Use maxItems 10 [skip ci]

CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.4.2 (unreleased)
+
+- Added notice when index created with little data
+- Fixed installation error with Postgres 12.0-12.2
+
 ## 0.4.1 (2023-03-21)
 
 - Improved performance of cosine distance

README.md

@@ -2,7 +2,11 @@
 
 Open-source vector similarity search for Postgres
 
-Supports exact and approximate nearest neighbor search for L2 distance, inner product, and cosine distance
+Supports
+
+- exact and approximate nearest neighbor search
+- L2 distance, inner product, and cosine distance
+- any [language](#languages) with a Postgres client
 
 [![Build Status](https://github.com/pgvector/pgvector/workflows/build/badge.svg?branch=master)](https://github.com/pgvector/pgvector/actions)
 
@@ -36,7 +40,7 @@ Create a vector column with 3 dimensions
 CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
 ```
 
-Insert values
+Insert vectors
 
 ```sql
 INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
@@ -124,7 +128,7 @@ SELECT embedding <-> '[3,1,2]' AS distance FROM items;
 For inner product, multiply by -1 (since `<#>` returns the negative inner product)
 
 ```sql
-SELECT -1 * (embedding <#> '[3,1,2]') AS inner_product FROM items;
+SELECT (embedding <#> '[3,1,2]') * -1 AS inner_product FROM items;
 ```
 
 For cosine similarity, use 1 - cosine distance
@@ -133,7 +137,7 @@ For cosine similarity, use 1 - cosine distance
 SELECT 1 - (embedding <=> '[3,1,2]') AS cosine_similarity FROM items;
 ```
 
-#### Averaging
+#### Aggregates
 
 Average vectors
 
@@ -153,15 +157,11 @@ By default, pgvector performs exact nearest neighbor search, which provides perf
 
 You can add an index to use approximate nearest neighbor search, which trades some recall for performance. Unlike typical indexes, you will see different results for queries after adding an approximate index.
 
-Two keys to achieving good recall are:
+Three keys to achieving good recall are:
 
 1. Create the index *after* the table has some data
-2. Choose an appropriate number of lists (lower is better for recall, higher is better for speed)
-
-A good place to start is:
-
-- `rows / 1000` for up to 1M rows
-- `sqrt(rows)` for over 1M rows
+2. Choose an appropriate number of lists - a good place to start is `rows / 1000` for up to 1M rows and `sqrt(rows)` for over 1M rows
+3. When querying, specify an appropriate number of [probes](#query-options) (higher is better for recall, lower is better for speed) - a good place to start is `lists / 10` for up to 1M rows and `sqrt(lists)` for over 1M rows
 
 Add an index for each distance function you want to use.
 
@@ -190,7 +190,7 @@ Vectors with up to 2,000 dimensions can be indexed.
 Specify the number of probes (1 by default)
 
 ```sql
-SET ivfflat.probes = 1;
+SET ivfflat.probes = 10;
 ```
 
 A higher value provides better recall at the cost of speed, and it can be set to the number of lists for exact nearest neighbor search (at which point the planner won’t use the index)
@@ -199,7 +199,7 @@ Use `SET LOCAL` inside a transaction to set it for a single query
 
 ```sql
 BEGIN;
-SET LOCAL ivfflat.probes = 1;
+SET LOCAL ivfflat.probes = 10;
 SELECT ...
 COMMIT;
 ```
@@ -279,8 +279,10 @@ Language | Libraries / Examples
 --- | ---
 C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
 C# | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
+Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
 Elixir | [pgvector-elixir](https://github.com/pgvector/pgvector-elixir)
 Go | [pgvector-go](https://github.com/pgvector/pgvector-go)
+Haskell | [pgvector-haskell](https://github.com/pgvector/pgvector-haskell)
 Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
 Julia | [pgvector-julia](https://github.com/pgvector/pgvector-julia)
 Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
@@ -291,6 +293,7 @@ Python | [pgvector-python](https://github.com/pgvector/pgvector-python)
 R | [pgvector-r](https://github.com/pgvector/pgvector-r)
 Ruby | [pgvector-ruby](https://github.com/pgvector/pgvector-ruby), [Neighbor](https://github.com/ankane/neighbor)
 Rust | [pgvector-rust](https://github.com/pgvector/pgvector-rust)
+Swift | [pgvector-swift](https://github.com/pgvector/pgvector-swift)
 
 ## Frequently Asked Questions
 
@@ -445,7 +448,7 @@ To request a new extension on other providers:
 - Google Cloud SQL - vote or comment on [this page](https://issuetracker.google.com/issues/265172065)
 - Azure Database - vote or comment on [this page](https://feedback.azure.com/d365community/idea/7b423322-6189-ed11-a81b-000d3ae49307)
 - DigitalOcean Managed Databases - vote or comment on [this page](https://ideas.digitalocean.com/app-framework-services/p/pgvector-extension-for-postgresql)
-- Render - vote or comment on [this page](https://feedback.render.com/features/p/add-pgvector-extension-to-postgresql)
+- Heroku Postgres - vote or comment on [this page](https://github.com/heroku/roadmap/issues/156)
 
 ## Upgrading
 

src/ivfbuild.c

@@ -431,8 +431,18 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	/* TODO Ensure within maintenance_work_mem */
 	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions);
 	if (buildstate->heap != NULL)
+	{
 		SampleRows(buildstate);
 
+		if (buildstate->samples->length < buildstate->lists)
+		{
+			ereport(NOTICE,
+					(errmsg("ivfflat index created with little data"),
+					 errdetail("this will cause poor recall"),
+					 errhint("drop the index until the table has more data")));
+		}
+	}
+
 	/* Calculate centers */
 	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers));
 

src/ivfflat.c

@@ -13,7 +13,6 @@
 #endif
 
 int			ivfflat_probes;
-int			ivfflat_bound;
 static relopt_kind ivfflat_relopt_kind;
 
 /*
@@ -33,10 +32,6 @@ _PG_init(void)
 	DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
 							"Valid range is 1..lists.", &ivfflat_probes,
 							1, 1, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
-
-	DefineCustomIntVariable("ivfflat.bound", "Sets the max results from index (experimental)",
-							NULL, &ivfflat_bound,
-							0, 0, INT_MAX, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
 /*

src/ivfflat.h

@@ -78,7 +78,6 @@
 
 /* Variables */
 extern int	ivfflat_probes;
-extern int	ivfflat_bound;
 
 /* Exported functions */
 PGDLLEXPORT void _PG_init(void);
@@ -188,6 +187,14 @@ typedef struct IvfflatScanList
 	double		distance;
 }			IvfflatScanList;
 
+typedef struct IvfflatScanItem
+{
+	pairingheap_node ph_node;
+	BlockNumber searchPage;
+	double		distance;
+	ItemPointerData tid;
+}			IvfflatScanItem;
+
 typedef struct IvfflatScanOpaqueData
 {
 	int			probes;
@@ -205,6 +212,13 @@ typedef struct IvfflatScanOpaqueData
 	FmgrInfo   *normprocinfo;
 	Oid			collation;
 
+	/* Items */
+	int			maxItems;
+	int			itemCount;
+	pairingheap *itemQueue;
+	IvfflatScanItem *items;
+	IvfflatScanItem **sortedItems;
+
 	/* Lists */
 	pairingheap *listQueue;
 	IvfflatScanList lists[FLEXIBLE_ARRAY_MEMBER];	/* must come last */

src/ivfscan.c

@@ -26,6 +26,21 @@ CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
 	return 0;
 }
 
+/*
+ * Compare item distances
+ */
+static int
+CompareItems(const pairingheap_node *a, const pairingheap_node *b, void *arg)
+{
+	if (((const IvfflatScanItem *) a)->distance > ((const IvfflatScanItem *) b)->distance)
+		return 1;
+
+	if (((const IvfflatScanItem *) a)->distance < ((const IvfflatScanItem *) b)->distance)
+		return -1;
+
+	return 0;
+}
+
 /*
  * Get lists and sort by distance
  */
@@ -111,13 +126,10 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	Datum		datum;
 	bool		isnull;
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
-	double		tuples = 0;
-
-#if PG_VERSION_NUM >= 120000
-	TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
-#else
-	TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc);
-#endif
+	int			i;
+	double		distance;
+	IvfflatScanItem *scanitem;
+	double		maxDistance = DBL_MAX;
 
 	/*
 	 * Reuse same set of shared buffers for scan
@@ -126,10 +138,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	 */
 	BufferAccessStrategy bas = GetAccessStrategy(BAS_BULKREAD);
 
-	/* Set the max number of results */
-	if (ivfflat_bound > 0)
-		tuplesort_set_bound(so->sortstate, ivfflat_bound);
-
 	/* Search closest probes lists */
 	while (!pairingheap_is_empty(so->listQueue))
 	{
@@ -147,25 +155,40 @@ GetScanItems(IndexScanDesc scan, Datum value)
 			{
 				itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
+				distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, datum, value));
 
-				/*
-				 * Add virtual tuple
-				 *
-				 * Use procinfo from the index instead of scan key for
-				 * performance
-				 */
-				ExecClearTuple(slot);
-				slot->tts_values[0] = FunctionCall2Coll(so->procinfo, so->collation, datum, value);
-				slot->tts_isnull[0] = false;
-				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
-				slot->tts_isnull[1] = false;
-				slot->tts_values[2] = Int32GetDatum((int) searchPage);
-				slot->tts_isnull[2] = false;
-				ExecStoreVirtualTuple(slot);
+				if (so->itemCount < so->maxItems)
+				{
+					scanitem = &so->items[so->itemCount];
+					scanitem->searchPage = searchPage;
+					scanitem->tid = itup->t_tid;
+					scanitem->distance = distance;
+					so->itemCount++;
 
-				tuplesort_puttupleslot(so->sortstate, slot);
+					/* Add to heap */
+					pairingheap_add(so->itemQueue, &scanitem->ph_node);
 
-				tuples++;
+					/* Calculate max distance */
+					if (so->itemCount == so->maxItems)
+					{
+						maxDistance = ((IvfflatScanItem *) pairingheap_first(so->itemQueue))->distance;
+						scanitem = &so->items[so->itemCount];
+					}
+				}
+				else if (distance < maxDistance)
+				{
+					/* Reuse */
+					scanitem->searchPage = searchPage;
+					scanitem->tid = itup->t_tid;
+					scanitem->distance = distance;
+					pairingheap_add(so->itemQueue, &scanitem->ph_node);
+
+					/* Remove */
+					scanitem = (IvfflatScanItem *) pairingheap_remove_first(so->itemQueue);
+
+					/* Update max distance */
+					maxDistance = ((IvfflatScanItem *) pairingheap_first(so->itemQueue))->distance;
+				}
 			}
 
 			searchPage = IvfflatPageGetOpaque(page)->nextblkno;
@@ -174,14 +197,10 @@ GetScanItems(IndexScanDesc scan, Datum value)
 		}
 	}
 
-	/* TODO Scan more lists */
-	if (tuples < 100)
-		ereport(DEBUG1,
-				(errmsg("index scan found few tuples"),
-				 errdetail("index may have been created without data or lists is too high"),
-				 errhint("recreate the index and possibly decrease lists")));
+	for (i = 0; i < so->itemCount; i++)
+		so->sortedItems[i] = (IvfflatScanItem *) pairingheap_remove_first(so->itemQueue);
 
-	tuplesort_performsort(so->sortstate);
+	Assert(pairingheap_is_empty(so->itemQueue));
 }
 
 /*
@@ -193,10 +212,6 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	IndexScanDesc scan;
 	IvfflatScanOpaque so;
 	int			lists;
-	AttrNumber	attNums[] = {1};
-	Oid			sortOperators[] = {Float8LessOperator};
-	Oid			sortCollations[] = {InvalidOid};
-	bool		nullsFirstFlags[] = {false};
 	int			probes = ivfflat_probes;
 
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
@@ -215,27 +230,14 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	so->collation = index->rd_indcollation[0];
 
-	/* Create tuple description for sorting */
-#if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(3);
-#else
-	so->tupdesc = CreateTemplateTupleDesc(3, false);
-#endif
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
-
-	/* Prep sort */
-	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
-
-#if PG_VERSION_NUM >= 120000
-	so->slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsMinimalTuple);
-#else
-	so->slot = MakeSingleTupleTableSlot(so->tupdesc);
-#endif
-
 	so->listQueue = pairingheap_allocate(CompareLists, scan);
 
+	so->maxItems = 10;
+	so->itemCount = 0;
+	so->itemQueue = pairingheap_allocate(CompareItems, scan);
+	so->items = palloc(sizeof(IvfflatScanItem) * (so->maxItems + 1));
+	so->sortedItems = palloc(sizeof(IvfflatScanItem *) * so->maxItems);
+
 	scan->opaque = so;
 
 	return scan;
@@ -249,13 +251,10 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 
-#if PG_VERSION_NUM >= 130000
-	if (!so->first)
-		tuplesort_reset(so->sortstate);
-#endif
-
 	so->first = true;
 	pairingheap_reset(so->listQueue);
+	pairingheap_reset(so->itemQueue);
+	so->itemCount = 0;
 
 	if (keys && scan->numberOfKeys > 0)
 		memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));
@@ -315,15 +314,18 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 			pfree(DatumGetPointer(value));
 	}
 
-	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
+	if (so->itemCount > 0)
 	{
-		ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
-		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
+		IvfflatScanItem *scanitem;
+
+		so->itemCount--;
+
+		scanitem = so->sortedItems[so->itemCount];
 
 #if PG_VERSION_NUM >= 120000
-		scan->xs_heaptid = *tid;
+		scan->xs_heaptid = scanitem->tid;
 #else
-		scan->xs_ctup.t_self = *tid;
+		scan->xs_ctup.t_self = scanitem->tid;
 #endif
 
 		if (BufferIsValid(so->buf))
@@ -335,7 +337,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		 *
 		 * https://www.postgresql.org/docs/current/index-locking.html
 		 */
-		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
+		so->buf = ReadBuffer(scan->indexRelation, scanitem->searchPage);
 
 		scan->xs_recheckorderby = false;
 		return true;
@@ -357,7 +359,10 @@ ivfflatendscan(IndexScanDesc scan)
 		ReleaseBuffer(so->buf);
 
 	pairingheap_free(so->listQueue);
-	tuplesort_end(so->sortstate);
+
+	pairingheap_free(so->itemQueue);
+	pfree(so->items);
+	pfree(so->sortedItems);
 
 	pfree(so);
 	scan->opaque = NULL;

src/vector.c

@@ -100,7 +100,7 @@ CheckStateArray(ArrayType *statearray, const char *caller)
 	return (float8 *) ARR_DATA_PTR(statearray);
 }
 
-#if PG_VERSION_NUM < 120000
+#if PG_VERSION_NUM < 120003
 static pg_noinline void
 float_overflow_error(void)
 {