Added support for inline filtering with HNSW [skip ci]

.github/workflows/build.yml

@@ -23,14 +23,14 @@ jobs:
           - postgres: 13
             os: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
           dev-files: true
       - run: make
         env:
-          PG_CFLAGS: ${{ matrix.postgres >= 18 && '-Wno-missing-field-initializers' || '' }} -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
+          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare ${{ matrix.postgres >= 18 && '-Wno-missing-field-initializers' || '' }}
       - run: |
           export PG_CONFIG=`which pg_config`
           sudo --preserve-env=PG_CONFIG make install
@@ -48,18 +48,18 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - postgres: 17
-            os: macos-15
+          - postgres: 18
+            os: macos-26
           - postgres: 14
-            os: macos-13
+            os: macos-15-intel
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
       - run: make
         env:
-          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-unknown-warning-option
+          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-unknown-warning-option ${{ matrix.postgres >= 18 && '-Wno-missing-field-initializers' || '' }}
       - run: make install
       - run: make installcheck
       - if: ${{ failure() }}
@@ -72,27 +72,35 @@ jobs:
           tar xf $TAG.tar.gz
           mv postgres-$TAG postgres
         env:
-          TAG: ${{ matrix.postgres == 17 && 'REL_17_2' || 'REL_14_15' }}
+          TAG: ${{ matrix.postgres == 18 && 'REL_18_0' || 'REL_14_19' }}
       - run: make prove_installcheck PROVE_FLAGS="-I ./postgres/src/test/perl -I ./test/perl"
         env:
           PERL5LIB: /Users/runner/perl5/lib/perl5
       - run: make clean && $(brew --prefix llvm@$LLVM_VERSION)/bin/scan-build --status-bugs make
         env:
-          LLVM_VERSION: ${{ matrix.os == 'macos-15' && 18 || 15 }}
+          LLVM_VERSION: ${{ matrix.os == 'macos-26' && 20 || 18 }}
           PG_CFLAGS: -DUSE_ASSERT_CHECKING
   windows:
-    runs-on: windows-latest
+    runs-on: ${{ matrix.os }}
     if: ${{ !startsWith(github.ref_name, 'mac') }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - postgres: 17
+            os: windows-2025
+          - postgres: 14
+            os: windows-2022
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - uses: ankane/setup-postgres@v1
         with:
-          postgres-version: 14
+          postgres-version: ${{ matrix.postgres }}
       - run: |
           call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat" && ^
           nmake /NOLOGO /F Makefile.win && ^
           nmake /NOLOGO /F Makefile.win install && ^
-          nmake /NOLOGO /F Makefile.win installcheck && ^
+          nmake /NOLOGO /F Makefile.win installcheck ${{ matrix.postgres != 17 && 'PG_REGRESS=$(PGROOT)\bin\pg_regress' || '' }} && ^
           nmake /NOLOGO /F Makefile.win clean && ^
           nmake /NOLOGO /F Makefile.win uninstall
         shell: cmd
@@ -125,10 +133,10 @@ jobs:
     if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - uses: ankane/setup-postgres-valgrind@v1
         with:
-          postgres-version: 17
+          postgres-version: 18
           check-ub: yes
       - run: make OPTFLAGS=""
       - run: sudo --preserve-env=PG_CONFIG make install

CHANGELOG.md

@@ -1,3 +1,12 @@
+## 0.9.0 (unreleased)
+
+- Added support for inline filtering with HNSW
+
+## 0.8.2 (unreleased)
+
+- Improved `install` target on Windows
+- Fixed `Index Searches` in `EXPLAIN` output for Postgres 18
+
 ## 0.8.1 (2025-09-04)
 
 - Added support for Postgres 18 rc1

Makefile.win

@@ -31,6 +31,9 @@ LIBDIR = $(PGROOT)\lib
 PKGLIBDIR = $(PGROOT)\lib
 SHAREDIR = $(PGROOT)\share
 
+# Use $(PGROOT)\bin\pg_regress for Postgres < 17
+PG_REGRESS = $(LIBDIR)\pgxs\src\test\regress\pg_regress
+
 CFLAGS = /nologo /I"$(INCLUDEDIR_SERVER)\port\win32_msvc" /I"$(INCLUDEDIR_SERVER)\port\win32" /I"$(INCLUDEDIR_SERVER)" /I"$(INCLUDEDIR)"
 
 CFLAGS = $(CFLAGS) $(PG_CFLAGS)
@@ -54,11 +57,11 @@ install: all
 	copy $(SHLIB) "$(PKGLIBDIR)"
 	copy $(EXTENSION).control "$(SHAREDIR)\extension"
 	copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
-	mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+	if not exist "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)" mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 	for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 installcheck:
-	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
+	"$(PG_REGRESS)" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
 
 uninstall:
 	del /f "$(PKGLIBDIR)\$(SHLIB)"

README.md

@@ -29,14 +29,14 @@ make install # may need sudo
 
 See the [installation notes](#installation-notes---linux-and-mac) if you run into issues
 
-You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), [pkg](#pkg), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
+You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), [pkg](#pkg), [APK](#apk), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
 
 ### Windows
 
 Ensure [C++ support in Visual Studio](https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170#download-and-install-the-tools) is installed and run `x64 Native Tools Command Prompt for VS [version]` as administrator. Then use `nmake` to build:
 
 ```cmd
-set "PGROOT=C:\Program Files\PostgreSQL\17"
+set "PGROOT=C:\Program Files\PostgreSQL\18"
 cd %TEMP%
 git clone --branch v0.8.1 https://github.com/pgvector/pgvector.git
 cd pgvector
@@ -467,6 +467,12 @@ If filtering by many different values, consider [partitioning](https://www.postg
 CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
 ```
 
+Or a composite HNSW index (added in 0.9.0)
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops, category_id);
+```
+
 ## Iterative Index Scans
 
 With approximate indexes, queries with filtering can return less results since filtering is applied *after* the index is scanned. Starting with 0.8.0, you can enable iterative index scans, which will automatically scan more of the index until enough results are found (or it reaches `hnsw.max_scan_tuples` or `ivfflat.max_probes`).
@@ -695,10 +701,10 @@ CREATE INDEX CONCURRENTLY ...
 
 ### Querying
 
-Use `EXPLAIN ANALYZE` to debug performance.
+Use `EXPLAIN (ANALYZE, BUFFERS)` to debug performance.
 
 ```sql
-EXPLAIN ANALYZE SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
+EXPLAIN (ANALYZE, BUFFERS) SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
 ```
 
 #### Exact Search
@@ -771,6 +777,8 @@ Use pgvector from any language with a Postgres client. You can even generate and
 
 Language | Libraries / Examples
 --- | ---
+Ada | [pgvector-ada](https://github.com/pgvector/pgvector-ada)
+Algol | [pgvector-algol](https://github.com/pgvector/pgvector-algol)
 C | [pgvector-c](https://github.com/pgvector/pgvector-c)
 C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
 C#, F#, Visual Basic | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
@@ -790,10 +798,13 @@ Lisp | [pgvector-lisp](https://github.com/pgvector/pgvector-lisp)
 Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
 Nim | [pgvector-nim](https://github.com/pgvector/pgvector-nim)
 OCaml | [pgvector-ocaml](https://github.com/pgvector/pgvector-ocaml)
+Pascal | [pgvector-pascal](https://github.com/pgvector/pgvector-pascal)
 Perl | [pgvector-perl](https://github.com/pgvector/pgvector-perl)
 PHP | [pgvector-php](https://github.com/pgvector/pgvector-php)
+Prolog | [pgvector-prolog](https://github.com/pgvector/pgvector-prolog)
 Python | [pgvector-python](https://github.com/pgvector/pgvector-python)
 R | [pgvector-r](https://github.com/pgvector/pgvector-r)
+Racket | [pgvector-racket](https://github.com/pgvector/pgvector-racket)
 Raku | [pgvector-raku](https://github.com/pgvector/pgvector-raku)
 Ruby | [pgvector-ruby](https://github.com/pgvector/pgvector-ruby), [Neighbor](https://github.com/ankane/neighbor)
 Rust | [pgvector-rust](https://github.com/pgvector/pgvector-rust)
@@ -812,7 +823,7 @@ Yes, pgvector uses the write-ahead log (WAL), which allows for replication and p
 
 #### What if I want to index vectors with more than 2,000 dimensions?
 
-You can use [half-precision indexing](#half-precision-indexing) to index up to 4,000 dimensions or [binary quantization](#binary-quantization) to index up to 64,000 dimensions. Another option is [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction).
+You can use [half-precision vectors](#half-precision-vectors) or [half-precision indexing](#half-precision-indexing) to index up to 4,000 dimensions or [binary quantization](#binary-quantization) to index up to 64,000 dimensions. Other options are [indexing subvectors](#indexing-subvectors) (for models that support it) or [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction).
 
 #### Can I store vectors with different dimensions in the same column?
 
@@ -1064,7 +1075,7 @@ 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/18/bin/pg_config
 ```
 
 Then re-run the installation instructions (run `make clean` before `make` if needed). If `sudo` is needed for `make install`, use:
@@ -1075,11 +1086,11 @@ 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/18/bin/pg_config`
+- Homebrew (arm64) - `/opt/homebrew/opt/postgresql@18/bin/pg_config`
+- Homebrew (x86-64) - `/usr/local/opt/postgresql@18/bin/pg_config`
 
-Note: Replace `17` with your Postgres server version
+Note: Replace `18` with your Postgres server version
 
 ### Missing Header
 
@@ -1088,10 +1099,10 @@ 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-18
 ```
 
-Note: Replace `17` with your Postgres server version
+Note: Replace `18` with your Postgres server version
 
 ### Missing SDK
 
@@ -1138,13 +1149,15 @@ 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-trixie
+docker pull pgvector/pgvector:pg18-trixie
 ```
 
-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 `18` with your Postgres server version, and run it the same way).
 
 Supported tags are:
 
+- `pg18-trixie`, `0.8.1-pg18-trixie`
+- `pg18-bookworm`, `0.8.1-pg18-bookworm`, `pg18`, `0.8.1-pg18`
 - `pg17-trixie`, `0.8.1-pg17-trixie`
 - `pg17-bookworm`, `0.8.1-pg17-bookworm`, `pg17`, `0.8.1-pg17`
 - `pg16-trixie`, `0.8.1-pg16-trixie`
@@ -1161,7 +1174,7 @@ You can also build the image manually:
 ```sh
 git clone --branch v0.8.1 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=18 -t myuser/pgvector .
 ```
 
 If you increase `maintenance_work_mem`, make sure `--shm-size` is at least that size to avoid an error with parallel HNSW index builds.
@@ -1178,7 +1191,7 @@ 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@18` and `postgresql@17` formulas
 
 ### PGXN
 
@@ -1193,22 +1206,22 @@ 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-18-pgvector
 ```
 
-Note: Replace `17` with your Postgres server version
+Note: Replace `18` 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_18
 # or
-sudo dnf install pgvector_17
+sudo dnf install pgvector_18
 ```
 
-Note: Replace `17` with your Postgres server version
+Note: Replace `18` with your Postgres server version
 
 ### pkg
 
@@ -1225,6 +1238,14 @@ cd /usr/ports/databases/pgvector
 make install
 ```
 
+### APK
+
+Install the Alpine package with:
+
+```sh
+apk add postgresql-pgvector
+```
+
 ### conda-forge
 
 With Conda Postgres, install from [conda-forge](https://anaconda.org/conda-forge/pgvector) with:
@@ -1267,6 +1288,7 @@ Thanks to:
 - [k-means++: The Advantage of Careful Seeding](https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf)
 - [Concept Decompositions for Large Sparse Text Data using Clustering](https://www.cs.utexas.edu/users/inderjit/public_papers/concept_mlj.pdf)
 - [Efficient and Robust Approximate Nearest Neighbor Search using Hierarchical Navigable Small World Graphs](https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf)
+- [HQANN: Efficient and Robust Similarity Search for Hybrid Queries with Structured and Unstructured Constraints](https://arxiv.org/pdf/2207.07940.pdf)
 
 ## History
 

sql/vector--0.8.1--0.9.0.sql

@@ -0,0 +1,10 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.9.0'" to load this file. \quit
+
+CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
+	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE OPERATOR CLASS vector_integer_ops
+	DEFAULT FOR TYPE integer USING hnsw AS
+	OPERATOR 2 = (integer, integer),
+	FUNCTION 4 hnsw_attribute_distance(integer, integer);

sql/vector.sql

@@ -916,3 +916,13 @@ 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);
+
+-- hnsw attributes
+
+CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
+	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE OPERATOR CLASS vector_integer_ops
+	DEFAULT FOR TYPE integer USING hnsw AS
+	OPERATOR 2 = (integer, integer),
+	FUNCTION 4 hnsw_attribute_distance(integer, integer);

src/hnsw.c

@@ -263,7 +263,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 3;
+	amroutine->amsupport = 4;
 	amroutine->amoptsprocnum = 0;
 	amroutine->amcanorder = false;
 	amroutine->amcanorderbyop = true;
@@ -274,7 +274,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 #endif
 	amroutine->amcanbackward = false;	/* can change direction mid-scan */
 	amroutine->amcanunique = false;
-	amroutine->amcanmulticol = false;
+	amroutine->amcanmulticol = true;
 	amroutine->amoptionalkey = true;
 	amroutine->amsearcharray = false;
 	amroutine->amsearchnulls = false;
@@ -334,3 +334,17 @@ hnswhandler(PG_FUNCTION_ARGS)
 
 	PG_RETURN_POINTER(amroutine);
 }
+
+/*
+ * Get the distance between two int4 attributes
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(hnsw_int4_attribute_distance);
+Datum
+hnsw_int4_attribute_distance(PG_FUNCTION_ARGS)
+{
+	int32		a = PG_GETARG_INT32(0);
+	int32		b = PG_GETARG_INT32(1);
+	double		distance = ((double) a) - ((double) b);
+
+	PG_RETURN_FLOAT8(distance);
+}

src/hnsw.h

@@ -19,6 +19,7 @@
 #define HNSW_DISTANCE_PROC 1
 #define HNSW_NORM_PROC 2
 #define HNSW_TYPE_INFO_PROC 3
+#define HNSW_ATTRIBUTE_DISTANCE_PROC 4
 
 #define HNSW_VERSION	1
 #define HNSW_MAGIC_NUMBER 0xA953A953
@@ -107,6 +108,8 @@
 #define HnswPtrPointer(hp) (hp).ptr
 #define HnswPtrOffset(hp) relptr_offset((hp).relptr)
 
+#define HnswUseIndexTuple(index) (IndexRelationGetNumberOfAttributes(index) > 1)
+
 /* Variables */
 extern int	hnsw_ef_search;
 extern int	hnsw_iterative_scan;
@@ -134,6 +137,7 @@ HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
 HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
 HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
 HnswPtrDeclare(char, DatumRelptr, DatumPtr);
+HnswPtrDeclare(IndexTupleData, IndexTupleRelptr, IndexTuplePtr);
 
 struct HnswElementData
 {
@@ -150,6 +154,7 @@ struct HnswElementData
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
 	DatumPtr	value;
+	IndexTuplePtr itup;
 	LWLock		lock;
 };
 
@@ -175,6 +180,7 @@ typedef struct HnswSearchCandidate
 	pairingheap_node w_node;
 	HnswElementPtr element;
 	double		distance;
+	bool		matches;
 }			HnswSearchCandidate;
 
 /* HNSW index options */
@@ -253,14 +259,16 @@ typedef struct HnswTypeInfo
 
 typedef struct HnswSupport
 {
-	FmgrInfo   *procinfo;
+	FmgrInfo   *procinfo[2];
 	FmgrInfo   *normprocinfo;
-	Oid			collation;
+	Oid		   *collation;
 }			HnswSupport;
 
 typedef struct HnswQuery
 {
 	Datum		value;
+	IndexTuple	itup;
+	ScanKeyData *keyData;
 }			HnswQuery;
 
 typedef struct HnswBuildState
@@ -289,6 +297,8 @@ typedef struct HnswBuildState
 	HnswGraph  *graph;
 	double		ml;
 	int			maxLevel;
+	bool		useIndexTuple;
+	TupleDesc	tupdesc;
 
 	/* Memory */
 	MemoryContext graphCtx;
@@ -417,30 +427,32 @@ bool		HnswCheckNorm(HnswSupport * support, 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, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples);
+List	   *HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, bool inMemory, 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, HnswSupport * support, int m, int efConstruction, bool existing, bool inMemory);
+HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, HnswQuery * q, Relation rel, HnswSupport * support, bool loadVec, bool inMemory);
 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);
+bool		HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, ItemPointer heaptid, bool building, TupleDesc tupdesc);
 void		HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, 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		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
+void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index);
+void		HnswLoadElement(HnswElement element, double *distance, bool *matches, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance);
+void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple);
 void		HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support);
+bool		HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support, TupleDesc tupdesc);
 bool		HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation index, int m, int lm, int lc);
 void		HnswInitLockTranche(void);
 const		HnswTypeInfo *HnswGetTypeInfo(Relation index);
 PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
+Size		HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple);
+bool		HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

src/hnswbuild.c

@@ -54,6 +54,10 @@
 #include "utils/datum.h"
 #include "utils/memutils.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"
 #else
@@ -148,6 +152,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 	Page		page;
 	HnswElementPtr iter = buildstate->graph->head;
 	char	   *base = buildstate->hnswarea;
+	bool		useIndexTuple = buildstate->useIndexTuple;
 
 	/* Calculate sizes */
 	maxSize = HNSW_MAX_SIZE;
@@ -167,7 +172,6 @@ CreateGraphPages(HnswBuildState * buildstate)
 		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
-		Pointer		valuePtr = HnswPtrAccess(base, element->value);
 
 		/* Update iterator */
 		iter = element->next;
@@ -176,7 +180,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 		MemSet(etup, 0, HNSW_TUPLE_ALLOC_SIZE);
 
 		/* Calculate sizes */
-		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(valuePtr));
+		etupSize = HnswGetElementTupleSize(base, element, useIndexTuple);
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
@@ -186,7 +190,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 					 errmsg("index tuple too large")));
 
-		HnswSetElementTuple(base, etup, element);
+		HnswSetElementTuple(base, etup, element, useIndexTuple);
 
 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
@@ -327,19 +331,18 @@ AddDuplicateInMemory(HnswElement element, HnswElement dup)
  * Find duplicate element
  */
 static bool
-FindDuplicateInMemory(char *base, HnswElement element)
+FindDuplicateInMemory(char *base, HnswElement element, bool useIndexTuple, TupleDesc tupdesc)
 {
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
-	Datum		value = HnswGetValue(base, element);
+	IndexTuple	itup = HnswPtrAccess(base, element->itup);
 
 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
-		Datum		neighborValue = HnswGetValue(base, neighborElement);
 
 		/* Exit early since ordered by distance */
-		if (!datumIsEqual(value, neighborValue, false, -1))
+		if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
 			return false;
 
 		/* Check for space */
@@ -366,7 +369,7 @@ AddElementInMemory(char *base, HnswGraph * graph, HnswElement element)
  * Update neighbors
  */
 static void
-UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
+UpdateNeighborsInMemory(char *base, Relation index, HnswSupport * support, HnswElement e, int m)
 {
 	for (int lc = e->level; lc >= 0; lc--)
 	{
@@ -388,7 +391,7 @@ 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, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, index, support);
 			LWLockRelease(&neighborElement->lock);
 		}
 	}
@@ -398,20 +401,20 @@ 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(HnswSupport * support, HnswElement element, int m, HnswElement entryPoint, HnswBuildState * buildstate)
 {
 	HnswGraph  *graph = buildstate->graph;
 	char	   *base = buildstate->hnswarea;
 
 	/* Look for duplicate */
-	if (FindDuplicateInMemory(base, element))
+	if (FindDuplicateInMemory(base, element, buildstate->useIndexTuple, buildstate->tupdesc))
 		return;
 
 	/* Add element */
 	AddElementInMemory(base, graph, element);
 
 	/* Update neighbors */
-	UpdateNeighborsInMemory(base, support, element, m);
+	UpdateNeighborsInMemory(base, buildstate->index, support, element, m);
 
 	/* Update entry point if needed (already have lock) */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -424,6 +427,7 @@ UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, int efCon
 static void
 InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 {
+	Relation	index = buildstate->index;
 	HnswGraph  *graph = buildstate->graph;
 	HnswSupport *support = &buildstate->support;
 	HnswElement entryPoint;
@@ -457,10 +461,10 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 	}
 
 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false, true);
 
 	/* Update graph in memory */
-	UpdateGraphInMemory(support, element, m, efConstruction, entryPoint, buildstate);
+	UpdateGraphInMemory(support, element, m, entryPoint, buildstate);
 
 	/* Release entry lock */
 	LWLockRelease(entryLock);
@@ -476,18 +480,20 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	HnswElement element;
 	HnswAllocator *allocator = &buildstate->allocator;
 	HnswSupport *support = &buildstate->support;
-	Size		valueSize;
-	Pointer		valuePtr;
 	LWLock	   *flushLock = &graph->flushLock;
 	char	   *base = buildstate->hnswarea;
-	Datum		value;
+	TupleDesc	tupdesc = buildstate->tupdesc;
+	IndexTuple	itup;
+	Size		itupSize;
+	IndexTuple	itupShared;
+	bool		unused;
 
-	/* Form index value */
-	if (!HnswFormIndexValue(&value, values, isnull, buildstate->typeInfo, support))
+	/* Form index tuple */
+	if (!HnswFormIndexTuple(&itup, values, isnull, buildstate->typeInfo, support, tupdesc))
 		return false;
 
-	/* Get datum size */
-	valueSize = VARSIZE_ANY(DatumGetPointer(value));
+	/* Get tuple size */
+	itupSize = IndexTupleSize(itup);
 
 	/* Ensure graph not flushed when inserting */
 	LWLockAcquire(flushLock, LW_SHARED);
@@ -497,7 +503,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	{
 		LWLockRelease(flushLock);
 
-		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
+		return HnswInsertTupleOnDisk(index, support, itup, heaptid, true, tupdesc);
 	}
 
 	/*
@@ -529,12 +535,12 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 
 		LWLockRelease(flushLock);
 
-		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
+		return HnswInsertTupleOnDisk(index, support, itup, heaptid, true, tupdesc);
 	}
 
 	/* Ok, we can proceed to allocate the element */
 	element = HnswInitElement(base, heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel, allocator);
-	valuePtr = HnswAlloc(allocator, valueSize);
+	itupShared = HnswAlloc(allocator, itupSize);
 
 	/*
 	 * We have now allocated the space needed for the element, so we don't
@@ -543,9 +549,10 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	 */
 	LWLockRelease(&graph->allocatorLock);
 
-	/* Copy the datum */
-	memcpy(valuePtr, DatumGetPointer(value), valueSize);
-	HnswPtrStore(base, element->value, valuePtr);
+	/* Copy the tuple */
+	memcpy(itupShared, itup, itupSize);
+	HnswPtrStore(base, element->itup, itupShared);
+	HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itupShared, 1, tupdesc, &unused)));
 
 	/* Create a lock for the element */
 	LWLockInitialize(&element->lock, hnsw_lock_tranche_id);
@@ -672,6 +679,19 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 				 errmsg("type not supported for hnsw index")));
 
+	/* TODO See if needed */
+	if (IndexRelationGetNumberOfKeyAttributes(index) > 2)
+		elog(ERROR, "index cannot have more than two columns");
+
+	if (!OidIsValid(index_getprocid(index, 1, HNSW_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, HNSW_ATTRIBUTE_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,
@@ -698,6 +718,8 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
+	buildstate->useIndexTuple = HnswUseIndexTuple(index);
+	buildstate->tupdesc = RelationGetDescr(index);
 
 	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
 												   "Hnsw build graph context",
@@ -1054,7 +1076,7 @@ ComputeParallelWorkers(Relation heap, Relation index)
  * Build graph
  */
 static void
-BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
+BuildGraph(HnswBuildState * buildstate)
 {
 	int			parallel_workers = 0;
 
@@ -1102,7 +1124,7 @@ BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
 
 	InitBuildState(buildstate, heap, index, indexInfo, forkNum);
 
-	BuildGraph(buildstate, forkNum);
+	BuildGraph(buildstate);
 
 	if (RelationNeedsWAL(index) || forkNum == INIT_FORKNUM)
 		log_newpage_range(index, forkNum, 0, RelationGetNumberOfBlocksInFork(index, forkNum), true);

src/hnswinsert.c

@@ -9,6 +9,10 @@
 #include "utils/datum.h"
 #include "utils/memutils.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 /*
  * Get the insert page
  */
@@ -156,9 +160,10 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 	BlockNumber newInsertPage = InvalidBlockNumber;
 	uint8		tupleVersion;
 	char	   *base = NULL;
+	bool		useIndexTuple = HnswUseIndexTuple(index);
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(HnswPtrAccess(base, e->value)));
+	etupSize = HnswGetElementTupleSize(base, e, useIndexTuple);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -166,7 +171,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 
 	/* Prepare element tuple */
 	etup = palloc0(etupSize);
-	HnswSetElementTuple(base, etup, e);
+	HnswSetElementTuple(base, etup, e, useIndexTuple);
 
 	/* Prepare neighbor tuple */
 	ntup = palloc0(ntupSize);
@@ -383,8 +388,9 @@ LoadElementsForInsert(HnswNeighborArray * neighbors, HnswQuery * q, int *idx, Re
 		HnswCandidate *hc = &neighbors->items[i];
 		HnswElement element = HnswPtrAccess(base, hc->element);
 		double		distance;
+		bool		matches;
 
-		HnswLoadElement(element, &distance, q, index, support, true, NULL);
+		HnswLoadElement(element, &distance, &matches, q, index, support, true, NULL);
 		hc->distance = distance;
 
 		/* Prune element if being deleted */
@@ -428,6 +434,8 @@ GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int
 		HnswQuery	q;
 
 		q.value = HnswGetValue(base, element);
+		q.itup = HnswPtrAccess(base, element->itup);
+		q.keyData = NULL;
 
 		LoadElementsForInsert(neighbors, &q, &idx, index, support);
 
@@ -633,21 +641,30 @@ AddDuplicateOnDisk(Relation index, HnswElement element, HnswElement dup, bool bu
  * Find duplicate element
  */
 static bool
-FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
+FindDuplicateOnDisk(Relation index, HnswElement element, bool building, TupleDesc tupdesc)
 {
 	char	   *base = NULL;
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
 	Datum		value = HnswGetValue(base, element);
+	IndexTuple	itup = HnswPtrAccess(base, element->itup);
 
 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
-		Datum		neighborValue = HnswGetValue(base, neighborElement);
 
-		/* Exit early since ordered by distance */
-		if (!datumIsEqual(value, neighborValue, false, -1))
-			return false;
+		if (HnswUseIndexTuple(index))
+		{
+			/* Exit early since ordered by distance */
+			if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
+				return false;
+		}
+		else
+		{
+			/* Exit early since ordered by distance */
+			if (!datumIsEqual(value, HnswGetValue(base, neighborElement), false, -1))
+				return false;
+		}
 
 		if (AddDuplicateOnDisk(index, element, neighborElement, building))
 			return true;
@@ -660,12 +677,12 @@ 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, HnswSupport * support, HnswElement element, int m, HnswElement entryPoint, bool building, TupleDesc tupdesc)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Look for duplicate */
-	if (FindDuplicateOnDisk(index, element, building))
+	if (FindDuplicateOnDisk(index, element, building, tupdesc))
 		return;
 
 	/* Add element */
@@ -687,7 +704,7 @@ 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, HnswSupport * support, IndexTuple itup, ItemPointer heaptid, bool building, TupleDesc tupdesc)
 {
 	HnswElement entryPoint;
 	HnswElement element;
@@ -695,6 +712,7 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 	int			efConstruction = HnswGetEfConstruction(index);
 	LOCKMODE	lockmode = ShareLock;
 	char	   *base = NULL;
+	bool		unused;
 
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -708,7 +726,8 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 
 	/* Create an element */
 	element = HnswInitElement(base, heaptid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
-	HnswPtrStore(base, element->value, DatumGetPointer(value));
+	HnswPtrStore(base, element->itup, itup);
+	HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -725,10 +744,10 @@ 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, support, m, efConstruction, false, false);
 
 	/* Update graph on disk */
-	UpdateGraphOnDisk(index, support, element, m, efConstruction, entryPoint, building);
+	UpdateGraphOnDisk(index, support, element, m, entryPoint, building, tupdesc);
 
 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -742,17 +761,18 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 static void
 HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid)
 {
-	Datum		value;
+	IndexTuple	itup;
 	const		HnswTypeInfo *typeInfo = HnswGetTypeInfo(index);
+	TupleDesc	tupdesc = RelationGetDescr(index);
 	HnswSupport support;
 
 	HnswInitSupport(&support, index);
 
-	/* Form index value */
-	if (!HnswFormIndexValue(&value, values, isnull, typeInfo, &support))
+	/* Form index tuple */
+	if (!HnswFormIndexTuple(&itup, values, isnull, typeInfo, &support, tupdesc))
 		return;
 
-	HnswInsertTupleOnDisk(index, &support, value, heaptid, false);
+	HnswInsertTupleOnDisk(index, &support, itup, heaptid, false, tupdesc);
 }
 
 /*

src/hnswscan.c

@@ -22,26 +22,30 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	int			m;
 	HnswElement entryPoint;
 	char	   *base = NULL;
+	bool		inMemory = false;
 	HnswQuery  *q = &so->q;
 
+	q->value = value;
+	q->itup = NULL;
+	q->keyData = scan->keyData;
+
 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);
 
-	q->value = value;
 	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, support, false, inMemory));
 
 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, NULL, NULL, true, NULL);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, inMemory, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
-	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, support, m, false, NULL, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &so->discarded : NULL, true, &so->tuples);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, support, m, false, NULL, inMemory, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &so->discarded : NULL, true, &so->tuples);
 }
 
 /*
@@ -72,7 +76,7 @@ ResumeScanItems(IndexScanDesc scan)
 		ep = lappend(ep, sc);
 	}
 
-	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
+	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, false, &so->v, &so->discarded, false, &so->tuples);
 }
 
 /*
@@ -96,7 +100,7 @@ GetScanValue(IndexScanDesc scan)
 
 		/* Normalize if needed */
 		if (so->support.normprocinfo != NULL)
-			value = HnswNormValue(so->typeInfo, so->support.collation, value);
+			value = HnswNormValue(so->typeInfo, so->support.collation[0], value);
 	}
 
 	return value;
@@ -193,6 +197,10 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 
 		/* Count index scan for stats */
 		pgstat_count_index_scan(scan->indexRelation);
+#if PG_VERSION_NUM >= 180000
+		if (scan->instrument)
+			scan->instrument->nsearches++;
+#endif
 
 		/* Safety check */
 		if (scan->orderByData == NULL)
@@ -279,7 +287,7 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		element = HnswPtrAccess(base, sc->element);
 
 		/* Move to next element if no valid heap TIDs */
-		if (element->heaptidsLength == 0)
+		if (!sc->matches || element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
 

src/hnswutils.c

@@ -15,6 +15,10 @@
 #include "utils/memdebug.h"
 #include "utils/rel.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #if PG_VERSION_NUM < 170000
 static inline uint64
 murmurhash64(uint64 data)
@@ -146,11 +150,39 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 void
 HnswInitSupport(HnswSupport * support, Relation index)
 {
-	support->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
-	support->collation = index->rd_indcollation[0];
+	support->procinfo[0] = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+
+	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
+		support->procinfo[1] = index_getprocinfo(index, 2, HNSW_ATTRIBUTE_DISTANCE_PROC);
+
+	support->collation = index->rd_indcollation;
 	support->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 }
 
+/*
+ * Get element tuple size
+ */
+Size
+HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple)
+{
+	Size		size;
+
+	if (useIndexTuple)
+	{
+		IndexTuple	itup = HnswPtrAccess(base, element->itup);
+
+		size = IndexTupleSize(itup);
+	}
+	else
+	{
+		Pointer		valuePtr = HnswPtrAccess(base, element->value);
+
+		size = VARSIZE_ANY(valuePtr);
+	}
+
+	return HNSW_ELEMENT_TUPLE_SIZE(size);
+}
+
 /*
  * Normalize value
  */
@@ -166,7 +198,38 @@ HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value)
 bool
 HnswCheckNorm(HnswSupport * support, Datum value)
 {
-	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation, value)) > 0;
+	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation[0], value)) > 0;
+}
+
+/*
+ * Check if index tuples are equal
+ */
+bool
+HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc)
+{
+	for (int i = 0; i < tupdesc->natts; i++)
+	{
+		bool		nullA;
+		bool		nullB;
+
+		Datum		datumA = index_getattr(a, i + 1, tupdesc, &nullA);
+		Datum		datumB = index_getattr(b, i + 1, tupdesc, &nullB);
+
+		if (nullA || nullB)
+		{
+			if (nullA != nullB)
+				return false;
+		}
+		else
+		{
+			Form_pg_attribute att = TupleDescAttr(tupdesc, i);
+
+			if (!datumIsEqual(datumA, datumB, att->attbyval, att->attlen))
+				return false;
+		}
+	}
+
+	return true;
 }
 
 /*
@@ -257,6 +320,7 @@ HnswInitElement(char *base, ItemPointer heaptid, int m, double ml, int maxLevel,
 	HnswInitNeighbors(base, element, m, allocator);
 
 	HnswPtrStore(base, element->value, (Pointer) NULL);
+	HnswPtrStore(base, element->itup, (IndexTuple) NULL);
 
 	return element;
 }
@@ -283,6 +347,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->offno = offno;
 	HnswPtrStore(base, element->neighbors, (HnswNeighborArrayPtr *) NULL);
 	HnswPtrStore(base, element->value, (Pointer) NULL);
+	HnswPtrStore(base, element->itup, (IndexTuple) NULL);
 	return element;
 }
 
@@ -395,11 +460,13 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 }
 
 /*
- * Form index value
+ * Form index tuple
  */
 bool
-HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support)
+HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support, TupleDesc tupdesc)
 {
+	Datum		newValues[2];
+
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
@@ -413,10 +480,14 @@ HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo *
 		if (!HnswCheckNorm(support, value))
 			return false;
 
-		value = HnswNormValue(typeInfo, support->collation, value);
+		value = HnswNormValue(typeInfo, support->collation[0], value);
 	}
 
-	*out = value;
+	newValues[0] = value;
+	for (int i = 1; i < tupdesc->natts; i++)
+		newValues[i] = values[i];
+
+	*out = index_form_tuple(tupdesc, newValues, isnull);
 
 	return true;
 }
@@ -425,10 +496,8 @@ HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo *
  * Set element tuple, except for neighbor info
  */
 void
-HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
+HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple)
 {
-	Pointer		valuePtr = HnswPtrAccess(base, element->value);
-
 	etup->type = HNSW_ELEMENT_TUPLE_TYPE;
 	etup->level = element->level;
 	etup->deleted = 0;
@@ -440,7 +509,19 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
+
+	if (useIndexTuple)
+	{
+		IndexTuple	itup = HnswPtrAccess(base, element->itup);
+
+		memcpy(&etup->data, itup, IndexTupleSize(itup));
+	}
+	else
+	{
+		Pointer		valuePtr = HnswPtrAccess(base, element->value);
+
+		memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
+	}
 }
 
 /*
@@ -482,7 +563,7 @@ HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
  * Load an element from a tuple
  */
 void
-HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec)
+HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index)
 {
 	element->level = etup->level;
 	element->deleted = etup->deleted;
@@ -506,26 +587,128 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	if (loadVec)
 	{
 		char	   *base = NULL;
-		Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
 
-		HnswPtrStore(base, element->value, DatumGetPointer(value));
+		if (HnswUseIndexTuple(index))
+		{
+			IndexTuple	itup = CopyIndexTuple((IndexTuple) &etup->data);
+			TupleDesc	tupdesc = RelationGetDescr(index);
+			bool		unused;
+
+			HnswPtrStore(base, element->itup, itup);
+			HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
+		}
+		else
+		{
+			Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
+
+			HnswPtrStore(base, element->value, DatumGetPointer(value));
+		}
 	}
 }
 
+/*
+ * Get the attribute distance
+ */
+static inline double
+AttributeDistance(double e)
+{
+	/* TODO Better bias */
+	/* must be >> max(w * g) + 1 / log10(2) */
+	double		bias = 4.32;
+
+	return e > 0 ? bias - 1.0 / log10(e + 1) : 0;
+}
+
 /*
  * Calculate the distance between values
  */
-static inline double
-HnswGetDistance(Datum a, Datum b, HnswSupport * support)
+static double
+HnswGetDistance(IndexTuple itup, Datum vec, HnswQuery * q, Relation index, HnswSupport * support, bool *matches)
 {
-	return DatumGetFloat8(FunctionCall2Coll(support->procinfo, support->collation, a, b));
+	double		g;
+
+	if (DatumGetPointer(q->value) == NULL)
+		g = 0;
+	else
+		g = DatumGetFloat8(FunctionCall2Coll(support->procinfo[0], support->collation[0], q->value, vec));
+
+	Assert(PointerIsValid(matches));
+	*matches = true;
+
+	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
+	{
+		double		w = 0.25;
+		double		e = 0.0;
+		TupleDesc	tupdesc = RelationGetDescr(index);
+
+		if (q->keyData)
+		{
+			/* TODO need to pass length of key data */
+			int			keyCount = 1;
+
+			for (int i = 0; i < keyCount; i++)
+			{
+				ScanKey		key = &q->keyData[i];
+				bool		isnull;
+				Datum		value = index_getattr(itup, key->sk_attno, tupdesc, &isnull);
+				bool		attnull = key->sk_flags & SK_ISNULL;
+
+				if (isnull || attnull)
+				{
+					if (isnull != attnull)
+					{
+						e += 1000;
+						*matches = false;
+					}
+				}
+				else if (!DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation, value, key->sk_argument)))
+				{
+					double		ei = fabs(DatumGetFloat8(FunctionCall2Coll(support->procinfo[key->sk_attno - 1], support->collation[key->sk_attno - 1], value, key->sk_argument)));
+
+					if (ei > 0)
+						e += ei;
+					else
+						/* Distance is zero for inequality */
+						e += 1000;
+
+					*matches = false;
+				}
+			}
+
+			return w * g + AttributeDistance(e);
+		}
+		else if (q->itup)
+		{
+			int			keyCount = IndexRelationGetNumberOfKeyAttributes(index) - 1;
+
+			for (int i = 0; i < keyCount; i++)
+			{
+				bool		isnull;
+				bool		attnull;
+				Datum		value = index_getattr(itup, i + 2, tupdesc, &isnull);
+				Datum		value2 = index_getattr(q->itup, i + 2, tupdesc, &attnull);
+
+				if (isnull || attnull)
+				{
+					if (isnull != attnull)
+						e += 1000;
+				}
+				else
+					e += fabs(DatumGetFloat8(FunctionCall2Coll(support->procinfo[i + 1], support->collation[i + 1], value, value2)));
+			}
+
+			return w * g + AttributeDistance(e);
+		}
+	}
+
+	return g;
 }
 
 /*
  * 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, double *distance, bool *matches, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance, HnswElement * element)
 {
 	Buffer		buf;
 	Page		page;
@@ -543,10 +726,23 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, Hns
 	/* Calculate distance */
 	if (distance != NULL)
 	{
-		if (DatumGetPointer(q->value) == NULL)
-			*distance = 0;
+		IndexTuple	itup = NULL;
+		Datum		value;
+
+		if (HnswUseIndexTuple(index))
+		{
+			TupleDesc	tupdesc = RelationGetDescr(index);
+			bool		unused;
+
+			itup = (IndexTuple) &etup->data;
+			value = index_getattr(itup, 1, tupdesc, &unused);
+		}
 		else
-			*distance = HnswGetDistance(q->value, PointerGetDatum(&etup->data), support);
+		{
+			value = PointerGetDatum(&etup->data);
+		}
+
+		*distance = HnswGetDistance(itup, value, q, index, support, matches);
 	}
 
 	/* Load element */
@@ -555,7 +751,7 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, Hns
 		if (*element == NULL)
 			*element = HnswInitElementFromBlock(blkno, offno);
 
-		HnswLoadElementFromTuple(*element, etup, true, loadVec);
+		HnswLoadElementFromTuple(*element, etup, true, loadVec, index);
 	}
 
 	UnlockReleaseBuffer(buf);
@@ -565,32 +761,34 @@ 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, double *distance, bool *matches, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
 {
-	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
+	HnswLoadElementImpl(element->blkno, element->offno, distance, matches, q, index, support, loadVec, maxDistance, &element);
 }
 
 /*
  * Get the distance for an element
  */
 static double
-GetElementDistance(char *base, HnswElement element, HnswQuery * q, HnswSupport * support)
+GetElementDistance(char *base, HnswElement element, bool *matches, HnswQuery * q, Relation index, HnswSupport * support)
 {
 	Datum		value = HnswGetValue(base, element);
+	IndexTuple	itup = HnswPtrAccess(base, element->itup);
 
-	return HnswGetDistance(q->value, value, support);
+	return HnswGetDistance(itup, value, q, index, support, matches);
 }
 
 /*
  * Allocate a search candidate
  */
 static HnswSearchCandidate *
-HnswInitSearchCandidate(char *base, HnswElement element, double distance)
+HnswInitSearchCandidate(char *base, HnswElement element, double distance, bool matches)
 {
 	HnswSearchCandidate *sc = palloc(sizeof(HnswSearchCandidate));
 
 	HnswPtrStore(base, sc->element, element);
 	sc->distance = distance;
+	sc->matches = matches;
 	return sc;
 }
 
@@ -598,17 +796,17 @@ HnswInitSearchCandidate(char *base, HnswElement element, double distance)
  * 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, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, bool inMemory)
 {
-	bool		inMemory = index == NULL;
 	double		distance;
+	bool		matches;
 
 	if (inMemory)
-		distance = GetElementDistance(base, entryPoint, q, support);
+		distance = GetElementDistance(base, entryPoint, &matches, q, index, support);
 	else
-		HnswLoadElement(entryPoint, &distance, q, index, support, loadVec, NULL);
+		HnswLoadElement(entryPoint, &distance, &matches, q, index, support, loadVec, NULL);
 
-	return HnswInitSearchCandidate(base, entryPoint, distance);
+	return HnswInitSearchCandidate(base, entryPoint, distance, matches);
 }
 
 /*
@@ -811,7 +1009,7 @@ 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, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
+HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, bool inMemory, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
 {
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
@@ -824,7 +1022,8 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 	int			lm = HnswGetLayerM(m, lc);
 	HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
 	int			unvisitedLength;
-	bool		inMemory = index == NULL;
+	uint64		additional = 0;
+	uint64		maxAdditional = q->keyData && lc == 0 ? 10000 : 0;
 
 	if (v == NULL)
 	{
@@ -865,6 +1064,10 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 		pairingheap_add(C, &sc->c_node);
 		pairingheap_add(W, &sc->w_node);
 
+		/* Do not count elements that do not match filter towards ef */
+		if (!sc->matches && ++additional <= maxAdditional)
+			continue;
+
 		/*
 		 * Do not count elements being deleted towards ef when vacuuming. It
 		 * would be ideal to do this for inserts as well, but this could
@@ -899,6 +1102,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			HnswElement eElement;
 			HnswSearchCandidate *e;
 			double		eDistance;
+			bool		eMatches;
 			bool		alwaysAdd = wlen < ef;
 
 			f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
@@ -906,7 +1110,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			if (inMemory)
 			{
 				eElement = unvisited[i].element;
-				eDistance = GetElementDistance(base, eElement, q, support);
+				eDistance = GetElementDistance(base, eElement, &eMatches, q, index, support);
 			}
 			else
 			{
@@ -916,7 +1120,7 @@ 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 || discarded != NULL ? NULL : &f->distance, &eElement);
+				HnswLoadElementImpl(blkno, offno, &eDistance, &eMatches, q, index, support, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);
 
 				if (eElement == NULL)
 					continue;
@@ -927,7 +1131,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 				if (discarded != NULL)
 				{
 					/* Create a new candidate */
-					e = HnswInitSearchCandidate(base, eElement, eDistance);
+					e = HnswInitSearchCandidate(base, eElement, eDistance, eMatches);
 					pairingheap_add(*discarded, &e->w_node);
 				}
 
@@ -939,7 +1143,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 				continue;
 
 			/* Create a new candidate */
-			e = HnswInitSearchCandidate(base, eElement, eDistance);
+			e = HnswInitSearchCandidate(base, eElement, eDistance, eMatches);
 			pairingheap_add(C, &e->c_node);
 			pairingheap_add(W, &e->w_node);
 
@@ -950,6 +1154,10 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			 */
 			if (CountElement(skipElement, eElement))
 			{
+				/* Do not count elements that do not match filter towards ef */
+				if (!e->matches && ++additional <= maxAdditional)
+					continue;
+
 				wlen++;
 
 				/* No need to decrement wlen */
@@ -1027,18 +1235,24 @@ CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
  * 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, Relation index, HnswSupport * support)
 {
 	HnswElement eElement = HnswPtrAccess(base, e->element);
-	Datum		eValue = HnswGetValue(base, eElement);
+	HnswQuery	q;
 	ListCell   *lc2;
 
+	q.value = HnswGetValue(base, eElement);
+	q.itup = HnswPtrAccess(base, eElement->itup);
+	q.keyData = NULL;
+
 	foreach(lc2, r)
 	{
 		HnswCandidate *ri = lfirst(lc2);
 		HnswElement riElement = HnswPtrAccess(base, ri->element);
 		Datum		riValue = HnswGetValue(base, riElement);
-		float		distance = HnswGetDistance(eValue, riValue, support);
+		IndexTuple	ritup = HnswPtrAccess(base, riElement->itup);
+		bool		matches;
+		float		distance = HnswGetDistance(ritup, riValue, &q, index, support, &matches);
 
 		if (distance <= e->distance)
 			return false;
@@ -1051,7 +1265,7 @@ 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, Relation index, HnswSupport * support, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
@@ -1085,7 +1299,7 @@ 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, index, support);
 		else if (list_length(added) > 0)
 		{
 			/* Keep Valgrind happy for in-memory, parallel builds */
@@ -1098,8 +1312,7 @@ 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, index, support);
 				if (!e->closer)
 					removedAny = true;
 			}
@@ -1111,7 +1324,7 @@ 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, index, support);
 					if (e->closer)
 						added = lappend(added, e);
 				}
@@ -1119,7 +1332,7 @@ 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, index, support);
 			if (e->closer)
 				added = lappend(added, e);
 		}
@@ -1196,7 +1409,7 @@ HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newE
 			c = lappend(c, &neighbors->items[i]);
 		c = lappend(c, &newHc);
 
-		SelectNeighbors(base, c, lm, support, &neighbors->closerSet, &newHc, &pruned, true);
+		SelectNeighbors(base, c, lm, index, support, &neighbors->closerSet, &newHc, &pruned, true);
 
 		/* Should not happen */
 		if (pruned == NULL)
@@ -1267,17 +1480,19 @@ 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, HnswSupport * support, int m, int efConstruction, bool existing, bool inMemory)
 {
 	List	   *ep;
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
 	HnswQuery	q;
+
 	HnswElement skipElement = existing ? element : NULL;
-	bool		inMemory = index == NULL;
 
 	q.value = HnswGetValue(base, element);
+	q.itup = HnswPtrAccess(base, element->itup);
+	q.keyData = NULL;
 
 	/* Precompute hash */
 	if (inMemory)
@@ -1288,13 +1503,13 @@ 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, support, true, inMemory));
 	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, NULL, NULL, true, NULL);
+		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement, inMemory, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
@@ -1313,7 +1528,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		List	   *lw = NIL;
 		ListCell   *lc2;
 
-		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
+		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, inMemory, NULL, NULL, true, NULL);
 
 		/* Convert search candidates to candidates */
 		foreach(lc2, w)
@@ -1337,7 +1552,7 @@ 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, index, support, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);
 
 		AddConnections(base, element, neighbors, lc);
 

src/hnswvacuum.c

@@ -9,6 +9,10 @@
 #include "storage/lmgr.h"
 #include "utils/memutils.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #if PG_VERSION_NUM >= 180000
 #define vacuum_delay_point() vacuum_delay_point(false)
 #endif
@@ -208,7 +212,7 @@ 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, support, m, efConstruction, true, false);
 
 	/* Zero memory for each element */
 	MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
@@ -260,7 +264,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
 
 		/* Load element */
-		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
+		HnswLoadElement(highestPoint, NULL, NULL, NULL, index, support, true, NULL);
 
 		/* Repair if needed */
 		if (NeedsUpdated(vacuumstate, highestPoint))
@@ -298,7 +302,7 @@ 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, NULL, index, support, true, NULL);
 
 			if (NeedsUpdated(vacuumstate, entryPoint))
 			{
@@ -374,7 +378,7 @@ RepairGraph(HnswVacuumState * vacuumstate)
 
 			/* Create an element */
 			element = HnswInitElementFromBlock(blkno, offno);
-			HnswLoadElementFromTuple(element, etup, false, true);
+			HnswLoadElementFromTuple(element, etup, false, true, index);
 
 			elements = lappend(elements, element);
 		}
@@ -444,6 +448,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	BlockNumber insertPage = InvalidBlockNumber;
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
+	bool		useIndexTuple = HnswUseIndexTuple(index);
 
 	/*
 	 * Wait for index scans to complete. Scans before this point may contain
@@ -525,7 +530,14 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
+			if (useIndexTuple)
+			{
+				IndexTuple	itup = (IndexTuple) &etup->data;
+
+				MemSet(itup, 0, IndexTupleSize(itup));
+			}
+			else
+				MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)

src/ivfbuild.c

@@ -20,6 +20,10 @@
 #include "utils/memutils.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"
 #else
@@ -138,7 +142,7 @@ SampleRows(IvfflatBuildState * buildstate)
  * Add tuple to sort
  */
 static void
-AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState * buildstate)
+AddTupleToSort(ItemPointer tid, Datum *values, IvfflatBuildState * buildstate)
 {
 	double		distance;
 	double		minDistance = DBL_MAX;
@@ -215,7 +219,7 @@ BuildCallback(Relation index, ItemPointer tid, Datum *values,
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Add tuple to sort */
-	AddTupleToSort(index, tid, values, buildstate);
+	AddTupleToSort(tid, values, buildstate);
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -470,8 +474,8 @@ CreateMetaPage(Relation index, int dimensions, int lists, ForkNumber forkNum)
  * Create list pages
  */
 static void
-CreateListPages(Relation index, VectorArray centers, int dimensions,
-				int lists, ForkNumber forkNum, ListInfo * *listInfo)
+CreateListPages(Relation index, VectorArray centers, int lists,
+				ForkNumber forkNum, ListInfo * *listInfo)
 {
 	Buffer		buf;
 	Page		page;
@@ -1004,7 +1008,7 @@ BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
 
 	/* Create pages */
 	CreateMetaPage(index, buildstate->dimensions, buildstate->lists, forkNum);
-	CreateListPages(index, buildstate->centers, buildstate->dimensions, buildstate->lists, forkNum, &buildstate->listInfo);
+	CreateListPages(index, buildstate->centers, buildstate->lists, forkNum, &buildstate->listInfo);
 	CreateEntryPages(buildstate, forkNum);
 
 	/* Write WAL for initialization fork since GenericXLog functions do not */

src/ivfflat.h

@@ -13,6 +13,10 @@
 #include "utils/tuplesort.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #if PG_VERSION_NUM >= 150000
 #include "common/pg_prng.h"
 #endif

src/ivfinsert.c

@@ -65,7 +65,7 @@ FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo
  * Insert a tuple into the index
  */
 static void
-InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
+InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid)
 {
 	const		IvfflatTypeInfo *typeInfo = IvfflatGetTypeInfo(index);
 	IndexTuple	itup;
@@ -204,7 +204,7 @@ ivfflatinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 	oldCtx = MemoryContextSwitchTo(insertCtx);
 
 	/* Insert tuple */
-	InsertTuple(index, values, isnull, heap_tid, heap);
+	InsertTuple(index, values, isnull, heap_tid);
 
 	/* Delete memory context */
 	MemoryContextSwitchTo(oldCtx);

src/ivfkmeans.c

@@ -13,6 +13,10 @@
 #include "utils/memutils.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 /*
  * Initialize with kmeans++
  *

src/ivfscan.c

@@ -355,6 +355,10 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 
 		/* Count index scan for stats */
 		pgstat_count_index_scan(scan->indexRelation);
+#if PG_VERSION_NUM >= 180000
+		if (scan->instrument)
+			scan->instrument->nsearches++;
+#endif
 
 		/* Safety check */
 		if (scan->orderByData == NULL)

test/t/045_hnsw_hqann.pl

@@ -0,0 +1,113 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @cs = ();
+my @expected;
+my $limit = 20;
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+my $nc = 1000;
+
+sub test_recall
+{
+	my ($min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		EXPLAIN ANALYZE SELECT i FROM tst WHERE c = $cs[0] ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Cond/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SELECT i FROM tst WHERE c = $cs[$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($dim), c int4);");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql], i % $nc FROM generate_series(1, 20000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my @r = ();
+	for (1 .. $dim)
+	{
+		push(@r, rand());
+	}
+	push(@queries, "[" . join(",", @r) . "]");
+	push(@cs, int(rand() * $nc));
+}
+
+# Get exact results
+@expected = ();
+for my $i (0 .. $#queries)
+{
+	my $res = $node->safe_psql("postgres", "SELECT i FROM tst WHERE c = $cs[$i] ORDER BY v <=> '$queries[$i]' LIMIT $limit;");
+	push(@expected, $res);
+}
+
+# Add index
+$node->safe_psql("postgres", qq(
+	SET maintenance_work_mem = '256MB';
+	SET max_parallel_maintenance_workers = 2;
+	CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, c);
+));
+
+# Test recall
+test_recall(0.99, '<=>');
+
+# Test vacuum
+$node->safe_psql("postgres", "DELETE FROM tst WHERE c > 5;");
+$node->safe_psql("postgres", "VACUUM tst;");
+
+# Test columns
+my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c);");
+like($stderr, qr/first column must be a vector/);
+
+($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c, v vector_cosine_ops);");
+like($stderr, qr/first column must be a vector/);
+
+($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, c, c);");
+like($stderr, qr/index cannot have more than two columns/);
+
+($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, v vector_cosine_ops);");
+like($stderr, qr/column 2 cannot be a vector/);
+
+done_testing();