to
-	 * retry if another update occurs before getting exclusive lock.
-	 */
-	neighbors = HnswLoadNeighbors(element, index, m, lm, lc);
-
-	/*
-	 * Could improve performance for vacuuming by checking neighbors against
-	 * list of elements being deleted to find index. It's important to exclude
-	 * already deleted elements for this since they can be replaced at any
-	 * time.
-	 */
-
-	if (neighbors->length < lm)
-		idx = -2;
-	else
-	{
-		HnswQuery	q;
-
-		q.value = HnswGetValue(base, element);
-
-		LoadElementsForInsert(neighbors, &q, &idx, index, support);
-
-		if (idx == -1)
-			HnswUpdateConnection(base, neighbors, newElement, distance, lm, &idx, index, support);
-	}
-
-	MemoryContextSwitchTo(oldCtx);
-	MemoryContextReset(updateCtx);
-
-	return idx;
-}
-
 /*
 * Check if connection already exists
 */
@@ -455,94 +360,14 @@ ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
 	return false;
 }

-/*
- * Update neighbor
- */
-static void
-UpdateNeighborOnDisk(HnswElement element, HnswElement newElement, int idx, int m, int lm, int lc, Relation index, bool checkExisting, bool building)
-{
-	Buffer		buf;
-	Page		page;
-	GenericXLogState *state;
-	HnswNeighborTuple ntup;
-	int			startIdx;
-	OffsetNumber offno = element->neighborOffno;
-
-	/* Register page */
-	buf = ReadBuffer(index, element->neighborPage);
-	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-	if (building)
-	{
-		state = NULL;
-		page = BufferGetPage(buf);
-	}
-	else
-	{
-		state = GenericXLogStart(index);
-		page = GenericXLogRegisterBuffer(state, buf, 0);
-	}
-
-	/* Get tuple */
-	ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, offno));
-
-	/* Calculate index for update */
-	startIdx = (element->level - lc) * m;
-
-	/* Check for existing connection */
-	if (checkExisting && ConnectionExists(newElement, ntup, startIdx, lm))
-		idx = -1;
-	else if (idx == -2)
-	{
-		/* Find free offset if still exists */
-		/* TODO Retry updating connections if not */
-		for (int j = 0; j < lm; j++)
-		{
-			if (!ItemPointerIsValid(&ntup->indextids[startIdx + j]))
-			{
-				idx = startIdx + j;
-				break;
-			}
-		}
-	}
-	else
-		idx += startIdx;
-
-	/* Make robust to issues */
-	if (idx >= 0 && idx < ntup->count)
-	{
-		ItemPointer indextid = &ntup->indextids[idx];
-
-		/* Update neighbor on the buffer */
-		ItemPointerSet(indextid, newElement->blkno, newElement->offno);
-
-		/* Commit */
-		if (building)
-			MarkBufferDirty(buf);
-		else
-			GenericXLogFinish(state);
-	}
-	else if (!building)
-		GenericXLogAbort(state);
-
-	UnlockReleaseBuffer(buf);
-}
-
 /*
 * Update neighbors
 */
 void
-HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e, int m, bool checkExisting, bool building)
+HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
 {
 	char	   *base = NULL;

-	/* Use separate memory context to improve performance for larger vectors */
-	MemoryContext updateCtx = GenerationContextCreate(CurrentMemoryContext,
-													  "Hnsw insert update context",
-#if PG_VERSION_NUM >= 150000
-													  128 * 1024, 128 * 1024,
-#endif
-													  128 * 1024);
-
 	for (int lc = e->level; lc >= 0; lc--)
 	{
 		int			lm = HnswGetLayerM(m, lc);
@@ -551,20 +376,96 @@ HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e,
 		for (int i = 0; i < neighbors->length; i++)
 		{
 			HnswCandidate *hc = &neighbors->items[i];
+			Buffer		buf;
+			Page		page;
+			GenericXLogState *state;
+			HnswNeighborTuple ntup;
+			int			idx = -1;
+			int			startIdx;
 			HnswElement neighborElement = HnswPtrAccess(base, hc->element);
-			int			idx;
+			OffsetNumber offno = neighborElement->neighborOffno;

-			idx = GetUpdateIndex(neighborElement, e, hc->distance, m, lm, lc, index, support, updateCtx);
+			/*
+			 * Get latest neighbors since they may have changed. Do not lock
+			 * yet since selecting neighbors can take time. Could use
+			 * optimistic locking to retry if another update occurs before
+			 * getting exclusive lock.
+			 */
+			HnswLoadNeighbors(neighborElement, index, m);
+
+			/*
+			 * Could improve performance for vacuuming by checking neighbors
+			 * against list of elements being deleted to find index. It's
+			 * important to exclude already deleted elements for this since
+			 * they can be replaced at any time.
+			 */
+
+			/* Select neighbors */
+			HnswUpdateConnection(NULL, e, hc, lm, lc, &idx, index, procinfo, collation);

 			/* New element was not selected as a neighbor */
 			if (idx == -1)
 				continue;

-			UpdateNeighborOnDisk(neighborElement, e, idx, m, lm, lc, index, checkExisting, building);
+			/* Register page */
+			buf = ReadBuffer(index, neighborElement->neighborPage);
+			LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+			if (building)
+			{
+				state = NULL;
+				page = BufferGetPage(buf);
+			}
+			else
+			{
+				state = GenericXLogStart(index);
+				page = GenericXLogRegisterBuffer(state, buf, 0);
+			}
+
+			/* Get tuple */
+			ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, offno));
+
+			/* Calculate index for update */
+			startIdx = (neighborElement->level - lc) * m;
+
+			/* Check for existing connection */
+			if (checkExisting && ConnectionExists(e, ntup, startIdx, lm))
+				idx = -1;
+			else if (idx == -2)
+			{
+				/* Find free offset if still exists */
+				/* TODO Retry updating connections if not */
+				for (int j = 0; j < lm; j++)
+				{
+					if (!ItemPointerIsValid(&ntup->indextids[startIdx + j]))
+					{
+						idx = startIdx + j;
+						break;
+					}
+				}
+			}
+			else
+				idx += startIdx;
+
+			/* Make robust to issues */
+			if (idx >= 0 && idx < ntup->count)
+			{
+				ItemPointer indextid = &ntup->indextids[idx];
+
+				/* Update neighbor on the buffer */
+				ItemPointerSet(indextid, e->blkno, e->offno);
+
+				/* Commit */
+				if (building)
+					MarkBufferDirty(buf);
+				else
+					GenericXLogFinish(state);
+			}
+			else if (!building)
+				GenericXLogAbort(state);
+
+			UnlockReleaseBuffer(buf);
 		}
 	}
-
-	MemoryContextDelete(updateCtx);
 }

 /*
@@ -654,7 +555,7 @@ FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
 * Update graph on disk
 */
 static void
-UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
+UpdateGraphOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;

@@ -670,7 +571,7 @@ UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, in
 		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);

 	/* Update neighbors */
-	HnswUpdateNeighborsOnDisk(index, support, element, m, false, building);
+	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, false, building);

 	/* Update entry point if needed */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -681,12 +582,14 @@ UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, in
 * Insert a tuple into the index
 */
 bool
-HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building)
+HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building)
 {
 	HnswElement entryPoint;
 	HnswElement element;
 	int			m;
 	int			efConstruction = HnswGetEfConstruction(index);
+	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	Oid			collation = index->rd_indcollation[0];
 	LOCKMODE	lockmode = ShareLock;
 	char	   *base = NULL;

@@ -701,7 +604,7 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 	HnswGetMetaPageInfo(index, &m, &entryPoint);

 	/* Create an element */
-	element = HnswInitElement(base, heaptid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
+	element = HnswInitElement(base, heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
 	HnswPtrStore(base, element->value, DatumGetPointer(value));

 	/* Prevent concurrent inserts when likely updating entry point */
@@ -719,10 +622,10 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 	}

 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, false);

 	/* Update graph on disk */
-	UpdateGraphOnDisk(index, support, element, m, efConstruction, entryPoint, building);
+	UpdateGraphOnDisk(index, procinfo, collation, element, m, efConstruction, entryPoint, building);

 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -734,19 +637,31 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 * Insert a tuple into the index
 */
 static void
-HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid)
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid)
 {
 	Datum		value;
 	const		HnswTypeInfo *typeInfo = HnswGetTypeInfo(index);
-	HnswSupport support;
+	FmgrInfo   *normprocinfo;
+	Oid			collation = index->rd_indcollation[0];

-	HnswInitSupport(&support, index);
+	/* Detoast once for all calls */
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));

-	/* Form index value */
-	if (!HnswFormIndexValue(&value, values, isnull, typeInfo, &support))
-		return;
+	/* Check value */
+	if (typeInfo->checkValue != NULL)
+		typeInfo->checkValue(DatumGetPointer(value));

-	HnswInsertTupleOnDisk(index, &support, value, heaptid, false);
+	/* Normalize if needed */
+	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	if (normprocinfo != NULL)
+	{
+		if (!HnswCheckNorm(normprocinfo, collation, value))
+			return;
+
+		value = HnswNormValue(typeInfo, collation, value);
+	}
+
+	HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, false);
 }

 /*
--- a/src/hnswscan.c
+++ b/src/hnswscan.c
@@ -1,5 +1,7 @@
 #include "postgres.h"

+#include <float.h>
+
 #include "access/relscan.h"
 #include "hnsw.h"
 #include "pgstat.h"
@@ -11,35 +13,69 @@
 * Algorithm 5 from paper
 */
 static List *
-GetScanItems(IndexScanDesc scan, Datum value)
+GetScanItems(IndexScanDesc scan, Datum q)
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 	Relation	index = scan->indexRelation;
-	HnswSupport *support = &so->support;
+	FmgrInfo   *procinfo = so->procinfo;
+	Oid			collation = so->collation;
 	List	   *ep;
 	List	   *w;
 	int			m;
 	HnswElement entryPoint;
 	char	   *base = NULL;
-	HnswQuery	q;
-
-	q.value = value;

 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);

+	so->q = q;
+	so->m = m;
+
 	if (entryPoint == NULL)
 		return NIL;

-	ep = list_make1(HnswEntryCandidate(base, entryPoint, &q, index, support, false));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, procinfo, collation, false));

 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, false, NULL);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL, NULL, NULL, true, NULL);
 		ep = w;
 	}

-	return HnswSearchLayer(base, &q, ep, hnsw_ef_search, 0, index, support, m, false, NULL);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL, &so->v, hnsw_streaming ? &so->discarded : NULL, true, &so->tuples);
+}
+
+/*
+ * Resume scan at ground level with discarded candidates
+ */
+static List *
+ResumeScanItems(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	Relation	index = scan->indexRelation;
+	FmgrInfo   *procinfo = so->procinfo;
+	Oid			collation = so->collation;
+	List	   *ep = NIL;
+	char	   *base = NULL;
+	int			batch_size = hnsw_ef_search;
+
+	if (pairingheap_is_empty(so->discarded))
+		return NIL;
+
+	/* Get next batch of candidates */
+	for (int i = 0; i < batch_size; i++)
+	{
+		HnswSearchCandidate *hc;
+
+		if (pairingheap_is_empty(so->discarded))
+			break;
+
+		hc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded));
+
+		ep = lappend(ep, hc);
+	}
+
+	return HnswSearchLayer(base, so->q, ep, batch_size, 0, index, procinfo, collation, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
 }

 /*
@@ -62,8 +98,8 @@ GetScanValue(IndexScanDesc scan)
 		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));

 		/* Normalize if needed */
-		if (so->support.normprocinfo != NULL)
-			value = HnswNormValue(so->typeInfo, so->support.collation, value);
+		if (so->normprocinfo != NULL)
+			value = HnswNormValue(so->typeInfo, so->collation, value);
 	}

 	return value;
@@ -83,12 +119,16 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
 	so->typeInfo = HnswGetTypeInfo(index);
 	so->first = true;
+	so->v.tids = NULL;
+	so->discarded = NULL;
 	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 									   "Hnsw scan temporary context",
 									   ALLOCSET_DEFAULT_SIZES);

 	/* Set support functions */
-	HnswInitSupport(&so->support, index);
+	so->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	so->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	so->collation = index->rd_indcollation[0];

 	scan->opaque = so;

@@ -103,7 +143,15 @@ hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int no
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;

+	if (so->v.tids != NULL)
+		tidhash_reset(so->v.tids);
+
+	if (so->discarded != NULL)
+		pairingheap_reset(so->discarded);
+
 	so->first = true;
+	so->tuples = 0;
+	so->previousDistance = -INFINITY;
 	MemoryContextReset(so->tmpCtx);

 	if (keys && scan->numberOfKeys > 0)
@@ -153,7 +201,7 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		 */
 		LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);

-		so->w = GetScanItems(scan, value);
+		HnswBench("scan iteration", so->w = GetScanItems(scan, value));

 		/* Release shared lock */
 		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
@@ -165,22 +213,97 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 #endif
 	}

-	while (list_length(so->w) > 0)
+	for (;;)
 	{
 		char	   *base = NULL;
-		HnswSearchCandidate *sc = llast(so->w);
-		HnswElement element = HnswPtrAccess(base, sc->element);
+		HnswSearchCandidate *hc;
+		HnswElement element;
 		ItemPointer heaptid;

+		if (list_length(so->w) == 0)
+		{
+			if (!hnsw_streaming)
+				break;
+
+			/* Empty index */
+			if (so->discarded == NULL)
+				break;
+
+			/* Reached max number of additional tuples */
+			if (hnsw_ef_stream != -1 && so->tuples >= hnsw_ef_search + hnsw_ef_stream)
+			{
+				if (pairingheap_is_empty(so->discarded))
+					break;
+
+				/* Return remaining tuples */
+				so->w = lappend(so->w, HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded)));
+			}
+			/* Prevent scans from consuming too much memory */
+			else if (MemoryContextMemAllocated(so->tmpCtx, false) > (Size) work_mem * 1024L)
+			{
+				if (pairingheap_is_empty(so->discarded))
+				{
+					ereport(NOTICE,
+							(errmsg("hnsw index scan exceeded work_mem after " INT64_FORMAT " tuples", so->tuples),
+							 errhint("Increase work_mem to scan more tuples.")));
+
+					break;
+				}
+
+				/* Return remaining tuples */
+				so->w = lappend(so->w, HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded)));
+			}
+			else
+			{
+				/*
+				 * Locking ensures when neighbors are read, the elements they
+				 * reference will not be deleted (and replaced) during the
+				 * iteration.
+				 *
+				 * Elements loaded into memory on previous iterations may have
+				 * been deleted (and replaced), so when reading neighbors, the
+				 * element version must be checked.
+				 */
+				LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+				HnswBench("scan iteration", so->w = ResumeScanItems(scan));
+
+				UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+#if defined(HNSW_MEMORY)
+				elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(so->tmpCtx, false) / 1024);
+#endif
+			}
+
+			if (list_length(so->w) == 0)
+				break;
+		}
+
+		hc = llast(so->w);
+		element = HnswPtrAccess(base, hc->element);
+
 		/* Move to next element if no valid heap TIDs */
 		if (element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
+
+			/* Mark memory as free for next iteration */
+			if (hnsw_streaming)
+			{
+				pfree(element);
+				pfree(hc);
+			}
+
 			continue;
 		}

 		heaptid = &element->heaptids[--element->heaptidsLength];

+		if (hc->distance < so->previousDistance)
+			continue;
+
+		so->previousDistance = hc->distance;
+
 		MemoryContextSwitchTo(oldCtx);

 		scan->xs_heaptid = *heaptid;
--- a/src/hnswutils.c
+++ b/src/hnswutils.c
@@ -100,6 +100,12 @@ hash_offset(Size offset)
 #define SH_DEFINE
 #include "lib/simplehash.h"

+typedef union
+{
+	HnswElement element;
+	ItemPointerData indextid;
+}			HnswUnvisited;
+
 /*
 * Get the max number of connections in an upper layer for each element in the index
 */
@@ -140,17 +146,6 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 	return index_getprocinfo(index, 1, procnum);
 }

-/*
- * Init support functions
- */
-void
-HnswInitSupport(HnswSupport * support, Relation index)
-{
-	support->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
-	support->collation = index->rd_indcollation[0];
-	support->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
-}
-
 /*
 * Normalize value
 */
@@ -164,9 +159,9 @@ HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value)
 * Check if non-zero norm
 */
 bool
-HnswCheckNorm(HnswSupport * support, Datum value)
+HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
 {
-	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation, value)) > 0;
+	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
 }

 /*
@@ -195,7 +190,7 @@ HnswInitPage(Buffer buf, Page page)
 /*
 * Allocate a neighbor array
 */
-HnswNeighborArray *
+static HnswNeighborArray *
 HnswInitNeighborArray(int lm, HnswAllocator * allocator)
 {
 	HnswNeighborArray *a = HnswAlloc(allocator, HNSW_NEIGHBOR_ARRAY_SIZE(lm));
@@ -251,6 +246,8 @@ HnswInitElement(char *base, ItemPointer heaptid, int m, double ml, int maxLevel,

 	element->level = level;
 	element->deleted = 0;
+	/* Start at one to make it easier to find issues */
+	element->version = 1;

 	HnswInitNeighbors(base, element, m, allocator);

@@ -392,33 +389,6 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 	UnlockReleaseBuffer(buf);
 }

-/*
- * Form index value
- */
-bool
-HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support)
-{
-	/* Detoast once for all calls */
-	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
-
-	/* Check value */
-	if (typeInfo->checkValue != NULL)
-		typeInfo->checkValue(DatumGetPointer(value));
-
-	/* Normalize if needed */
-	if (support->normprocinfo != NULL)
-	{
-		if (!HnswCheckNorm(support, value))
-			return false;
-
-		value = HnswNormValue(typeInfo, support->collation, value);
-	}
-
-	*out = value;
-
-	return true;
-}
-
 /*
 * Set element tuple, except for neighbor info
 */
@@ -430,6 +400,7 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
 	etup->type = HNSW_ELEMENT_TUPLE_TYPE;
 	etup->level = element->level;
 	etup->deleted = 0;
+	etup->version = element->version;
 	for (int i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (i < element->heaptidsLength)
@@ -472,6 +443,70 @@ HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
 	}

 	ntup->count = idx;
+	ntup->version = e->version;
+}
+
+/*
+ * Load neighbors from page
+ */
+static void
+LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
+{
+	char	   *base = NULL;
+
+	HnswNeighborTuple ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
+	int			neighborCount = (element->level + 2) * m;
+
+	Assert(HnswIsNeighborTuple(ntup));
+
+	HnswInitNeighbors(base, element, m, NULL);
+
+	/* Ensure expected neighbors */
+	if (ntup->count != neighborCount)
+		return;
+
+	for (int i = 0; i < neighborCount; i++)
+	{
+		HnswElement e;
+		int			level;
+		HnswCandidate *hc;
+		ItemPointer indextid;
+		HnswNeighborArray *neighbors;
+
+		indextid = &ntup->indextids[i];
+
+		if (!ItemPointerIsValid(indextid))
+			continue;
+
+		e = HnswInitElementFromBlock(ItemPointerGetBlockNumber(indextid), ItemPointerGetOffsetNumber(indextid));
+
+		/* Calculate level based on offset */
+		level = element->level - i / m;
+		if (level < 0)
+			level = 0;
+
+		neighbors = HnswGetNeighbors(base, element, level);
+		hc = &neighbors->items[neighbors->length++];
+		HnswPtrStore(base, hc->element, e);
+	}
+}
+
+/*
+ * Load neighbors
+ */
+void
+HnswLoadNeighbors(HnswElement element, Relation index, int m)
+{
+	Buffer		buf;
+	Page		page;
+
+	buf = ReadBuffer(index, element->neighborPage);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+
+	LoadNeighborsFromPage(element, index, page, m);
+
+	UnlockReleaseBuffer(buf);
 }

 /*
@@ -482,6 +517,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 {
 	element->level = etup->level;
 	element->deleted = etup->deleted;
+	element->version = etup->version;
 	element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
 	element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
 	element->heaptidsLength = 0;
@@ -507,20 +543,11 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	}
 }

-/*
- * Calculate the distance between values
- */
-static inline double
-HnswGetDistance(Datum a, Datum b, HnswSupport * support)
-{
-	return DatumGetFloat8(FunctionCall2Coll(support->procinfo, support->collation, a, b));
-}
-
 /*
 * Load an element and optionally get its distance from q
 */
 static void
-HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance, HnswElement * element)
+HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, float *maxDistance, HnswElement * element)
 {
 	Buffer		buf;
 	Page		page;
@@ -538,10 +565,10 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, Hns
 	/* Calculate distance */
 	if (distance != NULL)
 	{
-		if (DatumGetPointer(q->value) == NULL)
+		if (DatumGetPointer(*q) == NULL)
 			*distance = 0;
 		else
-			*distance = HnswGetDistance(q->value, PointerGetDatum(&etup->data), support);
+			*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
 	}

 	/* Load element */
@@ -560,37 +587,36 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, Hns
 * Load an element and optionally get its distance from q
 */
 void
-HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
+HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, float *maxDistance)
 {
-	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
+	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, procinfo, collation, loadVec, maxDistance, &element);
 }

 /*
 * Get the distance for an element
 */
-static double
-GetElementDistance(char *base, HnswElement element, HnswQuery * q, HnswSupport * support)
+static float
+GetElementDistance(char *base, HnswElement element, Datum q, FmgrInfo *procinfo, Oid collation)
 {
 	Datum		value = HnswGetValue(base, element);

-	return HnswGetDistance(q->value, value, support);
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, value));
 }

 /*
 * Create a candidate for the entry point
 */
 HnswSearchCandidate *
-HnswEntryCandidate(char *base, HnswElement entryPoint, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec)
+HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
 {
-	HnswSearchCandidate *sc = palloc(sizeof(HnswSearchCandidate));
-	bool		inMemory = index == NULL;
+	HnswSearchCandidate *hc = palloc(sizeof(HnswSearchCandidate));

-	HnswPtrStore(base, sc->element, entryPoint);
-	if (inMemory)
-		sc->distance = GetElementDistance(base, entryPoint, q, support);
+	HnswPtrStore(base, hc->element, entryPoint);
+	if (index == NULL)
+		hc->distance = GetElementDistance(base, entryPoint, q, procinfo, collation);
 	else
-		HnswLoadElement(entryPoint, &sc->distance, q, index, support, loadVec, NULL);
-	return sc;
+		HnswLoadElement(entryPoint, &hc->distance, &q, index, procinfo, collation, loadVec, NULL);
+	return hc;
 }

 /*
@@ -608,6 +634,21 @@ CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, v
 	return 0;
 }

+/*
+ * Compare discarded candidate distances
+ */
+static int
+CompareNearestDiscardedCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
+{
+	if (HnswGetSearchCandidateConst(w_node, a)->distance < HnswGetSearchCandidateConst(w_node, b)->distance)
+		return 1;
+
+	if (HnswGetSearchCandidateConst(w_node, a)->distance > HnswGetSearchCandidateConst(w_node, b)->distance)
+		return -1;
+
+	return 0;
+}
+
 /*
 * Compare candidate distances
 */
@@ -627,9 +668,9 @@ CompareFurthestCandidates(const pairingheap_node *a, const pairingheap_node *b,
 * Init visited
 */
 static inline void
-InitVisited(char *base, visited_hash * v, bool inMemory, int ef, int m)
+InitVisited(char *base, visited_hash * v, Relation index, int ef, int m)
 {
-	if (!inMemory)
+	if (index != NULL)
 		v->tids = tidhash_create(CurrentMemoryContext, ef * m * 2, NULL);
 	else if (base != NULL)
 		v->offsets = offsethash_create(CurrentMemoryContext, ef * m * 2, NULL);
@@ -641,9 +682,9 @@ InitVisited(char *base, visited_hash * v, bool inMemory, int ef, int m)
 * Add to visited
 */
 static inline void
-AddToVisited(char *base, visited_hash * v, HnswElementPtr elementPtr, bool inMemory, bool *found)
+AddToVisited(char *base, visited_hash * v, HnswElementPtr elementPtr, Relation index, bool *found)
 {
-	if (!inMemory)
+	if (index != NULL)
 	{
 		HnswElement element = HnswPtrAccess(base, elementPtr);
 		ItemPointerData indextid;
@@ -704,57 +745,48 @@ HnswLoadUnvisitedFromMemory(char *base, HnswElement element, HnswUnvisited * unv
 		HnswCandidate *hc = &localNeighborhood->items[i];
 		bool		found;

-		AddToVisited(base, v, hc->element, true, &found);
+		AddToVisited(base, v, hc->element, NULL, &found);

 		if (!found)
 			unvisited[(*unvisitedLength)++].element = HnswPtrAccess(base, hc->element);
 	}
 }

-/*
- * Load neighbor index TIDs
- */
-bool
-HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation index, int m, int lm, int lc)
-{
-	Buffer		buf;
-	Page		page;
-	HnswNeighborTuple ntup;
-	int			start;
-
-	buf = ReadBuffer(index, element->neighborPage);
-	LockBuffer(buf, BUFFER_LOCK_SHARE);
-	page = BufferGetPage(buf);
-
-	ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
-
-	/* Ensure expected neighbors */
-	if (ntup->count != (element->level + 2) * m)
-	{
-		UnlockReleaseBuffer(buf);
-		return false;
-	}
-
-	/* Copy to minimize lock time */
-	start = (element->level - lc) * m;
-	memcpy(indextids, ntup->indextids + start, lm * sizeof(ItemPointerData));
-
-	UnlockReleaseBuffer(buf);
-	return true;
-}
-
 /*
 * Load unvisited neighbors from disk
 */
 static void
 HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *unvisitedLength, visited_hash * v, Relation index, int m, int lm, int lc)
 {
+	Buffer		buf;
+	Page		page;
+	HnswNeighborTuple ntup;
+	int			start;
 	ItemPointerData indextids[HNSW_MAX_M * 2];

 	*unvisitedLength = 0;

-	if (!HnswLoadNeighborTids(element, indextids, index, m, lm, lc))
+	buf = ReadBuffer(index, element->neighborPage);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+
+	ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
+
+	/*
+	 * Ensure the neighbor tuple has not been deleted or replaced between
+	 * index scan iterations
+	 */
+	if (ntup->version != element->version || ntup->count != (element->level + 2) * m)
+	{
+		UnlockReleaseBuffer(buf);
 		return;
+	}
+
+	/* Copy to minimize lock time */
+	start = (element->level - lc) * m;
+	memcpy(&indextids, ntup->indextids + start, lm * sizeof(ItemPointerData));
+
+	UnlockReleaseBuffer(buf);

 	for (int i = 0; i < lm; i++)
 	{
@@ -775,25 +807,36 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
 * Algorithm 2 from paper
 */
 List *
-HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement)
+HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
 {
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
 	pairingheap *W = pairingheap_allocate(CompareFurthestCandidates, NULL);
 	int			wlen = 0;
-	visited_hash v;
+	visited_hash vh;
 	ListCell   *lc2;
 	HnswNeighborArray *localNeighborhood = NULL;
 	Size		neighborhoodSize = 0;
 	int			lm = HnswGetLayerM(m, lc);
 	HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
 	int			unvisitedLength;
-	bool		inMemory = index == NULL;

-	InitVisited(base, &v, inMemory, ef, m);
+	if (v == NULL)
+	{
+		v = &vh;
+		initVisited = true;
+	}
+
+	if (initVisited)
+	{
+		InitVisited(base, v, index, ef, m);
+
+		if (discarded != NULL)
+			*discarded = pairingheap_allocate(CompareNearestDiscardedCandidates, NULL);
+	}

 	/* Create local memory for neighborhood if needed */
-	if (inMemory)
+	if (index == NULL)
 	{
 		neighborhoodSize = HNSW_NEIGHBOR_ARRAY_SIZE(lm);
 		localNeighborhood = palloc(neighborhoodSize);
@@ -802,20 +845,26 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 	/* Add entry points to v, C, and W */
 	foreach(lc2, ep)
 	{
-		HnswSearchCandidate *sc = (HnswSearchCandidate *) lfirst(lc2);
+		HnswSearchCandidate *hc = (HnswSearchCandidate *) lfirst(lc2);
 		bool		found;

-		AddToVisited(base, &v, sc->element, inMemory, &found);
+		if (initVisited)
+		{
+			AddToVisited(base, v, hc->element, index, &found);

-		pairingheap_add(C, &sc->c_node);
-		pairingheap_add(W, &sc->w_node);
+			if (tuples != NULL)
+				(*tuples)++;
+		}
+
+		pairingheap_add(C, &hc->c_node);
+		pairingheap_add(W, &hc->w_node);

 		/*
 		 * Do not count elements being deleted towards ef when vacuuming. It
 		 * would be ideal to do this for inserts as well, but this could
 		 * affect insert performance.
 		 */
-		if (CountElement(skipElement, HnswPtrAccess(base, sc->element)))
+		if (CountElement(skipElement, HnswPtrAccess(base, hc->element)))
 			wlen++;
 	}

@@ -830,24 +879,27 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in

 		cElement = HnswPtrAccess(base, c->element);

-		if (inMemory)
-			HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, &v, lc, localNeighborhood, neighborhoodSize);
+		if (index == NULL)
+			HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, v, lc, localNeighborhood, neighborhoodSize);
 		else
-			HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, &v, index, m, lm, lc);
+			HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, v, index, m, lm, lc);
+
+		if (tuples != NULL)
+			(*tuples) += unvisitedLength;

 		for (int i = 0; i < unvisitedLength; i++)
 		{
 			HnswElement eElement;
 			HnswSearchCandidate *e;
-			double		eDistance;
+			float		eDistance;
 			bool		alwaysAdd = wlen < ef;

 			f = HnswGetSearchCandidate(w_node, pairingheap_first(W));

-			if (inMemory)
+			if (index == NULL)
 			{
 				eElement = unvisited[i].element;
-				eDistance = GetElementDistance(base, eElement, q, support);
+				eDistance = GetElementDistance(base, eElement, q, procinfo, collation);
 			}
 			else
 			{
@@ -857,16 +909,22 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in

 				/* Avoid any allocations if not adding */
 				eElement = NULL;
-				HnswLoadElementImpl(blkno, offno, &eDistance, q, index, support, inserting, alwaysAdd ? NULL : &f->distance, &eElement);
-
-				if (eElement == NULL)
-					continue;
+				HnswLoadElementImpl(blkno, offno, &eDistance, &q, index, procinfo, collation, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);
 			}

-			if (!(eDistance < f->distance || alwaysAdd))
-				continue;
+			if (eElement == NULL || !(eDistance < f->distance || alwaysAdd))
+			{
+				if (discarded != NULL)
+				{
+					/* Create a new candidate */
+					e = palloc(sizeof(HnswSearchCandidate));
+					HnswPtrStore(base, e->element, eElement);
+					e->distance = eDistance;
+					pairingheap_add(*discarded, &e->w_node);
+				}

-			Assert(!eElement->deleted);
+				continue;
+			}

 			/* Make robust to issues */
 			if (eElement->level < lc)
@@ -890,7 +948,12 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in

 				/* No need to decrement wlen */
 				if (wlen > ef)
-					pairingheap_remove_first(W);
+				{
+					HnswSearchCandidate *d = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));
+
+					if (discarded != NULL)
+						pairingheap_add(*discarded, &d->w_node);
+				}
 			}
 		}
 	}
@@ -898,9 +961,9 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 	/* Add each element of W to w */
 	while (!pairingheap_is_empty(W))
 	{
-		HnswSearchCandidate *sc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));
+		HnswSearchCandidate *hc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));

-		w = lappend(w, sc);
+		w = lappend(w, hc);
 	}

 	return w;
@@ -954,22 +1017,32 @@ CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
 	return 0;
 }

+/*
+ * Calculate the distance between elements
+ */
+static float
+HnswGetDistance(char *base, HnswElement a, HnswElement b, FmgrInfo *procinfo, Oid collation)
+{
+	Datum		aValue = HnswGetValue(base, a);
+	Datum		bValue = HnswGetValue(base, b);
+
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, aValue, bValue));
+}
+
 /*
 * Check if an element is closer to q than any element from R
 */
 static bool
-CheckElementCloser(char *base, HnswCandidate * e, List *r, HnswSupport * support)
+CheckElementCloser(char *base, HnswCandidate * e, List *r, FmgrInfo *procinfo, Oid collation)
 {
 	HnswElement eElement = HnswPtrAccess(base, e->element);
-	Datum		eValue = HnswGetValue(base, eElement);
 	ListCell   *lc2;

 	foreach(lc2, r)
 	{
 		HnswCandidate *ri = lfirst(lc2);
 		HnswElement riElement = HnswPtrAccess(base, ri->element);
-		Datum		riValue = HnswGetValue(base, riElement);
-		float		distance = HnswGetDistance(eValue, riValue, support);
+		float		distance = HnswGetDistance(base, eElement, riElement, procinfo, collation);

 		if (distance <= e->distance)
 			return false;
@@ -982,14 +1055,15 @@ CheckElementCloser(char *base, HnswCandidate * e, List *r, HnswSupport * support
 * Algorithm 4 from paper
 */
 static List *
-SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
+SelectNeighbors(char *base, List *c, int lm, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
 	HnswCandidate **wd;
 	int			wdlen = 0;
 	int			wdoff = 0;
-	bool		mustCalculate = !(*closerSet);
+	HnswNeighborArray *neighbors = HnswGetNeighbors(base, e2, lc);
+	bool		mustCalculate = !neighbors->closerSet;
 	List	   *added = NIL;
 	bool		removedAny = false;

@@ -1016,7 +1090,7 @@ SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closer

 		/* Use previous state of r and wd to skip work when possible */
 		if (mustCalculate)
-			e->closer = CheckElementCloser(base, e, r, support);
+			e->closer = CheckElementCloser(base, e, r, procinfo, collation);
 		else if (list_length(added) > 0)
 		{
 			/* Keep Valgrind happy for in-memory, parallel builds */
@@ -1029,7 +1103,7 @@ SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closer
 			 */
 			if (e->closer)
 			{
-				e->closer = CheckElementCloser(base, e, added, support);
+				e->closer = CheckElementCloser(base, e, added, procinfo, collation);

 				if (!e->closer)
 					removedAny = true;
@@ -1042,7 +1116,7 @@ SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closer
 				 */
 				if (removedAny)
 				{
-					e->closer = CheckElementCloser(base, e, r, support);
+					e->closer = CheckElementCloser(base, e, r, procinfo, collation);
 					if (e->closer)
 						added = lappend(added, e);
 				}
@@ -1050,7 +1124,7 @@ SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closer
 		}
 		else if (e == newCandidate)
 		{
-			e->closer = CheckElementCloser(base, e, r, support);
+			e->closer = CheckElementCloser(base, e, r, procinfo, collation);
 			if (e->closer)
 				added = lappend(added, e);
 		}
@@ -1066,7 +1140,7 @@ SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closer
 	}

 	/* Cached value can only be used in future if sorted deterministically */
-	*closerSet = sortCandidates;
+	neighbors->closerSet = sortCandidates;

 	/* Keep pruned connections */
 	while (wdoff < wdlen && list_length(r) < lm)
@@ -1101,16 +1175,18 @@ AddConnections(char *base, HnswElement element, List *neighbors, int lc)
 * Update connections
 */
 void
-HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support)
+HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int lm, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation)
 {
-	HnswCandidate newHc;
+	HnswElement hce = HnswPtrAccess(base, hc->element);
+	HnswNeighborArray *currentNeighbors = HnswGetNeighbors(base, hce, lc);
+	HnswCandidate hc2;

-	HnswPtrStore(base, newHc.element, newElement);
-	newHc.distance = distance;
+	HnswPtrStore(base, hc2.element, element);
+	hc2.distance = hc->distance;

-	if (neighbors->length < lm)
+	if (currentNeighbors->length < lm)
 	{
-		neighbors->items[neighbors->length++] = newHc;
+		currentNeighbors->items[currentNeighbors->length++] = hc2;

 		/* Track update */
 		if (updateIdx != NULL)
@@ -1119,26 +1195,54 @@ HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newE
 	else
 	{
 		/* Shrink connections */
-		List	   *c = NIL;
 		HnswCandidate *pruned = NULL;

-		/* Add candidates */
-		for (int i = 0; i < neighbors->length; i++)
-			c = lappend(c, &neighbors->items[i]);
-		c = lappend(c, &newHc);
+		/* Load elements on insert */
+		if (index != NULL)
+		{
+			Datum		q = HnswGetValue(base, hce);

-		SelectNeighbors(base, c, lm, support, &neighbors->closerSet, &newHc, &pruned, true);
+			for (int i = 0; i < currentNeighbors->length; i++)
+			{
+				HnswCandidate *hc3 = &currentNeighbors->items[i];
+				HnswElement hc3Element = HnswPtrAccess(base, hc3->element);
+
+				if (HnswPtrIsNull(base, hc3Element->value))
+					HnswLoadElement(hc3Element, &hc3->distance, &q, index, procinfo, collation, true, NULL);
+				else
+					hc3->distance = GetElementDistance(base, hc3Element, q, procinfo, collation);
+
+				/* Prune element if being deleted */
+				if (hc3Element->heaptidsLength == 0)
+				{
+					pruned = &currentNeighbors->items[i];
+					break;
+				}
+			}
+		}

-		/* Should not happen */
 		if (pruned == NULL)
-			return;
+		{
+			List	   *c = NIL;
+
+			/* Add candidates */
+			for (int i = 0; i < currentNeighbors->length; i++)
+				c = lappend(c, &currentNeighbors->items[i]);
+			c = lappend(c, &hc2);
+
+			SelectNeighbors(base, c, lm, lc, procinfo, collation, hce, &hc2, &pruned, true);
+
+			/* Should not happen */
+			if (pruned == NULL)
+				return;
+		}

 		/* Find and replace the pruned element */
-		for (int i = 0; i < neighbors->length; i++)
+		for (int i = 0; i < currentNeighbors->length; i++)
 		{
-			if (HnswPtrEqual(base, neighbors->items[i].element, pruned->element))
+			if (HnswPtrEqual(base, currentNeighbors->items[i].element, pruned->element))
 			{
-				neighbors->items[i] = newHc;
+				currentNeighbors->items[i] = hc2;

 				/* Track update */
 				if (updateIdx != NULL)
@@ -1198,20 +1302,17 @@ PrecomputeHash(char *base, HnswElement element)
 * Algorithm 1 from paper
 */
 void
-HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing)
+HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing)
 {
 	List	   *ep;
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	HnswQuery	q;
+	Datum		q = HnswGetValue(base, element);
 	HnswElement skipElement = existing ? element : NULL;
-	bool		inMemory = index == NULL;
-
-	q.value = HnswGetValue(base, element);

 	/* Precompute hash */
-	if (inMemory)
+	if (index == NULL)
 		PrecomputeHash(base, element);

 	/* No neighbors if no entry point */
@@ -1219,13 +1320,13 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		return;

 	/* Get entry point and level */
-	ep = list_make1(HnswEntryCandidate(base, entryPoint, &q, index, support, true));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, procinfo, collation, true));
 	entryLevel = entryPoint->level;

 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, true, skipElement, NULL, NULL, true, NULL);
 		ep = w;
 	}

@@ -1244,7 +1345,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		List	   *lw = NIL;
 		ListCell   *lc2;

-		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement, NULL, NULL, true, NULL);

 		/* Convert search candidates to candidates */
 		foreach(lc2, w)
@@ -1260,7 +1361,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint

 		/* Elements being deleted or skipped can help with search */
 		/* but should be removed before selecting neighbors */
-		if (!inMemory)
+		if (index != NULL)
 			lw = RemoveElements(base, lw, skipElement);

 		/*
@@ -1268,7 +1369,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		 * sortCandidates to true for in-memory builds to enable closer
 		 * caching, but there does not seem to be a difference in performance.
 		 */
-		neighbors = SelectNeighbors(base, lw, lm, support, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);
+		neighbors = SelectNeighbors(base, lw, lm, lc, procinfo, collation, element, NULL, NULL, false);

 		AddConnections(base, element, neighbors, lc);

--- a/src/hnswvacuum.c
+++ b/src/hnswvacuum.c
@@ -184,12 +184,13 @@ static void
 RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
 {
 	Relation	index = vacuumstate->index;
-	HnswSupport *support = &vacuumstate->support;
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
 	int			m = vacuumstate->m;
 	int			efConstruction = vacuumstate->efConstruction;
+	FmgrInfo   *procinfo = vacuumstate->procinfo;
+	Oid			collation = vacuumstate->collation;
 	BufferAccessStrategy bas = vacuumstate->bas;
 	HnswNeighborTuple ntup = vacuumstate->ntup;
 	Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
@@ -204,7 +205,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	element->heaptidsLength = 0;

 	/* Find neighbors for element, skipping itself */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, true);
+	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, true);

 	/* Zero memory for each element */
 	MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
@@ -228,7 +229,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	UnlockReleaseBuffer(buf);

 	/* Update neighbors */
-	HnswUpdateNeighborsOnDisk(index, support, element, m, true, false);
+	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, true, false);
 }

 /*
@@ -238,7 +239,6 @@ static void
 RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 {
 	Relation	index = vacuumstate->index;
-	HnswSupport *support = &vacuumstate->support;
 	HnswElement highestPoint = &vacuumstate->highestPoint;
 	HnswElement entryPoint;
 	MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
@@ -256,7 +256,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);

 		/* Load element */
-		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
+		HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL);

 		/* Repair if needed */
 		if (NeedsUpdated(vacuumstate, highestPoint))
@@ -294,7 +294,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 			 * is outdated, this can remove connections at higher levels in
 			 * the graph until they are repaired, but this should be fine.
 			 */
-			HnswLoadElement(entryPoint, NULL, NULL, index, support, true, NULL);
+			HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL);

 			if (NeedsUpdated(vacuumstate, entryPoint))
 			{
@@ -527,6 +527,14 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 			for (int i = 0; i < ntup->count; i++)
 				ItemPointerSetInvalid(&ntup->indextids[i]);

+			/* Increment version */
+			/* This is used to avoid incorrect reads for iterative scans */
+			/* Reserve some bits for future use */
+			etup->version++;
+			if (etup->version > 15)
+				etup->version = 1;
+			ntup->version = etup->version;
+
 			/*
 			 * We modified the tuples in place, no need to call
 			 * PageIndexTupleOverwrite
@@ -573,13 +581,13 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	vacuumstate->callback_state = callback_state;
 	vacuumstate->efConstruction = HnswGetEfConstruction(index);
 	vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
+	vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	vacuumstate->collation = index->rd_indcollation[0];
 	vacuumstate->ntup = palloc0(HNSW_TUPLE_ALLOC_SIZE);
 	vacuumstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 												"Hnsw vacuum temporary context",
 												ALLOCSET_DEFAULT_SIZES);

-	HnswInitSupport(&vacuumstate->support, index);
-
 	/* Get m from metapage */
 	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);

--- a/src/ivfbuild.c
+++ b/src/ivfbuild.c
@@ -138,7 +138,7 @@ SampleRows(IvfflatBuildState * buildstate)
 * Add tuple to sort
 */
 static void
-AddTupleToSort(Relation index, ItemPointer tid, Datum *values, bool *isnull, IvfflatBuildState * buildstate)
+AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState * buildstate)
 {
 	double		distance;
 	double		minDistance = DBL_MAX;
@@ -184,11 +184,6 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, bool *isnull, Ivf
 	slot->tts_isnull[1] = false;
 	slot->tts_values[2] = value;
 	slot->tts_isnull[2] = false;
-	for (int i = 1; i < buildstate->tupdesc->natts; i++)
-	{
-		slot->tts_values[2 + i] = values[i];
-		slot->tts_isnull[2 + i] = isnull[i];
-	}
 	ExecStoreVirtualTuple(slot);

 	/*
@@ -220,7 +215,7 @@ BuildCallback(Relation index, ItemPointer tid, Datum *values,
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);

 	/* Add tuple to sort */
-	AddTupleToSort(index, tid, values, isnull, buildstate);
+	AddTupleToSort(index, tid, values, buildstate);

 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -231,20 +226,19 @@ BuildCallback(Relation index, ItemPointer tid, Datum *values,
 * Get index tuple from sort state
 */
 static inline void
-GetNextTuple(Tuplesortstate *sortstate, TupleDesc tupdesc, TupleTableSlot *slot, Datum *values, bool *isnull, IndexTuple *itup, int *list)
+GetNextTuple(Tuplesortstate *sortstate, TupleDesc tupdesc, TupleTableSlot *slot, IndexTuple *itup, int *list)
 {
+	Datum		value;
+	bool		isnull;
+
 	if (tuplesort_gettupleslot(sortstate, true, false, slot, NULL))
 	{
-		bool		unused;
-
-		*list = DatumGetInt32(slot_getattr(slot, 1, &unused));
-
-		for (int i = 0; i < tupdesc->natts; i++)
-			values[i] = slot_getattr(slot, 3 + i, &isnull[i]);
+		*list = DatumGetInt32(slot_getattr(slot, 1, &isnull));
+		value = slot_getattr(slot, 3, &isnull);

 		/* Form the index tuple */
-		*itup = index_form_tuple(tupdesc, values, isnull);
-		(*itup)->t_tid = *((ItemPointer) DatumGetPointer(slot_getattr(slot, 2, &unused)));
+		*itup = index_form_tuple(tupdesc, &value, &isnull);
+		(*itup)->t_tid = *((ItemPointer) DatumGetPointer(slot_getattr(slot, 2, &isnull)));
 	}
 	else
 		*list = -1;
@@ -260,16 +254,14 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
 	IndexTuple	itup = NULL;	/* silence compiler warning */
 	int64		inserted = 0;

-	TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsMinimalTuple);
-	TupleDesc	tupdesc = buildstate->tupdesc;
-	Datum	   *values = palloc(tupdesc->natts * sizeof(Datum));
-	bool	   *isnull = palloc(tupdesc->natts * sizeof(bool));
+	TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsMinimalTuple);
+	TupleDesc	tupdesc = RelationGetDescr(index);

 	pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_LOAD);

 	pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_TOTAL, buildstate->indtuples);

-	GetNextTuple(buildstate->sortstate, tupdesc, slot, values, isnull, &itup, &list);
+	GetNextTuple(buildstate->sortstate, tupdesc, slot, &itup, &list);

 	for (int i = 0; i < buildstate->centers->length; i++)
 	{
@@ -305,7 +297,7 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)

 			pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE, ++inserted);

-			GetNextTuple(buildstate->sortstate, tupdesc, slot, values, isnull, &itup, &list);
+			GetNextTuple(buildstate->sortstate, tupdesc, slot, &itup, &list);
 		}

 		insertPage = BufferGetBlockNumber(buf);
@@ -315,9 +307,6 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
 		/* Set the start and insert pages */
 		IvfflatUpdateList(index, buildstate->listInfo[i], insertPage, InvalidBlockNumber, startPage, forkNum);
 	}
-
-	pfree(values);
-	pfree(isnull);
 }

 /*
@@ -330,7 +319,6 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->index = index;
 	buildstate->indexInfo = indexInfo;
 	buildstate->typeInfo = IvfflatGetTypeInfo(index);
-	buildstate->tupdesc = RelationGetDescr(index);

 	buildstate->lists = IvfflatGetLists(index);
 	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
@@ -341,19 +329,6 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 				 errmsg("type not supported for ivfflat index")));

-	/* TODO See if needed */
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 3)
-		elog(ERROR, "index cannot have more than three columns");
-
-	if (!OidIsValid(index_getprocid(index, 1, IVFFLAT_DISTANCE_PROC)))
-		elog(ERROR, "first column must be a vector");
-
-	for (int i = 1; i < IndexRelationGetNumberOfKeyAttributes(index); i++)
-	{
-		if (OidIsValid(index_getprocid(index, i + 1, IVFFLAT_DISTANCE_PROC)))
-			elog(ERROR, "column %d cannot be a vector", i + 1);
-	}
-
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)
 		ereport(ERROR,
@@ -381,13 +356,12 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 				 errmsg("dimensions must be greater than one for this opclass")));

 	/* Create tuple description for sorting */
-	buildstate->sortdesc = CreateTemplateTupleDesc(2 + buildstate->tupdesc->natts);
-	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
-	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
-	for (int i = 0; i < buildstate->tupdesc->natts; i++)
-		TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) (3 + i), NULL, buildstate->tupdesc->attrs[i].atttypid, -1, 0);
+	buildstate->tupdesc = CreateTemplateTupleDesc(3);
+	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
+	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
+	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 3, "vector", RelationGetDescr(index)->attrs[0].atttypid, -1, 0);

-	buildstate->slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsVirtual);
+	buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsVirtual);

 	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions, buildstate->typeInfo->itemSize(buildstate->dimensions));
 	buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
@@ -659,7 +633,7 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
 	InitBuildState(&buildstate, ivfspool->heap, ivfspool->index, indexInfo);
 	memcpy(buildstate.centers->items, ivfcenters, buildstate.centers->itemsize * buildstate.centers->maxlen);
 	buildstate.centers->length = buildstate.centers->maxlen;
-	ivfspool->sortstate = InitBuildSortState(buildstate.sortdesc, sortmem, coordinate);
+	ivfspool->sortstate = InitBuildSortState(buildstate.tupdesc, sortmem, coordinate);
 	buildstate.sortstate = ivfspool->sortstate;
 	scan = table_beginscan_parallel(ivfspool->heap,
 									ParallelTableScanFromIvfflatShared(ivfshared));
@@ -976,7 +950,7 @@ AssignTuples(IvfflatBuildState * buildstate)
 	}

 	/* Begin serial/leader tuplesort */
-	buildstate->sortstate = InitBuildSortState(buildstate->sortdesc, maintenance_work_mem, coordinate);
+	buildstate->sortstate = InitBuildSortState(buildstate->tupdesc, maintenance_work_mem, coordinate);

 	/* Add tuples to sort */
 	if (buildstate->heap != NULL)
--- a/src/ivfflat.c
+++ b/src/ivfflat.c
@@ -69,8 +69,6 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	GenericCosts costs;
 	int			lists;
 	double		ratio;
-	double		sequentialRatio = 0.5;
-	double		startupPages;
 	double		spc_seq_page_cost;
 	Relation	index;

@@ -87,8 +85,6 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,

 	MemSet(&costs, 0, sizeof(costs));

-	genericcostestimate(root, path, loop_count, &costs);
-
 	index = index_open(path->indexinfo->indexoid, NoLock);
 	IvfflatGetMetaPageInfo(index, &lists, NULL);
 	index_close(index, NoLock);
@@ -98,26 +94,34 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	if (ratio > 1.0)
 		ratio = 1.0;

+	/*
+	 * This gives us the subset of tuples to visit. This value is passed into
+	 * the generic cost estimator to determine the number of pages to visit
+	 * during the index scan.
+	 */
+	costs.numIndexTuples = path->indexinfo->tuples * ratio;
+
+	genericcostestimate(root, path, loop_count, &costs);
+
 	get_tablespace_page_costs(path->indexinfo->reltablespace, NULL, &spc_seq_page_cost);

-	/* Change some page cost from random to sequential */
-	costs.indexTotalCost -= sequentialRatio * costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);
-
-	/* Startup cost is cost before returning the first row */
-	costs.indexStartupCost = costs.indexTotalCost * ratio;
-
 	/* Adjust cost if needed since TOAST not included in seq scan cost */
-	startupPages = costs.numIndexPages * ratio;
-	if (startupPages > path->indexinfo->rel->pages && ratio < 0.5)
+	if (costs.numIndexPages > path->indexinfo->rel->pages && ratio < 0.5)
 	{
-		/* Change rest of page cost from random to sequential */
-		costs.indexStartupCost -= (1 - sequentialRatio) * startupPages * (costs.spc_random_page_cost - spc_seq_page_cost);
+		/* Change all page cost from random to sequential */
+		costs.indexTotalCost -= costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);

 		/* Remove cost of extra pages */
-		costs.indexStartupCost -= (startupPages - path->indexinfo->rel->pages) * spc_seq_page_cost;
+		costs.indexTotalCost -= (costs.numIndexPages - path->indexinfo->rel->pages) * spc_seq_page_cost;
+	}
+	else
+	{
+		/* Change some page cost from random to sequential */
+		costs.indexTotalCost -= 0.5 * costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);
 	}

-	*indexStartupCost = costs.indexStartupCost;
+	/* Use total cost since most work happens before first tuple is returned */
+	*indexStartupCost = costs.indexTotalCost;
 	*indexTotalCost = costs.indexTotalCost;
 	*indexSelectivity = costs.indexSelectivity;
 	*indexCorrelation = costs.indexCorrelation;
@@ -167,7 +171,7 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	amroutine->amcanorderbyop = true;
 	amroutine->amcanbackward = false;	/* can change direction mid-scan */
 	amroutine->amcanunique = false;
-	amroutine->amcanmulticol = true;
+	amroutine->amcanmulticol = false;
 	amroutine->amoptionalkey = true;
 	amroutine->amsearcharray = false;
 	amroutine->amsearchnulls = false;
--- a/src/ivfflat.h
+++ b/src/ivfflat.h
@@ -165,7 +165,6 @@ typedef struct IvfflatBuildState
 	Relation	index;
 	IndexInfo  *indexInfo;
 	const		IvfflatTypeInfo *typeInfo;
-	TupleDesc	tupdesc;

 	/* Settings */
 	int			dimensions;
@@ -199,7 +198,7 @@ typedef struct IvfflatBuildState

 	/* Sorting */
 	Tuplesortstate *sortstate;
-	TupleDesc	sortdesc;
+	TupleDesc	tupdesc;
 	TupleTableSlot *slot;

 	/* Memory */
--- a/src/ivfinsert.c
+++ b/src/ivfinsert.c
@@ -78,8 +78,6 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	BlockNumber insertPage = InvalidBlockNumber;
 	ListInfo	listInfo;
 	BlockNumber originalInsertPage;
-	TupleDesc	tupdesc = RelationGetDescr(index);
-	Datum	   *newValues = palloc(tupdesc->natts * sizeof(Datum));

 	/* Detoast once for all calls */
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -100,16 +98,12 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	IvfflatGetMetaPageInfo(index, NULL, NULL);

 	/* Find the insert page - sets the page and list info */
-	FindInsertPage(index, &value, &insertPage, &listInfo);
+	FindInsertPage(index, values, &insertPage, &listInfo);
 	Assert(BlockNumberIsValid(insertPage));
 	originalInsertPage = insertPage;

-	newValues[0] = value;
-	for (int i = 1; i < tupdesc->natts; i++)
-		newValues[i] = values[i];
-
 	/* Form tuple */
-	itup = index_form_tuple(tupdesc, newValues, isnull);
+	itup = index_form_tuple(RelationGetDescr(index), &value, isnull);
 	itup->t_tid = *heap_tid;

 	/* Get tuple size */
--- a/src/ivfscan.c
+++ b/src/ivfscan.c
@@ -104,34 +104,6 @@ GetScanLists(IndexScanDesc scan, Datum value)
 	}
 }

-/*
- * Check if matches scan keys
- */
-static bool
-MatchesScanKeys(IndexScanDesc scan, IndexTuple itup, TupleDesc tupdesc)
-{
-	for (int i = 0; i < scan->numberOfKeys; i++)
-	{
-		ScanKey		key = &scan->keyData[i];
-		bool		attnull = key->sk_flags & SK_ISNULL;
-		bool		isnull;
-		Datum		value = index_getattr(itup, key->sk_attno, tupdesc, &isnull);
-
-		if (isnull || attnull)
-		{
-			if (isnull != attnull)
-				return false;
-		}
-		else
-		{
-			if (!DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation, value, key->sk_argument)))
-				return false;
-		}
-	}
-
-	return true;
-}
-
 /*
 * Get items
 */
@@ -168,10 +140,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				ItemId		itemid = PageGetItemId(page, offno);

 				itup = (IndexTuple) PageGetItem(page, itemid);
-
-				if (!MatchesScanKeys(scan, itup, tupdesc))
-					continue;
-
 				datum = index_getattr(itup, 1, tupdesc, &isnull);

 				/*
--- a/test/t/002_ivfflat_vacuum.pl
+++ b/test/t/002_ivfflat_vacuum.pl
@@ -6,7 +6,13 @@ use Test::More;

 my $dim = 3;

-my $array_sql = join(",", ('random()') x $dim);
+my @r = ();
+for (1 .. $dim)
+{
+	my $v = int(rand(1000)) + 1;
+	push(@r, "i % $v");
+}
+my $array_sql = join(", ", @r);

 # Initialize node
 my $node = PostgreSQL::Test::Cluster->new('node');
@@ -17,20 +23,19 @@ $node->start;
 $node->safe_psql("postgres", "CREATE EXTENSION vector;");
 $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
 $node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");

 # Get size
 my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");

-# Store values
-$node->safe_psql("postgres", "CREATE TABLE tmp AS SELECT * FROM tst;");
-
 # Delete all, vacuum, and insert same data
 $node->safe_psql("postgres", "DELETE FROM tst;");
 $node->safe_psql("postgres", "VACUUM tst;");
-$node->safe_psql("postgres", "INSERT INTO tst SELECT * FROM tmp;");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);

 # Check size
 my $new_size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
--- a/test/t/017_hnsw_filtering.pl
+++ b/test/t/017_hnsw_filtering.pl
@@ -41,7 +41,8 @@ my $c = int(rand() * $nc);
 my $explain = $node->safe_psql("postgres", qq(
 	EXPLAIN ANALYZE SELECT i FROM tst WHERE c = $c ORDER BY v <-> '$query' LIMIT $limit;
 ));
-like($explain, qr/Seq Scan/);
+# TODO Do not use index
+like($explain, qr/Index Scan using idx/);

 # Test attribute filtering with few rows removed
 $explain = $node->safe_psql("postgres", qq(
@@ -59,7 +60,8 @@ like($explain, qr/Index Scan using idx/);
 $explain = $node->safe_psql("postgres", qq(
 	EXPLAIN ANALYZE SELECT i FROM tst WHERE c < 1 ORDER BY v <-> '$query' LIMIT $limit;
 ));
-like($explain, qr/Seq Scan/);
+# TODO Do not use index
+like($explain, qr/Index Scan using idx/);

 # Test attribute filtering with few rows removed like
 $explain = $node->safe_psql("postgres", qq(
--- a/test/t/039_hnsw_cost.pl
+++ b/test/t/039_hnsw_cost.pl
@@ -17,11 +17,12 @@ $node->safe_psql("postgres", "CREATE EXTENSION vector;");
 for my $dim (@dims)
 {
 	my $array_sql = join(",", ('random()') x $dim);
+	my $n = $dim == 384 ? 2000 : 1000;

 	# Create table and index
 	$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 	$node->safe_psql("postgres",
-		"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 2000) i;"
+		"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, $n) i;"
 	);
 	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v vector_l2_ops);");
 	$node->safe_psql("postgres", "ANALYZE tst;");
@@ -39,21 +40,6 @@ for my $dim (@dims)
 	));
 	like($explain, qr/Index Scan using idx/);

-	# 3x the rows are needed for distance filters
-	# since the planner uses DEFAULT_INEQ_SEL for the selectivity (should be 1)
-	# Recreate index for performance
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-	$node->safe_psql("postgres",
-		"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(2001, 6000) i;"
-	);
-	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v vector_l2_ops);");
-	$node->safe_psql("postgres", "ANALYZE tst;");
-
-	$explain = $node->safe_psql("postgres", qq(
-		EXPLAIN ANALYZE SELECT i FROM tst WHERE v <-> '$query' < 1 ORDER BY v <-> '$query' LIMIT $limit;
-	));
-	like($explain, qr/Index Scan using idx/);
-
 	$node->safe_psql("postgres", "DROP TABLE tst;");
 }

--- a/test/t/040_ivfflat_cost.pl
+++ b/test/t/040_ivfflat_cost.pl
@@ -39,11 +39,6 @@ for my $dim (@dims)
 	));
 	like($explain, qr/Index Scan using idx/);

-	$explain = $node->safe_psql("postgres", qq(
-		EXPLAIN ANALYZE SELECT i FROM tst WHERE v <-> '$query' < 1 ORDER BY v <-> '$query' LIMIT $limit;
-	));
-	like($explain, qr/Index Scan using idx/);
-
 	$node->safe_psql("postgres", "DROP TABLE tst;");
 }

--- a/test/t/041_hnsw_streaming.pl
+++ b/test/t/041_hnsw_streaming.pl
@@ -0,0 +1,66 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4 PRIMARY KEY, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+$node->safe_psql("postgres", qq(
+	SET maintenance_work_mem = '128MB';
+	SET max_parallel_maintenance_workers = 2;
+	CREATE INDEX ON tst USING hnsw (v vector_l2_ops)
+));
+
+my $count = $node->safe_psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.streaming = on;
+	SET work_mem = '8MB';
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+is($count, 10);
+
+foreach ((30000, 50000, 70000))
+{
+	my $ef_stream = $_;
+	my $expected = $ef_stream / 10000;
+	my $sum = 0;
+
+	for my $i (1 .. 20)
+	{
+		$count = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET hnsw.streaming = on;
+			SET hnsw.ef_stream = $ef_stream;
+			SET work_mem = '8MB';
+			SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst WHERE i = $i) LIMIT 11) t;
+		));
+		$sum += $count;
+	}
+
+	my $avg = $sum / 20;
+	cmp_ok($avg, '>', $expected - 2);
+	cmp_ok($avg, '<', $expected + 2);
+}
+
+my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.streaming = on;
+	SET work_mem = '2MB';
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+like($stderr, qr/hnsw index scan exceeded work_mem after \d+ tuples/);
+
+done_testing();
--- a/test/t/041_ivfflat_inline_filtering.pl
+++ b/test/t/041_ivfflat_inline_filtering.pl
@@ -1,197 +0,0 @@
-use strict;
-use warnings FATAL => 'all';
-use PostgreSQL::Test::Cluster;
-use PostgreSQL::Test::Utils;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @where = ();
-my @expected;
-my $limit = 20;
-my $dim = 3;
-my $array_sql = join(",", ('random()') x $dim);
-my $nc = 100;
-my $nc2 = 10;
-
-sub test_recall
-{
-	my ($probes, $min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	for my $j (0 .. 2)
-	{
-		my $explain = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SET ivfflat.probes = $probes;
-			EXPLAIN ANALYZE SELECT i FROM tst WHERE $where[$j] ORDER BY v $operator '$queries[$j]' LIMIT $limit;
-		));
-		like($explain, qr/Index Cond/);
-	}
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SET ivfflat.probes = $probes;
-			SELECT i FROM tst WHERE $where[$i] ORDER BY v $operator '$queries[$i]' LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-		my %actual_set = map { $_ => 1 } @actual_ids;
-
-		is(scalar(@actual_ids), $limit);
-
-		my @expected_ids = split("\n", $expected[$i]);
-
-		foreach (@expected_ids)
-		{
-			if (exists($actual_set{$_}))
-			{
-				$correct++;
-			}
-			$total++;
-		}
-	}
-
-	cmp_ok($correct / $total, ">=", $min, $operator);
-}
-
-# Initialize node
-$node = PostgreSQL::Test::Cluster->new('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim), c int4, c2 int4);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql], i % $nc, i % $nc2 FROM generate_series(1, 50000) i;"
-);
-
-# Generate queries
-for my $i (1 .. 100)
-{
-	my @r = ();
-	for (1 .. $dim)
-	{
-		push(@r, rand());
-	}
-	push(@queries, "[" . join(",", @r) . "]");
-
-	if ($i % 3 == 0)
-	{
-		my $c = int(rand() * $nc);
-		push(@where, "c = $c");
-	}
-	elsif ($i % 3 == 1)
-	{
-		my $c2 = int(rand() * $nc2);
-		push(@where, "c2 = $c2");
-	}
-	else
-	{
-		# use c2 to ensure results
-		my $c2 = int(rand() * $nc2);
-		push(@where, "c = $c2 AND c2 = $c2");
-	}
-}
-
-# Add index
-$node->safe_psql("postgres", qq(
-	CREATE INDEX ON tst USING ivfflat (v vector_l2_ops, c, c2) WITH (lists = 100);
-));
-
-# Insert more rows
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql], i % $nc, i % $nc2 FROM generate_series(1, 50000) i;"
-);
-
-# Get exact results
-@expected = ();
-for my $i (0 .. $#queries)
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_indexscan = off;
-		SELECT i FROM tst WHERE $where[$i] ORDER BY v <-> '$queries[$i]' LIMIT $limit;
-	));
-	push(@expected, $res);
-}
-
-# Test recall
-test_recall(10, 0.99, '<->');
-
-# Test vacuum
-$node->safe_psql("postgres", "DELETE FROM tst WHERE c > 5;");
-$node->safe_psql("postgres", "VACUUM tst;");
-
-# Test less than
-my $explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c < 10 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(c < 10\)/);
-
-# Test less than or equal
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c <= 10 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(c <= 10\)/);
-
-# Test greater than or equal
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c >= 90 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(c >= 90\)/);
-
-# Test greater than
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c > 90 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(c > 90\)/);
-
-# Test multiple attribute columns
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c = 1 AND c2 = 1 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(\(c = 1\) AND \(c2 = 1\)\)/);
-
-# Test only last attribute column
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c2 = 1 ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Cond: \(c2 = 1\)/);
-
-# Test only vector column
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$queries[0]' LIMIT $limit;
-));
-like($explain, qr/Index Scan/);
-
-# Test only attribute columns
-$explain = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	EXPLAIN ANALYZE SELECT i FROM tst WHERE c = 1;
-));
-like($explain, qr/Seq Scan/);
-
-# Test columns
-my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (c);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (c, v vector_cosine_ops);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_cosine_ops, c, c, c);");
-like($stderr, qr/index cannot have more than three columns/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_cosine_ops, v vector_cosine_ops);");
-like($stderr, qr/column 2 cannot be a vector/);
-
-done_testing();
--- a/test/t/042_hnsw_streaming_recall.pl
+++ b/test/t/042_hnsw_streaming_recall.pl
@@ -0,0 +1,131 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+my @cs = (100, 1000);
+
+sub test_recall
+{
+	my ($c, $ef_search, $min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET hnsw.ef_search = $ef_search;
+		SET hnsw.streaming = on;
+		EXPLAIN ANALYZE SELECT i FROM tst WHERE i % $c = 0 ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Scan using idx on tst/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET hnsw.ef_search = $ef_search;
+			SET hnsw.streaming = on;
+			SELECT i FROM tst WHERE i % $c = 0 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 = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my @r = ();
+	for (1 .. $dim)
+	{
+		push(@r, rand());
+	}
+	push(@queries, "[" . join(",", @r) . "]");
+}
+
+# Check each index type
+my @operators = ("<->", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_cosine_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	$node->safe_psql("postgres", qq(
+		SET maintenance_work_mem = '128MB';
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
+
+	foreach (@cs)
+	{
+		my $c = $_;
+
+		# Get exact results
+		@expected = ();
+		foreach (@queries)
+		{
+			my $res = $node->safe_psql("postgres", qq(
+				SET enable_indexscan = off;
+				WITH top AS (
+					SELECT v $operator '$_' AS distance FROM tst WHERE i % $c = 0 ORDER BY distance LIMIT $limit
+				)
+				SELECT i FROM tst WHERE (v $operator '$_') <= (SELECT MAX(distance) FROM top)
+			));
+			push(@expected, $res);
+		}
+
+		if ($c == 100)
+		{
+			test_recall($c, 40, 0.99, $operator);
+		}
+		else
+		{
+			if ($operator eq "<->")
+			{
+				test_recall($c, 40, 0.99, $operator);
+			}
+			else
+			{
+				test_recall($c, 40, 0.99, $operator);
+			}
+		}
+	}
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();
Author	SHA1	Message	Date
Andrew Kane	c01e76f2fa	Added strict ordering [skip ci]	2024-09-28 11:31:03 -07:00
Andrew Kane	ab57217f48	Added todo [skip ci]	2024-09-28 10:07:09 -07:00
Andrew Kane	5f6e031ccc	Updated changelog [skip ci]	2024-09-28 09:57:05 -07:00
Andrew Kane	1a1221f905	Merge branch 'master' into hnsw-streaming	2024-09-26 08:34:07 -07:00
Andrew Kane	40c3e402c7	Removed todo [skip ci]	2024-09-25 17:29:20 -07:00
Andrew Kane	058248fdcc	Improved cost code [skip ci]	2024-09-25 17:23:29 -07:00
Andrew Kane	73c5145b77	Use int for ef [skip ci]	2024-09-25 16:52:41 -07:00
Andrew Kane	ec4a23fe49	Added cost estimation [skip ci]	2024-09-25 16:45:04 -07:00
Andrew Kane	38207f5640	Merge branch 'master' into hnsw-streaming	2024-09-25 16:09:09 -07:00
Andrew Kane	4e35c6abe3	Updated readme [skip ci]	2024-09-24 23:24:48 -07:00
Andrew Kane	11e4d040d9	Fixed test [skip ci]	2024-09-24 19:38:06 -07:00
Andrew Kane	b2fa625255	Fixed crash with empty index [skip ci]	2024-09-23 09:42:31 -07:00
Andrew Kane	a8e699c927	Improved message [skip ci]	2024-09-22 22:31:48 -07:00
Andrew Kane	91541fece6	Fixed example [skip ci]	2024-09-22 22:28:39 -07:00
Andrew Kane	f3de487da2	Started readme updates [skip ci]	2024-09-22 22:26:27 -07:00
Andrew Kane	721d4b7e3f	Improved test for ef_stream [skip ci]	2024-09-22 18:51:38 -07:00
Andrew Kane	28066d8fe4	Added test for ef_stream [skip ci]	2024-09-22 18:35:35 -07:00
Andrew Kane	495041e43b	Added option to limit tuples [skip ci]	2024-09-22 18:10:19 -07:00
Andrew Kane	52c385c03a	Only pass discarded when streaming [skip ci]	2024-09-22 17:47:10 -07:00
Andrew Kane	80cbd32dab	Added streaming option for HNSW	2024-09-22 12:02:48 -07:00