redis/hnsw.c

2555 lines
98 KiB
C

/* HNSW (Hierarchical Navigable Small World) Implementation.
*
* Based on the paper by Yu. A. Malkov, D. A. Yashunin.
*
* Many details of this implementation, not covered in the paper, were
* obtained simulating different workloads and checking the connection
* quality of the graph.
*
* Notably, this implementation:
*
* 1. Only uses bi-directional links, implementing strategies in order to
* link new nodes even when candidates are full, and our new node would
* be not close enough to replace old links in candidate.
*
* 2. We normalize on-insert, making cosine similarity and dot product the
* same. This means we can't use euclidian distance or alike here.
* Together with quantization, this provides an important speedup that
* makes HNSW more practical.
*
* 3. The quantization used is int8. And it is performed per-vector, so the
* "range" (max abs value) is also stored alongside with the quantized data.
*
* 4. This library implements true elements deletion, not just marking the
* element as deleted, but removing it (we can do it since our links are
* bidirectional), and reliking the nodes orphaned of one link among
* them.
*
* Copyright(C) 2024-2025 Salvatore Sanfilippo. All Rights Reserved.
*/
#define _DEFAULT_SOURCE
#define _POSIX_C_SOURCE 200809L
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <stdint.h>
#include <float.h> /* for INFINITY if not in math.h */
#include <assert.h>
#include "hnsw.h"
#if 0
#define debugmsg printf
#else
#define debugmsg if(0) printf
#endif
#ifndef INFINITY
#define INFINITY (1.0/0.0)
#endif
#define MIN(a,b) ((a) < (b) ? (a) : (b))
/* Algorithm parameters. */
#define HNSW_M 16 /* Number of max connections per node. Note that
* layer zero has twice as many. Also note that
* when a new node is added, we will populate
* even layer 0 links to just HNSW_M neighbors, so
* initially half layer 0 slots will be empty. */
#define HNSW_M0 (HNSW_M*2) /* Maximum number of connections for layer 0 */
#define HNSW_P 0.25 /* Probability of level increase. */
#define HNSW_MAX_LEVEL 16 /* Max level nodes can reach. */
#define HNSW_EF_C 200 /* Default size of dynamic candidate list while
* inserting a new node, in case 0 is passed to
* the 'ef' argument while inserting. This is also
* used when deleting nodes for the search step
* needed sometimes to reconnect nodes that remain
* orphaned of one link. */
static void (*hfree)(void *p) = free;
static void *(*hmalloc)(size_t s) = malloc;
static void *(*hrealloc)(void *old, size_t s) = realloc;
void hnsw_set_allocator(void (*free_ptr)(void*), void *(*malloc_ptr)(size_t),
void *(*realloc_ptr)(void*, size_t))
{
hfree = free_ptr;
hmalloc = malloc_ptr;
hrealloc = realloc_ptr;
}
// Get a warning if you use the libc allocator functions for mistake.
#define malloc use_hmalloc_instead
#define realloc use_hrealloc_instead
#define free use_hfree_instead
/* ============================== Prototypes ================================ */
void hnsw_cursor_element_deleted(HNSW *index, hnswNode *deleted);
/* ============================ Priority queue ================================
* We need a priority queue to take an ordered list of candidates. Right now
* it is implemented as a linear array, since it is relatively small.
*
* You may find it to be odd that we take the best element (smaller distance)
* at the end of the array, but this way popping from the pqueue is O(1), as
* we need to just decrement the count, and this is a very used operation
* in a critical code path. This makes the priority queue implementation a
* bit more complex in the insertion, but for good reasons. */
/* Maximum number of candidates we'll ever need (cit. Bill Gates). */
#define HNSW_MAX_CANDIDATES 256
typedef struct {
hnswNode *node;
float distance;
} pqitem;
typedef struct {
pqitem *items; /* Array of items. */
uint32_t count; /* Current number of items. */
uint32_t cap; /* Maximum capacity. */
} pqueue;
/* The HNSW algorithms access the pqueue conceptually from nearest (index 0)
* to farest (larger indexes) node, so the following macros are used to
* access the pqueue in this fashion, even if the internal order is
* actually reversed. */
#define pq_get_node(q,i) ((q)->items[(q)->count-(i+1)].node)
#define pq_get_distance(q,i) ((q)->items[(q)->count-(i+1)].distance)
/* Create a new priority queue with given capacity. Adding to the
* pqueue only retains 'capacity' elements with the shortest distance. */
pqueue *pq_new(uint32_t capacity) {
pqueue *pq = hmalloc(sizeof(*pq));
if (!pq) return NULL;
pq->items = hmalloc(sizeof(pqitem) * capacity);
if (!pq->items) {
hfree(pq);
return NULL;
}
pq->count = 0;
pq->cap = capacity;
return pq;
}
/* Free a priority queue. */
void pq_free(pqueue *pq) {
if (!pq) return;
hfree(pq->items);
hfree(pq);
}
/* Insert maintaining distance order (higher distances first). */
void pq_push(pqueue *pq, hnswNode *node, float distance) {
if (pq->count < pq->cap) {
/* Queue not full: shift right from high distances to make room. */
uint32_t i = pq->count;
while (i > 0 && pq->items[i-1].distance < distance) {
pq->items[i] = pq->items[i-1];
i--;
}
pq->items[i].node = node;
pq->items[i].distance = distance;
pq->count++;
} else {
/* Queue full: if new item is worse than worst, ignore it. */
if (distance >= pq->items[0].distance) return;
/* Otherwise shift left from low distances to drop worst. */
uint32_t i = 0;
while (i < pq->cap-1 && pq->items[i+1].distance > distance) {
pq->items[i] = pq->items[i+1];
i++;
}
pq->items[i].node = node;
pq->items[i].distance = distance;
}
}
/* Remove and return the top (closest) element, which is at count-1
* since we store elements with higher distances first.
* Runs in constant time. */
hnswNode *pq_pop(pqueue *pq, float *distance) {
if (pq->count == 0) return NULL;
pq->count--;
*distance = pq->items[pq->count].distance;
return pq->items[pq->count].node;
}
/* Get distance of the furthest element.
* An empty priority queue has infinite distance as its furthest element,
* note that this behavior is needed by the algorithms below. */
float pq_max_distance(pqueue *pq) {
if (pq->count == 0) return INFINITY;
return pq->items[0].distance;
}
/* ============================ HNSW algorithm ============================== */
/* Dot product: our vectors are already normalized.
* Version for not quantized vectors of floats. */
float vectors_distance_float(const float *x, const float *y, uint32_t dim) {
/* Use two accumulators to reduce dependencies among multiplications.
* This provides a clear speed boost in Apple silicon, but should be
* help in general. */
float dot0 = 0.0f, dot1 = 0.0f;
uint32_t i;
// Process 8 elements per iteration, 50/50 with the two accumulators.
for (i = 0; i + 7 < dim; i += 8) {
dot0 += x[i] * y[i] +
x[i+1] * y[i+1] +
x[i+2] * y[i+2] +
x[i+3] * y[i+3];
dot1 += x[i+4] * y[i+4] +
x[i+5] * y[i+5] +
x[i+6] * y[i+6] +
x[i+7] * y[i+7];
}
/* Handle the remaining elements. These are a minority in the case
* of a smal vector, don't optimze this part. */
for (; i < dim; i++) dot0 += x[i] * y[i];
/* The following line may be counter intuitive. The dot product of
* normalized vectors is equivalent to their cosine similarity. The
* cosine will be from -1 (vectors facing opposite directions in the
* N-dim space) to 1 (vectors are facing in the same direction).
*
* We kinda want a "score" of distance from 0 to 2 (this is a distance
* function and we want minimize the distance for K-NN searches), so we
* can't just add 1: that would return a number in the 0-2 range, with
* 0 meaning opposite vectors and 2 identical vectors, so this is
* similarity, not distance.
*
* Returning instead (1 - dotprod) inverts the meaning: 0 is identical
* and 2 is opposite, hence it is their distance.
*
* Why don't normalize the similarity right now, and return from 0 to
* 1? Because division is costly. */
return 1.0f - (dot0 + dot1);
}
/* Q8 quants dotproduct. We do integer math and later fix it by range. */
float vectors_distance_q8(const int8_t *x, const int8_t *y, uint32_t dim,
float range_a, float range_b) {
// Handle zero vectors special case.
if (range_a == 0 || range_b == 0) {
/* Zero vector distance from anything is 1.0
* (since 1.0 - dot_product where dot_product = 0). */
return 1.0f;
}
/* Each vector is quantized from [-max_abs, +max_abs] to [-127, 127]
* where range = 2*max_abs. */
const float scale_product = (range_a/127) * (range_b/127);
int32_t dot0 = 0, dot1 = 0;
uint32_t i;
// Process 8 elements at a time for better pipeline utilization.
for (i = 0; i + 7 < dim; i += 8) {
dot0 += ((int32_t)x[i]) * ((int32_t)y[i]) +
((int32_t)x[i+1]) * ((int32_t)y[i+1]) +
((int32_t)x[i+2]) * ((int32_t)y[i+2]) +
((int32_t)x[i+3]) * ((int32_t)y[i+3]);
dot1 += ((int32_t)x[i+4]) * ((int32_t)y[i+4]) +
((int32_t)x[i+5]) * ((int32_t)y[i+5]) +
((int32_t)x[i+6]) * ((int32_t)y[i+6]) +
((int32_t)x[i+7]) * ((int32_t)y[i+7]);
}
// Handle remaining elements.
for (; i < dim; i++) dot0 += ((int32_t)x[i]) * ((int32_t)y[i]);
// Convert to original range.
float dotf = (dot0 + dot1) * scale_product;
float distance = 1.0f - dotf;
// Clamp distance to [0, 2].
if (distance < 0) distance = 0;
else if (distance > 2) distance = 2;
return distance;
}
static inline int popcount64(uint64_t x) {
x = (x & 0x5555555555555555) + ((x >> 1) & 0x5555555555555555);
x = (x & 0x3333333333333333) + ((x >> 2) & 0x3333333333333333);
x = (x & 0x0F0F0F0F0F0F0F0F) + ((x >> 4) & 0x0F0F0F0F0F0F0F0F);
x = (x & 0x00FF00FF00FF00FF) + ((x >> 8) & 0x00FF00FF00FF00FF);
x = (x & 0x0000FFFF0000FFFF) + ((x >> 16) & 0x0000FFFF0000FFFF);
x = (x & 0x00000000FFFFFFFF) + ((x >> 32) & 0x00000000FFFFFFFF);
return x;
}
/* Binary vectors distance. */
float vectors_distance_bin(const uint64_t *x, const uint64_t *y, uint32_t dim) {
uint32_t len = (dim+63)/64;
uint32_t opposite = 0;
for (uint32_t j = 0; j < len; j++) {
uint64_t xor = x[j]^y[j];
opposite += popcount64(xor);
}
return (float)opposite*2/dim;
}
/* Dot product between nodes. Will call the right version depending on the
* quantization used. */
float hnsw_distance(HNSW *index, hnswNode *a, hnswNode *b) {
switch(index->quant_type) {
case HNSW_QUANT_NONE:
return vectors_distance_float(a->vector,b->vector,index->vector_dim);
case HNSW_QUANT_Q8:
return vectors_distance_q8(a->vector,b->vector,index->vector_dim,a->quants_range,b->quants_range);
case HNSW_QUANT_BIN:
return vectors_distance_bin(a->vector,b->vector,index->vector_dim);
default:
assert(1 != 1);
return 0;
}
}
/* This do Q8 'range' quantization.
* For people looking at this code thinking: Oh, I could use min/max
* quants instead! Well: I tried with min/max normalization but the dot
* product needs to accumulate the sum for later correction, and it's slower. */
void quantize_to_q8(float *src, int8_t *dst, uint32_t dim, float *rangeptr) {
float max_abs = 0;
for (uint32_t j = 0; j < dim; j++) {
if (src[j] > max_abs) max_abs = src[j];
if (-src[j] > max_abs) max_abs = -src[j];
}
if (max_abs == 0) {
if (rangeptr) *rangeptr = 0;
memset(dst, 0, dim);
return;
}
const float scale = 127.0f / max_abs; // Scale to map to [-127, 127].
for (uint32_t j = 0; j < dim; j++) {
dst[j] = (int8_t)roundf(src[j] * scale);
}
if (rangeptr) *rangeptr = max_abs; // Return max_abs instead of 2*max_abs.
}
/* Binary quantization of vector 'src' to 'dst'. We use full words of
* 64 bit as smallest unit, we will just set all the unused bits to 0
* so that they'll be the same in all the vectors, and when xor+popcount
* is used to compute the distance, such bits are not considered. This
* allows to go faster. */
void quantize_to_bin(float *src, uint64_t *dst, uint32_t dim) {
memset(dst,0,(dim+63)/64*sizeof(uint64_t));
for (uint32_t j = 0; j < dim; j++) {
uint32_t word = j/64;
uint32_t bit = j&63;
/* Since cosine similarity checks the vector direction and
* not magnitudo, we do likewise in the binary quantization and
* just remember if the component is positive or negative. */
if (src[j] > 0) dst[word] |= 1ULL<<bit;
}
}
/* L2 normalization of the float vector.
*
* Store the L2 value on 'l2ptr' if not NULL. This way the process
* can be reversed even if some precision will be lost. */
void hnsw_normalize_vector(float *x, float *l2ptr, uint32_t dim) {
float l2 = 0;
uint32_t i;
for (i = 0; i + 3 < dim; i += 4) {
l2 += x[i]*x[i] +
x[i+1]*x[i+1] +
x[i+2]*x[i+2] +
x[i+3]*x[i+3];
}
for (; i < dim; i++) l2 += x[i]*x[i];
if (l2 == 0) return; // All zero vector, can't normalize.
l2 = sqrtf(l2);
if (l2ptr) *l2ptr = l2;
for (i = 0; i < dim; i++) x[i] /= l2;
}
/* Helper function to generate random level. */
uint32_t random_level() {
static const int threshold = HNSW_P * RAND_MAX;
uint32_t level = 0;
while (rand() < threshold && level < HNSW_MAX_LEVEL)
level += 1;
return level;
}
/* Create new HNSW index, quantized or not. */
HNSW *hnsw_new(uint32_t vector_dim, uint32_t quant_type) {
HNSW *index = hmalloc(sizeof(HNSW));
if (!index) return NULL;
index->quant_type = quant_type;
index->enter_point = NULL;
index->max_level = 0;
index->vector_dim = vector_dim;
index->node_count = 0;
index->last_id = 0;
index->head = NULL;
index->cursors = NULL;
/* Initialize epochs array. */
for (int i = 0; i < HNSW_MAX_THREADS; i++)
index->current_epoch[i] = 0;
/* Initialize locks. */
if (pthread_rwlock_init(&index->global_lock, NULL) != 0) {
hfree(index);
return NULL;
}
for (int i = 0; i < HNSW_MAX_THREADS; i++) {
if (pthread_mutex_init(&index->slot_locks[i], NULL) != 0) {
/* Clean up previously initialized mutexes. */
for (int j = 0; j < i; j++)
pthread_mutex_destroy(&index->slot_locks[j]);
pthread_rwlock_destroy(&index->global_lock);
hfree(index);
return NULL;
}
}
/* Initialize atomic variables. */
index->next_slot = 0;
index->version = 0;
return index;
}
/* Fill 'vec' with the node vector, de-normalizing and de-quantizing it
* as needed. Note that this function will return an approximated version
* of the original vector. */
void hnsw_get_node_vector(HNSW *index, hnswNode *node, float *vec) {
if (index->quant_type == HNSW_QUANT_NONE) {
memcpy(vec,node->vector,index->vector_dim*sizeof(float));
} else if (index->quant_type == HNSW_QUANT_Q8) {
int8_t *quants = node->vector;
for (uint32_t j = 0; j < index->vector_dim; j++)
vec[j] = (quants[j]*node->quants_range)/127;
} else if (index->quant_type == HNSW_QUANT_BIN) {
uint64_t *bits = node->vector;
for (uint32_t j = 0; j < index->vector_dim; j++) {
uint32_t word = j/64;
uint32_t bit = j&63;
vec[j] = (bits[word] & (1ULL<<bit)) ? 1.0f : -1.0f;
}
}
// De-normalize.
if (index->quant_type != HNSW_QUANT_BIN) {
for (uint32_t j = 0; j < index->vector_dim; j++)
vec[j] *= node->l2;
}
}
/* Return the number of bytes needed to represent a vector in the index,
* that is function of the dimension of the vectors and the quantization
* type used. */
uint32_t hnsw_quants_bytes(HNSW *index) {
switch(index->quant_type) {
case HNSW_QUANT_NONE: return index->vector_dim * sizeof(float);
case HNSW_QUANT_Q8: return index->vector_dim;
case HNSW_QUANT_BIN: return (index->vector_dim+63)/64*8;
default: assert(0 && "Quantization type not supported.");
}
}
/* Create new node. Returns NULL on out of memory.
* It is possible to pass the vector as floats or, in case this index
* was already stored on disk and is being loaded, or serialized and
* transmitted in any form, the already quantized version in
* 'qvector'.
*
* Only vector or qvector should be non-NULL. The reason why passing
* a quantized vector is useful, is that because re-normalizing and
* re-quantizing several times the same vector may accumulate rounding
* errors. So if you work with quantized indexes, you should save
* the quantized indexes.
*
* Note that, together with qvector, the quantization range is needed,
* since this library uses per-vector quantization. In case of quantized
* vectors the l2 is considered to be '1', so if you want to restore
* the right l2 (to use the API that returns an approximation of the
* original vector) make sure to save the l2 on disk and set it back
* after the node creation (see later for the serialization API that
* handles this and more). */
hnswNode *hnsw_node_new(HNSW *index, uint64_t id, const float *vector, const int8_t *qvector, float qrange, uint32_t level, int normalize) {
hnswNode *node = hmalloc(sizeof(hnswNode)+(sizeof(hnswNodeLayer)*(level+1)));
if (!node) return NULL;
if (id == 0) id = ++index->last_id;
node->level = level;
node->id = id;
node->next = NULL;
node->vector = NULL;
node->l2 = 1; // Default in case of already quantized vectors. It is
// up to the caller to fill this later, if needed.
/* Initialize visited epoch array. */
for (int i = 0; i < HNSW_MAX_THREADS; i++)
node->visited_epoch[i] = 0;
if (qvector == NULL) {
/* Copy input vector. */
node->vector = hmalloc(sizeof(float) * index->vector_dim);
if (!node->vector) {
hfree(node);
return NULL;
}
memcpy(node->vector, vector, sizeof(float) * index->vector_dim);
if (normalize)
hnsw_normalize_vector(node->vector,&node->l2,index->vector_dim);
/* Handle quantization. */
if (index->quant_type != HNSW_QUANT_NONE) {
void *quants = hmalloc(hnsw_quants_bytes(index));
if (quants == NULL) {
hfree(node->vector);
hfree(node);
return NULL;
}
// Quantize.
switch(index->quant_type) {
case HNSW_QUANT_Q8:
quantize_to_q8(node->vector,quants,index->vector_dim,&node->quants_range);
break;
case HNSW_QUANT_BIN:
quantize_to_bin(node->vector,quants,index->vector_dim);
break;
default:
assert(0 && "Quantization type not handled.");
break;
}
// Discard the full precision vector.
hfree(node->vector);
node->vector = quants;
}
} else {
// We got the already quantized vector. Just copy it.
assert(index->quant_type != HNSW_QUANT_NONE);
uint32_t vector_bytes = hnsw_quants_bytes(index);
node->vector = hmalloc(vector_bytes);
node->quants_range = qrange;
if (node->vector == NULL) {
hfree(node);
return NULL;
}
memcpy(node->vector,qvector,vector_bytes);
}
/* Initialize each layer. */
for (uint32_t i = 0; i <= level; i++) {
uint32_t max_links = (i == 0) ? HNSW_M0 : HNSW_M;
node->layers[i].max_links = max_links;
node->layers[i].num_links = 0;
node->layers[i].worst_distance = 0;
node->layers[i].worst_idx = 0;
node->layers[i].links = hmalloc(sizeof(hnswNode*) * max_links);
if (!node->layers[i].links) {
for (uint32_t j = 0; j < i; j++) hfree(node->layers[j].links);
hfree(node->vector);
hfree(node);
return NULL;
}
}
return node;
}
/* Free a node. */
void hnsw_node_free(hnswNode *node) {
if (!node) return;
for (uint32_t i = 0; i <= node->level; i++)
hfree(node->layers[i].links);
hfree(node->vector);
hfree(node);
}
/* Free the entire index. */
void hnsw_free(HNSW *index,void(*free_value)(void*value)) {
if (!index) return;
hnswNode *current = index->head;
while (current) {
hnswNode *next = current->next;
if (free_value) free_value(current->value);
hnsw_node_free(current);
current = next;
}
/* Destroy locks */
pthread_rwlock_destroy(&index->global_lock);
for (int i = 0; i < HNSW_MAX_THREADS; i++) {
pthread_mutex_destroy(&index->slot_locks[i]);
}
hfree(index);
}
/* Add node to linked list of nodes. We may need to scan the whole
* HNSW graph for several reasons. The list is doubly linked since we
* also need the ability to remove a node without scanning the whole thing. */
void hnsw_add_node(HNSW *index, hnswNode *node) {
node->next = index->head;
node->prev = NULL;
if (index->head)
index->head->prev = node;
index->head = node;
index->node_count++;
}
/* Search the specified layer starting from the specified entry point
* to collect 'ef' nodes that are near to 'query'.
*
* This function implements optional hybrid search, so that each node
* can be accepted or not based on its associated value. In this case
* a callback 'filter_callback' should be passed, together with a maximum
* effort for the search (number of candidates to evaluate), since even
* with a a low "EF" value we risk that there are too few nodes that satisfy
* the provided filter, and we could trigger a full scan. */
pqueue *search_layer_with_filter(
HNSW *index, hnswNode *query, hnswNode *entry_point,
uint32_t ef, uint32_t layer, uint32_t slot,
int (*filter_callback)(void *value, void *privdata),
void *filter_privdata, uint32_t max_candidates)
{
// Mark visited nodes with a never seen epoch.
index->current_epoch[slot]++;
pqueue *candidates = pq_new(HNSW_MAX_CANDIDATES);
pqueue *results = pq_new(ef);
if (!candidates || !results) {
if (candidates) pq_free(candidates);
if (results) pq_free(results);
return NULL;
}
// Take track of the total effort: only used when filtering via
// a callback to have a bound effort.
uint32_t evaluated_candidates = 1;
// Add entry point.
float dist = hnsw_distance(index, query, entry_point);
pq_push(candidates, entry_point, dist);
if (filter_callback == NULL ||
filter_callback(entry_point->value, filter_privdata))
{
pq_push(results, entry_point, dist);
}
entry_point->visited_epoch[slot] = index->current_epoch[slot];
// Process candidates.
while (candidates->count > 0) {
// Max effort. If zero, we keep scanning.
if (filter_callback &&
max_candidates &&
evaluated_candidates >= max_candidates) break;
float cur_dist;
hnswNode *current = pq_pop(candidates, &cur_dist);
evaluated_candidates++;
float furthest = pq_max_distance(results);
if (results->count >= ef && cur_dist > furthest) break;
/* Check neighbors. */
for (uint32_t i = 0; i < current->layers[layer].num_links; i++) {
hnswNode *neighbor = current->layers[layer].links[i];
if (neighbor->visited_epoch[slot] == index->current_epoch[slot])
continue; // Already visited during this scan.
neighbor->visited_epoch[slot] = index->current_epoch[slot];
float neighbor_dist = hnsw_distance(index, query, neighbor);
furthest = pq_max_distance(results);
if (filter_callback == NULL) {
/* Original HNSW logic when no filtering:
* Add to results if better than current max or
* results not full. */
if (neighbor_dist < furthest || results->count < ef) {
pq_push(candidates, neighbor, neighbor_dist);
pq_push(results, neighbor, neighbor_dist);
}
} else {
/* With filtering: we add candidates even if doesn't match
* the filter, in order to continue to explore the graph. */
if (neighbor_dist < furthest || candidates->count < ef) {
pq_push(candidates, neighbor, neighbor_dist);
}
/* Add results only if passes filter. */
if (filter_callback(neighbor->value, filter_privdata)) {
if (neighbor_dist < furthest || results->count < ef) {
pq_push(results, neighbor, neighbor_dist);
}
}
}
}
}
pq_free(candidates);
return results;
}
/* Just a wrapper without hybrid search callback. */
pqueue *search_layer(HNSW *index, hnswNode *query, hnswNode *entry_point,
uint32_t ef, uint32_t layer, uint32_t slot)
{
return search_layer_with_filter(index, query, entry_point, ef, layer, slot,
NULL, NULL, 0);
}
/* This function is used in order to initialize a node allocated in the
* function stack with the specified vector. The idea is that we can
* easily use hnsw_distance() from a vector and the HNSW nodes this way:
*
* hnswNode myQuery;
* hnsw_init_tmp_node(myIndex,&myQuery,0,some_vector);
* hnsw_distance(&myQuery, some_hnsw_node);
*
* Make sure to later free the node with:
*
* hnsw_free_tmp_node(&myQuery,some_vector);
* You have to pass the vector to the free function, because sometimes
* hnsw_init_tmp_node() may just avoid allocating a vector at all,
* just reusing 'some_vector' pointer.
*
* Return 0 on out of memory, 1 on success.
*/
int hnsw_init_tmp_node(HNSW *index, hnswNode *node, int is_normalized, const float *vector) {
node->vector = NULL;
/* Work on a normalized query vector if the input vector is
* not normalized. */
if (!is_normalized) {
node->vector = hmalloc(sizeof(float)*index->vector_dim);
if (node->vector == NULL) return 0;
memcpy(node->vector,vector,sizeof(float)*index->vector_dim);
hnsw_normalize_vector(node->vector,NULL,index->vector_dim);
} else {
node->vector = (float*)vector;
}
/* If quantization is enabled, our query fake node should be
* quantized as well. */
if (index->quant_type != HNSW_QUANT_NONE) {
void *quants = hmalloc(hnsw_quants_bytes(index));
if (quants == NULL) {
if (node->vector != vector) hfree(node->vector);
return 0;
}
switch(index->quant_type) {
case HNSW_QUANT_Q8:
quantize_to_q8(node->vector, quants, index->vector_dim, &node->quants_range);
break;
case HNSW_QUANT_BIN:
quantize_to_bin(node->vector, quants, index->vector_dim);
}
if (node->vector != vector) hfree(node->vector);
node->vector = quants;
}
return 1;
}
/* Free the stack allocated node initialized by hnsw_init_tmp_node(). */
void hnsw_free_tmp_node(hnswNode *node, const float *vector) {
if (node->vector != vector) hfree(node->vector);
}
/* Return approximated K-NN items. Note that neighbors and distances
* arrays must have space for at least 'k' items.
* norm_query should be set to 1 if the query vector is already
* normalized, otherwise, if 0, the function will copy the vector,
* L2-normalize the copy and search using the normalized version.
*
* If the filter_privdata callback is passed, only elements passing the
* specified filter (invoked with privdata and the value associated
* to the node as arguments) are returned. In such case, if max_candidates
* is not NULL, it represents the maximum number of nodes to explore, since
* the search may be otherwise unbound if few or no elements pass the
* filter. */
int hnsw_search_with_filter
(HNSW *index, const float *query_vector, uint32_t k,
hnswNode **neighbors, float *distances, uint32_t slot,
int query_vector_is_normalized,
int (*filter_callback)(void *value, void *privdata),
void *filter_privdata, uint32_t max_candidates)
{
if (!index || !query_vector || !neighbors || k == 0) return -1;
if (!index->enter_point) return 0; // Empty index.
/* Use a fake node that holds the query vector, this way we can
* use our normal node to node distance functions when checking
* the distance between query and graph nodes. */
hnswNode query;
if (hnsw_init_tmp_node(index,&query,query_vector_is_normalized,query_vector) == 0) return -1;
// Start searching from the entry point.
hnswNode *curr_ep = index->enter_point;
/* Start from higher layer to layer 1 (layer 0 is handled later)
* in the next section. Descend to the most similar node found
* so far. */
for (int lc = index->max_level; lc > 0; lc--) {
pqueue *results = search_layer(index, &query, curr_ep, 1, lc, slot);
if (!results) continue;
if (results->count > 0) {
curr_ep = pq_get_node(results,0);
}
pq_free(results);
}
/* Search bottom layer (the most densely populated) with ef = k */
pqueue *results = search_layer_with_filter(
index, &query, curr_ep, k, 0, slot, filter_callback,
filter_privdata, max_candidates);
if (!results) {
hnsw_free_tmp_node(&query, query_vector);
return -1;
}
/* Copy results. */
uint32_t found = MIN(k, results->count);
for (uint32_t i = 0; i < found; i++) {
neighbors[i] = pq_get_node(results,i);
if (distances) {
distances[i] = pq_get_distance(results,i);
}
}
pq_free(results);
hnsw_free_tmp_node(&query, query_vector);
return found;
}
/* Wrapper to hnsw_search_with_filter() when no filter is needed. */
int hnsw_search(HNSW *index, const float *query_vector, uint32_t k,
hnswNode **neighbors, float *distances, uint32_t slot,
int query_vector_is_normalized)
{
return hnsw_search_with_filter(index,query_vector,k,neighbors,
distances,slot,query_vector_is_normalized,
NULL,NULL,0);
}
/* Rescan a node and update the wortst neighbor index.
* The followinng two functions are variants of this function to be used
* when links are added or removed: they may do less work than a full scan. */
void hnsw_update_worst_neighbor(HNSW *index, hnswNode *node, uint32_t layer) {
float worst_dist = 0;
uint32_t worst_idx = 0;
for (uint32_t i = 0; i < node->layers[layer].num_links; i++) {
float dist = hnsw_distance(index, node, node->layers[layer].links[i]);
if (dist > worst_dist) {
worst_dist = dist;
worst_idx = i;
}
}
node->layers[layer].worst_distance = worst_dist;
node->layers[layer].worst_idx = worst_idx;
}
/* Update node worst neighbor distance information when a new neighbor
* is added. */
void hnsw_update_worst_neighbor_on_add(HNSW *index, hnswNode *node, uint32_t layer, uint32_t added_index, float distance) {
(void) index; // Unused but here for API symmetry.
if (node->layers[layer].num_links == 1 || // First neighbor?
distance > node->layers[layer].worst_distance) // New worst?
{
node->layers[layer].worst_distance = distance;
node->layers[layer].worst_idx = added_index;
}
}
/* Update node worst neighbor distance information when a linked neighbor
* is removed. */
void hnsw_update_worst_neighbor_on_remove(HNSW *index, hnswNode *node, uint32_t layer, uint32_t removed_idx)
{
if (node->layers[layer].num_links == 0) {
node->layers[layer].worst_distance = 0;
node->layers[layer].worst_idx = 0;
} else if (removed_idx == node->layers[layer].worst_idx) {
hnsw_update_worst_neighbor(index,node,layer);
} else if (removed_idx < node->layers[layer].worst_idx) {
// Just update index if we removed element before worst.
node->layers[layer].worst_idx--;
}
}
/* We have a list of candidate nodes to link to the new node, when iserting
* one. This function selects which nodes to link and performs the linking.
*
* Parameters:
*
* - 'candidates' is the priority queue of potential good nodes to link to the
* new node 'new_node'.
* - 'required_links' is as many links we would like our new_node to get
* at the specified layer.
* - 'aggressive' changes the startegy used to find good neighbors as follows:
*
* This function is called with aggressive=0 for all the layers, including
* layer 0. When called like that, it will use the diversity of links and
* quality of links checks before linking our new node with some candidate.
*
* However if the insert function finds that at layer 0, with aggressive=0,
* few connections were made, it calls this function again with agressiveness
* levels greater up to 2.
*
* At aggressive=1, the diversity checks are disabled, and the candidate
* node for linking is accepted even if it is nearest to an already accepted
* neighbor than it is to the new node.
*
* When we link our new node by replacing the link of a candidate neighbor
* that already has the max number of links, inevitably some other node loses
* a connection (to make space for our new node link). In this case:
*
* 1. If such "dropped" node would remain with too little links, we try with
* some different neighbor instead, however as the 'aggressive' paramter
* has incremental values (0, 1, 2) we are more and more willing to leave
* the dropped node with fever connections.
* 2. If aggressive=2, we will scan the candidate neighbor node links to
* find a different linked-node to replace, one better connected even if
* its distance is not the worse.
*
* Note: this function is also called during deletion of nodes in order to
* provide certain nodes with additional links.
*/
void select_neighbors(HNSW *index, pqueue *candidates, hnswNode *new_node,
uint32_t layer, uint32_t required_links, int aggressive)
{
uint32_t max_links = (layer == 0) ? HNSW_M0 : HNSW_M;
for (uint32_t i = 0; i < candidates->count; i++) {
hnswNode *neighbor = pq_get_node(candidates,i);
if (neighbor == new_node) continue; // Don't link node with itself.
/* Use our cached distance among the new node and the candidate. */
float dist = pq_get_distance(candidates,i);
/* First of all, since our links are all bidirectional, if the
* new node for any reason has no longer room, or if it accumulated
* the required number of links, return ASAP. */
if (new_node->layers[layer].num_links >= new_node->layers[layer].max_links ||
new_node->layers[layer].num_links >= required_links) return;
/* If aggressive is true, it is possible that the new node
* already got some link among the candidates (see the top comment,
* this function gets re-called in case of too few links).
* So we need to check if this candidate is already linked to
* the new node. */
if (aggressive) {
int duplicated = 0;
for (uint32_t j = 0; j < new_node->layers[layer].num_links; j++) {
if (new_node->layers[layer].links[j] == neighbor) {
duplicated = 1;
break;
}
}
if (duplicated) continue;
}
/* Diversity check. We accept new candidates
* only if there is no element already accepted that is nearest
* to the candidate than the new element itself.
* However this check is disabled if we have pressure to find
* new links (aggressive != 0) */
if (!aggressive) {
int diversity_failed = 0;
for (uint32_t j = 0; j < new_node->layers[layer].num_links; j++) {
float link_dist = hnsw_distance(index, neighbor,
new_node->layers[layer].links[j]);
if (link_dist < dist) {
diversity_failed = 1;
break;
}
}
if (diversity_failed) continue;
}
/* If potential neighbor node has space, simply add the new link.
* We will have space as well. */
uint32_t n = neighbor->layers[layer].num_links;
if (n < max_links) {
/* Link candidate to new node. */
neighbor->layers[layer].links[n] = new_node;
neighbor->layers[layer].num_links++;
/* Update candidate worst link info. */
hnsw_update_worst_neighbor_on_add(index,neighbor,layer,n,dist);
/* Link new node to candidate. */
uint32_t new_links = new_node->layers[layer].num_links;
new_node->layers[layer].links[new_links] = neighbor;
new_node->layers[layer].num_links++;
/* Update new node worst link info. */
hnsw_update_worst_neighbor_on_add(index,new_node,layer,new_links,dist);
continue;
}
/* ====================================================================
* Replacing existing candidate neighbor link step.
* ================================================================== */
/* If we are here, our accepted candidate for linking is full.
*
* If new node is more distant to candidate than its current worst link
* then we skip it: we would not be able to establish a bidirectional
* connection without compromising link quality of candidate.
*
* At aggressiveness > 0 we don't care about this check. */
if (!aggressive && dist >= neighbor->layers[layer].worst_distance)
continue;
/* We can add it: we are ready to replace the candidate neighbor worst
* link with the new node, assuming certain conditions are met. */
hnswNode *worst_node = neighbor->layers[layer].links[neighbor->layers[layer].worst_idx];
/* The worst node linked to our candidate may remain too disconnected
* if we remove the candidate node as its link. Let's check if
* this is the case: */
if (aggressive == 0 &&
worst_node->layers[layer].num_links <= HNSW_M/2)
continue;
/* Aggressive level = 1. It's ok if the node remains with just
* HNSW_M/4 links. */
else if (aggressive == 1 &&
worst_node->layers[layer].num_links <= HNSW_M/4)
continue;
/* If aggressive is set to 2, then the new node we are adding failed
* to find enough neighbors. We can't insert an almost orphaned new
* node, so let's see if the target node has some other link
* that is well connected in the graph: we could drop it instead
* of the worst link. */
if (aggressive == 2 && worst_node->layers[layer].num_links <= HNSW_M/4)
{
/* Let's see if we can find at least a candidate link that
* would remain with a few connections. Track the one
* that is the farest away (worst distance) from our candidate
* neighbor (in order to remove the less interesting link). */
worst_node = NULL;
uint32_t worst_idx = 0;
float max_dist = 0;
for (uint32_t j = 0; j < neighbor->layers[layer].num_links; j++) {
hnswNode *to_drop = neighbor->layers[layer].links[j];
/* Skip this if it would remain too disconnected as well.
*
* NOTE about HNSW_M/4 min connections requirement:
*
* It is not too strict, since leaving a node with just a
* single link does not just leave it too weakly connected, but
* also sometimes creates cycles with few disconnected
* nodes linked among them. */
if (to_drop->layers[layer].num_links <= HNSW_M/4) continue;
float link_dist = hnsw_distance(index, neighbor, to_drop);
if (worst_node == NULL || link_dist > max_dist) {
worst_node = to_drop;
max_dist = link_dist;
worst_idx = j;
}
}
if (worst_node != NULL) {
/* We found a node that we can drop. Let's pretend this is
* the worst node of the candidate to unify the following
* code path. Later we will fix the worst node info anyway. */
neighbor->layers[layer].worst_distance = max_dist;
neighbor->layers[layer].worst_idx = worst_idx;
} else {
/* Otherwise we have no other option than reallocating
* the max number of links for this target node, and
* ensure at least a few connections for our new node.
*
* XXX: Implement this part. */
debugmsg("Node overbooking needed: allocate more\n");
continue;
}
}
// Remove backlink from the worst node of our candidate.
for (uint64_t j = 0; j < worst_node->layers[layer].num_links; j++) {
if (worst_node->layers[layer].links[j] == neighbor) {
memmove(&worst_node->layers[layer].links[j],
&worst_node->layers[layer].links[j+1],
(worst_node->layers[layer].num_links - j - 1) * sizeof(hnswNode*));
worst_node->layers[layer].num_links--;
hnsw_update_worst_neighbor_on_remove(index,worst_node,layer,j);
break;
}
}
/* Replace worst link with the new node. */
neighbor->layers[layer].links[neighbor->layers[layer].worst_idx] = new_node;
/* Update the worst link in the target node, at this point
* the link that we replaced may no longer be the worst. */
hnsw_update_worst_neighbor(index,neighbor,layer);
// Add new node -> candidate link.
uint32_t new_links = new_node->layers[layer].num_links;
new_node->layers[layer].links[new_links] = neighbor;
new_node->layers[layer].num_links++;
// Update new node worst link.
hnsw_update_worst_neighbor_on_add(index,new_node,layer,new_links,dist);
}
}
/* This function implements node reconnection after a node deletion in HNSW.
* When a node is deleted, other nodes at the specified layer lose one
* connection (all the neighbors of the deleted node). This function attempts
* to pair such nodes together in a way that maximizes connection quality
* among the M nodes that were former neighbors of our deleted node.
*
* The algorithm works by first building a distance matrix among the nodes:
*
* N0 N1 N2 N3
* N0 0 1.2 0.4 0.9
* N1 1.2 0 0.8 0.5
* N2 0.4 0.8 0 1.1
* N3 0.9 0.5 1.1 0
*
* For each potential pairing (i,j) we compute a score that combines:
* 1. The direct cosine distance between the two nodes
* 2. The average distance to other nodes that would no longer be
* available for pairing if we select this pair
*
* We want to balance local node-to-node requirements and global requirements.
* For instance sometimes connecting A with B, while optimal, would leave
* C and D to be connected without other choices, and this could be a very
* bad connection. Maybe instead A and C and B and D are both relatively high
* quality connections.
*
* The formula used to calculate the score of each connection is:
*
* score[i,j] = W1*(2-distance[i,j]) + W2*((new_avg_i + new_avg_j)/2)
* where new_avg_x is the average of distances in row x excluding distance[i,j]
*
* So the score is directly proportional to the SIMILARITY of the two nodes
* and also directly proportional to the DISTANCE of the potential other
* connections that we lost by pairign i,j. So we have a cost for missed
* opportunities, or better, in this case, a reward if the missing
* opportunities are not so good (big average distance).
*
* W1 and W2 are weights (defaults: 0.7 and 0.3) that determine the relative
* importance of immediate connection quality vs future pairing potential.
*
* After the initial pairing phase, any nodes that couldn't be paired
* (due to odd count or existing connections) are handled by searching
* the broader graph using the standard HNSW neighbor selection logic.
*/
void hnsw_reconnect_nodes(HNSW *index, hnswNode **nodes, int count, uint32_t layer) {
if (count <= 0) return;
debugmsg("Reconnecting %d nodes\n", count);
/* Step 1: Build the distance matrix between all nodes.
* Since distance(i,j) = distance(j,i), we only compute the upper triangle
* and mirror it to the lower triangle. */
float *distances = hmalloc(count * count * sizeof(float));
if (!distances) return;
for (int i = 0; i < count; i++) {
distances[i*count + i] = 0; // Distance to self is 0
for (int j = i+1; j < count; j++) {
float dist = hnsw_distance(index, nodes[i], nodes[j]);
distances[i*count + j] = dist; // Upper triangle.
distances[j*count + i] = dist; // Lower triangle.
}
}
/* Step 2: Calculate row averages (will be used in scoring):
* please note that we just calculate row averages and not
* colums averages since the matrix is symmetrical, so those
* are the same: check the image in the top comment if you have any
* doubt about this. */
float *row_avgs = hmalloc(count * sizeof(float));
if (!row_avgs) {
hfree(distances);
return;
}
for (int i = 0; i < count; i++) {
float sum = 0;
int valid_count = 0;
for (int j = 0; j < count; j++) {
if (i != j) {
sum += distances[i*count + j];
valid_count++;
}
}
row_avgs[i] = valid_count ? sum / valid_count : 0;
}
/* Step 3: Build scoring matrix. What we do here is to combine how
* good is a given i,j nodes connection, with how badly connecting
* i,j will affect the remaining quality of connections left to
* pair the other nodes. */
float *scores = hmalloc(count * count * sizeof(float));
if (!scores) {
hfree(distances);
hfree(row_avgs);
return;
}
/* Those weights were obtained manually... No guarantee that they
* are optimal. However with these values the algorithm is certain
* better than its greedy version that just attempts to pick the
* best pair each time (verified experimentally). */
const float W1 = 0.7; // Weight for immediate distance.
const float W2 = 0.3; // Weight for future potential.
for (int i = 0; i < count; i++) {
for (int j = 0; j < count; j++) {
if (i == j) {
scores[i*count + j] = -1; // Invalid pairing.
continue;
}
// Check for existing connection between i and j.
int already_linked = 0;
for (uint32_t k = 0; k < nodes[i]->layers[layer].num_links; k++)
{
if (nodes[i]->layers[layer].links[k] == nodes[j]) {
scores[i*count + j] = -1; // Already linked.
already_linked = 1;
break;
}
}
if (already_linked) continue;
float dist = distances[i*count + j];
/* Calculate new averages excluding this pair.
* Handle edge case where we might have too few elements.
* Note that it would be not very smart to recompute the average
* each time scanning the row, we can remove the element
* and adjust the average without it. */
float new_avg_i = 0, new_avg_j = 0;
if (count > 2) {
new_avg_i = (row_avgs[i] * (count-1) - dist) / (count-2);
new_avg_j = (row_avgs[j] * (count-1) - dist) / (count-2);
}
/* Final weighted score: the more similar i,j, the better
* the score. The more distant are the pairs we lose by
* connecting i,j, the better the score. */
scores[i*count + j] = W1*(2-dist) + W2*((new_avg_i + new_avg_j)/2);
}
}
// Step 5: Pair nodes greedily based on scores.
int *used = hmalloc(count*sizeof(int));
memset(used,0,count*sizeof(int));
if (!used) {
hfree(distances);
hfree(row_avgs);
hfree(scores);
return;
}
/* Scan the matrix looking each time for the potential
* link with the best score. */
while(1) {
float max_score = -1;
int best_j = -1, best_i = -1;
// Seek best score i,j values.
for (int i = 0; i < count; i++) {
if (used[i]) continue; // Already connected.
/* No space left? Not possible after a node deletion but makes
* this function more future-proof. */
if (nodes[i]->layers[layer].num_links >=
nodes[i]->layers[layer].max_links) continue;
for (int j = 0; j < count; j++) {
if (i == j) continue; // Same node, skip.
if (used[j]) continue; // Already connected.
float score = scores[i*count + j];
if (score < 0) continue; // Invalid link.
/* If the target node has space, and its score is better
* than any other seen so far... remember it is the best. */
if (score > max_score &&
nodes[j]->layers[layer].num_links <
nodes[j]->layers[layer].max_links)
{
// Track the best connection found so far.
max_score = score;
best_j = j;
best_i = i;
}
}
}
// Possible link found? Connect i and j.
if (best_j != -1) {
debugmsg("[%d] linking %d with %d: %f\n", layer, (int)best_i, (int)best_j, max_score);
// Link i -> j.
int link_idx = nodes[best_i]->layers[layer].num_links;
nodes[best_i]->layers[layer].links[link_idx] = nodes[best_j];
nodes[best_i]->layers[layer].num_links++;
// Update worst distance if needed.
float dist = distances[best_i*count + best_j];
hnsw_update_worst_neighbor_on_add(index,nodes[best_i],layer,link_idx,dist);
// Link j -> i.
link_idx = nodes[best_j]->layers[layer].num_links;
nodes[best_j]->layers[layer].links[link_idx] = nodes[best_i];
nodes[best_j]->layers[layer].num_links++;
// Update worst distance if needed.
hnsw_update_worst_neighbor_on_add(index,nodes[best_j],layer,link_idx,dist);
// Mark connection as used.
used[best_i] = used[best_j] = 1;
} else {
break; // No more valid connections available.
}
}
/* Step 6: Handle remaining unpaired nodes using the standard HNSW
* neighbor selection. */
for (int i = 0; i < count; i++) {
if (used[i]) continue;
// Skip if node is already at max connections.
if (nodes[i]->layers[layer].num_links >=
nodes[i]->layers[layer].max_links)
continue;
debugmsg("[%d] Force linking %d\n", layer, i);
/* First, try with local nodes as candidates.
* Some candidate may have space. */
pqueue *candidates = pq_new(count);
if (!candidates) continue;
/* Add all the local nodes having some space as candidates
* to be linked with this node. */
for (int j = 0; j < count; j++) {
if (i != j && // Must not be itself.
nodes[j]->layers[layer].num_links < // Must not be full.
nodes[j]->layers[layer].max_links)
{
float dist = distances[i*count + j];
pq_push(candidates, nodes[j], dist);
}
}
/* Try local candidates first with aggressive = 1.
* So we will link only if there is space.
* We want one link more than the links we already have. */
uint32_t wanted_links = nodes[i]->layers[layer].num_links+1;
if (candidates->count > 0) {
select_neighbors(index, candidates, nodes[i], layer,
wanted_links, 1);
debugmsg("Final links after attempt with local nodes: %d (wanted: %d)\n", (int)nodes[i]->layers[layer].num_links, wanted_links);
}
// If still no connection, search the broader graph.
if (nodes[i]->layers[layer].num_links != wanted_links) {
debugmsg("No force linking possible with local candidats\n");
pq_free(candidates);
// Find entry point for target layer by descending through levels.
hnswNode *curr_ep = index->enter_point;
for (uint32_t lc = index->max_level; lc > layer; lc--) {
pqueue *results = search_layer(index, nodes[i], curr_ep, 1, lc, 0);
if (results) {
if (results->count > 0) {
curr_ep = pq_get_node(results,0);
}
pq_free(results);
}
}
if (curr_ep) {
/* Search this layer for candidates.
* Use the defalt EF_C in this case, since it's not an
* "insert" operation, and we don't know the user
* specified "EF". */
candidates = search_layer(index, nodes[i], curr_ep, HNSW_EF_C, layer, 0);
if (candidates) {
/* Try to connect with aggressiveness proportional to the
* node linking condition. */
int aggressiveness =
nodes[i]->layers[layer].num_links > HNSW_M / 2 ? 1 : 2;
select_neighbors(index, candidates, nodes[i], layer, wanted_links, aggressiveness);
debugmsg("Final links with broader search: %d (wanted: %d)\n", (int)nodes[i]->layers[layer].num_links, wanted_links);
pq_free(candidates);
}
}
} else {
pq_free(candidates);
}
}
// Cleanup.
hfree(distances);
hfree(row_avgs);
hfree(scores);
hfree(used);
}
/* This is an helper function in order to support node deletion.
* It's goal is just to:
*
* 1. Remove the node from the bidirectional links of neighbors in the graph.
* 2. Remove the node from the linked list of nodes.
* 3. Fix the entry point in the graph. We just select one of the neighbors
* of the deleted node at a lower level. If none is found, we do
* a full scan.
* 4. The node itself amd its aux value field are NOT freed. It's up to the
* caller to do it, by using hnsw_node_free().
* 5. The node associated value (node->value) is NOT freed.
*
* Why this function will not free the node? Because in node updates it
* could be a good idea to reuse the node allocation for different reasons
* (currently not implemented).
* In general it is more future-proof to be able to reuse the node if
* needed. Right now this library reuses the node only when links are
* not touched (see hnsw_update() for more information). */
void hnsw_unlink_node(HNSW *index, hnswNode *node) {
if (!index || !node) return;
index->version++; // This node may be missing in an already compiled list
// of neighbors. Make optimistic concurrent inserts fail.
/* Remove all bidirectional links at each level.
* Note that in this implementation all the
* links are guaranteed to be bedirectional. */
/* For each level of the deleted node... */
for (uint32_t level = 0; level <= node->level; level++) {
/* For each linked node of the deleted node... */
for (uint32_t i = 0; i < node->layers[level].num_links; i++) {
hnswNode *linked = node->layers[level].links[i];
/* Find and remove the backlink in the linked node */
for (uint32_t j = 0; j < linked->layers[level].num_links; j++) {
if (linked->layers[level].links[j] == node) {
/* Remove by shifting remaining links left */
memmove(&linked->layers[level].links[j],
&linked->layers[level].links[j + 1],
(linked->layers[level].num_links - j - 1) * sizeof(hnswNode*));
linked->layers[level].num_links--;
hnsw_update_worst_neighbor_on_remove(index,linked,level,j);
break;
}
}
}
}
/* Update cursors pointing at this element. */
if (index->cursors) hnsw_cursor_element_deleted(index,node);
/* Update the previous node's next pointer. */
if (node->prev) {
node->prev->next = node->next;
} else {
/* If there's no previous node, this is the head. */
index->head = node->next;
}
/* Update the next node's prev pointer. */
if (node->next) node->next->prev = node->prev;
/* Update node count. */
index->node_count--;
/* If this node was the enter_point, we need to update it. */
if (node == index->enter_point) {
/* Reset entry point - we'll find a new one (unless the HNSW is
* now empty) */
index->enter_point = NULL;
index->max_level = 0;
/* Step 1: Try to find a replacement by scanning levels
* from top to bottom. Under normal conditions, if there is
* any other node at the same level, we have a link. Anyway
* we descend levels to find any neighbor at the higher level
* possible. */
for (int level = node->level; level >= 0; level--) {
if (node->layers[level].num_links > 0) {
index->enter_point = node->layers[level].links[0];
break;
}
}
/* Step 2: If no links were found at any level, do a full scan.
* This should never happen in practice if the HNSW is not
* empty. */
if (!index->enter_point) {
uint32_t new_max_level = 0;
hnswNode *current = index->head;
while (current) {
if (current != node && current->level >= new_max_level) {
new_max_level = current->level;
index->enter_point = current;
}
current = current->next;
}
}
/* Update max_level. */
if (index->enter_point)
index->max_level = index->enter_point->level;
}
/* Clear the node's links but don't free the node itself */
node->prev = node->next = NULL;
}
/* Higher level API for hnsw_unlink_node() + hnsw_reconnect_nodes() actual work.
* This will get the write lock, will delete the node, free it,
* reconnect the node neighbors among themselves, and unlock again.
* If free_value function pointer is not NULL, then the function provided is
* used to free node->value. */
void hnsw_delete_node(HNSW *index, hnswNode *node, void(*free_value)(void*value)) {
pthread_rwlock_wrlock(&index->global_lock);
hnsw_unlink_node(index,node);
if (free_value && node->value) free_value(node->value);
/* Relink all the nodes orphaned of this node link.
* Do it for all the levels. */
for (unsigned int j = 0; j <= node->level; j++) {
hnsw_reconnect_nodes(index, node->layers[j].links,
node->layers[j].num_links, j);
}
hnsw_node_free(node);
pthread_rwlock_unlock(&index->global_lock);
}
/* ============================ Threaded API ================================
* Concurent readers should use the following API to get a slot assigned
* (and a lock, too), do their read-only call, and unlock the slot.
*
* There is a reason why read operations don't implement opaque transparent
* locking directly on behalf of the user: when we return a result set
* with hnsw_search(), we report a set of nodes. The caller will do something
* with the nodes and the associated values, so the unlocking of the
* slot should happen AFTER the result was already used, otherwise we may
* have changes to the HNSW nodes as the result is being accessed. */
/* Try to acquire a read slot. Returns the slot number (0 to HNSW_MAX_THREADS-1)
* on success, -1 on error (pthread mutex errors). */
int hnsw_acquire_read_slot(HNSW *index) {
/* First try a non-blocking approach on all slots. */
for (uint32_t i = 0; i < HNSW_MAX_THREADS; i++) {
if (pthread_mutex_trylock(&index->slot_locks[i]) == 0) {
if (pthread_rwlock_rdlock(&index->global_lock) != 0) {
pthread_mutex_unlock(&index->slot_locks[i]);
return -1;
}
return i;
}
}
/* All trylock attempts failed, use atomic increment to select slot. */
uint32_t slot = index->next_slot++ % HNSW_MAX_THREADS;
/* Try to lock the selected slot. */
if (pthread_mutex_lock(&index->slot_locks[slot]) != 0) return -1;
/* Get read lock. */
if (pthread_rwlock_rdlock(&index->global_lock) != 0) {
pthread_mutex_unlock(&index->slot_locks[slot]);
return -1;
}
return slot;
}
/* Release a previously acquired read slot: note that it is important that
* nodes returned by hnsw_search() are accessed while the read lock is
* still active, to be sure that nodes are not freed. */
void hnsw_release_read_slot(HNSW *index, int slot) {
if (slot < 0 || slot >= HNSW_MAX_THREADS) return;
pthread_rwlock_unlock(&index->global_lock);
pthread_mutex_unlock(&index->slot_locks[slot]);
}
/* ============================ Nodes insertion =============================
* We have an optimistic API separating the read-only candidates search
* and the write side (actual node insertion). We internally also use
* this API to provide the plain hnsw_insert() function for code unification. */
struct InsertContext {
pqueue *level_queues[HNSW_MAX_LEVEL]; /* Candidates for each level. */
hnswNode *node; /* Pre-allocated node ready for insertion */
uint64_t version; /* Index version at preparation time. This is used
* for CAS-like locking during change commit. */
};
/* Optimistic insertion API.
*
* WARNING: Note that this is an internal function: users should call
* hnsw_prepare_insert() instead.
*
* This is how it works: you use hnsw_prepare_insert() and it will return
* a context where good candidate neighbors are already pre-selected.
* This step only uses read locks.
*
* Then finally you try to actually commit the new node with
* hnsw_try_commit_insert(): this time we will require a write lock, but
* for less time than it would be otherwise needed if using directly
* hnsw_insert(). When you try to commit the write, if no node was deleted in
* the meantime, your operation will succeed, otherwise it will fail, and
* you should try to just use the hnsw_insert() API, since there is
* contention.
*
* See hnsw_node_new() for information about 'vector' and 'qvector'
* arguments, and which one to pass. */
InsertContext *hnsw_prepare_insert_nolock(HNSW *index, const float *vector,
const int8_t *qvector, float qrange, uint64_t id, void *value,
int slot, int ef)
{
InsertContext *ctx = hmalloc(sizeof(*ctx));
if (!ctx) return NULL;
memset(ctx, 0, sizeof(*ctx));
ctx->version = index->version;
/* Crete a new node that we may be able to insert into the
* graph later, when calling the commit function. */
uint32_t level = random_level();
ctx->node = hnsw_node_new(index, id, vector, qvector, qrange, level, 1);
if (!ctx->node) {
hfree(ctx);
return NULL;
}
ctx->node->value = value;
hnswNode *curr_ep = index->enter_point;
/* Empty graph, no need to collect candidates. */
if (curr_ep == NULL) return ctx;
/* Phase 1: Find good entry point on the highest level of the new
* node we are going to insert. */
for (unsigned int lc = index->max_level; lc > level; lc--) {
pqueue *results = search_layer(index, ctx->node, curr_ep, 1, lc, slot);
if (results) {
if (results->count > 0) curr_ep = pq_get_node(results,0);
pq_free(results);
}
}
/* Phase 2: Collect a set of potential connections for each layer of
* the new node. */
for (int lc = MIN(level, index->max_level); lc >= 0; lc--) {
pqueue *candidates =
search_layer(index, ctx->node, curr_ep, ef, lc, slot);
if (!candidates) continue;
curr_ep = (candidates->count > 0) ? pq_get_node(candidates,0) : curr_ep;
ctx->level_queues[lc] = candidates;
}
return ctx;
}
/* External API for hnsw_prepare_insert_nolock(), handling locking. */
InsertContext *hnsw_prepare_insert(HNSW *index, const float *vector,
const int8_t *qvector, float qrange, uint64_t id, void *value,
int ef)
{
InsertContext *ctx;
int slot = hnsw_acquire_read_slot(index);
ctx = hnsw_prepare_insert_nolock(index,vector,qvector,qrange,id,value,slot,ef);
hnsw_release_read_slot(index,slot);
return ctx;
}
/* Free an insert context and all its resources. */
void hnsw_free_insert_context(InsertContext *ctx) {
if (!ctx) return;
for (uint32_t i = 0; i < HNSW_MAX_LEVEL; i++) {
if (ctx->level_queues[i]) pq_free(ctx->level_queues[i]);
}
if (ctx->node) hnsw_node_free(ctx->node);
hfree(ctx);
}
/* Commit a prepared insert operation. This function is a low level API that
* should not be called by the user. See instead hnsw_try_commit_insert(), that
* will perform the CAS check and acquire the write lock.
*
* See the top comment in hnsw_prepare_insert() for more information
* on the optimistic insertion API.
*
* This function can't fail and always returns the pointer to the
* just inserted node. Out of memory is not possible since no critical
* allocation is never performed in this code path: we populate links
* on already allocated nodes. */
hnswNode *hnsw_commit_insert_nolock(HNSW *index, InsertContext *ctx) {
hnswNode *node = ctx->node;
/* Handle first node case. */
if (index->enter_point == NULL) {
index->version++; // First node, make concurrent inserts fail.
index->enter_point = node;
index->max_level = node->level;
hnsw_add_node(index, node);
ctx->node = NULL; // So hnsw_free_insert_context() will not free it.
hnsw_free_insert_context(ctx);
return node;
}
/* Connect the node with near neighbors at each level. */
for (int lc = MIN(node->level,index->max_level); lc >= 0; lc--) {
if (ctx->level_queues[lc] == NULL) continue;
/* Try to provide HNSW_M connections to our node. The call
* is not guaranteed to be able to provide all the links we would
* like to have for the new node: they must be bi-directional, obey
* certain quality checks, and so forth, so later there are further
* calls to force the hand a bit if needed.
*
* Let's start with aggressiveness = 0. */
select_neighbors(index, ctx->level_queues[lc], node, lc, HNSW_M, 0);
/* Layer 0 and too few connections? Let's be more aggressive. */
if (lc == 0 && node->layers[0].num_links < HNSW_M/2) {
select_neighbors(index, ctx->level_queues[lc], node, lc,
HNSW_M, 1);
/* Still too few connections? Let's go to
* aggressiveness level '2' in linking strategy. */
if (node->layers[0].num_links < HNSW_M/4) {
select_neighbors(index, ctx->level_queues[lc], node, lc,
HNSW_M/4, 2);
}
}
}
/* If new node level is higher than current max, update entry point. */
if (node->level > index->max_level) {
index->version++; // Entry point changed, make concurrent inserts fail.
index->enter_point = node;
index->max_level = node->level;
}
/* Add node to the linked list. */
hnsw_add_node(index, node);
ctx->node = NULL; // So hnsw_free_insert_context() will not free the node.
hnsw_free_insert_context(ctx);
return node;
}
/* If the context obtained with hnsw_prepare_insert() is still valid
* (nodes not deleted in the meantime) then add the new node to the HNSW
* index and return its pointer. Otherwise NULL is returned and the operation
* should be either performed with the blocking API hnsw_insert() or attempted
* again. */
hnswNode *hnsw_try_commit_insert(HNSW *index, InsertContext *ctx) {
/* Check if the version changed since preparation. Note that we
* should access index->version under the write lock in order to
* be sure we can safely commit the write: this is just a fast-path
* in order to return ASAP without acquiring the write lock in case
* the version changed. */
if (ctx->version != index->version) {
hnsw_free_insert_context(ctx);
return NULL;
}
/* Try to acquire write lock. */
if (pthread_rwlock_wrlock(&index->global_lock) != 0) {
hnsw_free_insert_context(ctx);
return NULL;
}
/* Check version again under write lock. */
if (ctx->version != index->version) {
pthread_rwlock_unlock(&index->global_lock);
hnsw_free_insert_context(ctx);
return NULL;
}
/* Commit the change: note that it's up to hnsw_commit_insert_nolock()
* to free the insertion context. */
hnswNode *node = hnsw_commit_insert_nolock(index, ctx);
/* Release the write lock. */
pthread_rwlock_unlock(&index->global_lock);
return node;
}
/* Insert a new element into the graph.
* See hnsw_node_new() for information about 'vector' and 'qvector'
* arguments, and which one to pass.
*
* Return NULL on out of memory during insert. Otherwise the newly
* inserted node pointer is returned. */
hnswNode *hnsw_insert(HNSW *index, const float *vector, const int8_t *qvector, float qrange, uint64_t id, void *value, int ef) {
/* Write lock. We acquire the write lock even for the prepare()
* operation (that is a read-only operation) since we want this function
* to don't fail in the check-and-set stage of commit().
*
* Basically here we are using the optimistic API in a non-optimistinc
* way in order to have a single insertion code in the implementation. */
if (pthread_rwlock_wrlock(&index->global_lock) != 0) return NULL;
// Prepare the insertion - note we pass slot 0 since we're single threaded.
InsertContext *ctx = hnsw_prepare_insert_nolock(index, vector, qvector,
qrange, id, value, 0, ef);
if (!ctx) {
pthread_rwlock_unlock(&index->global_lock);
return NULL;
}
// Commit the prepared insertion without version checking.
hnswNode *node = hnsw_commit_insert_nolock(index, ctx);
// Release write lock and return our node pointer.
pthread_rwlock_unlock(&index->global_lock);
return node;
}
/* Helper function for qsort call in hnsw_should_reuse_node(). */
static int compare_floats(const float *a, const float *b) {
if (*a < *b) return 1;
if (*a > *b) return -1;
return 0;
}
/* This function determines if a node can be reused with a new vector by:
*
* 1. Computing average of worst 25% of current distances.
* 2. Checking if at least 50% of new distances stay below this threshold.
* 3. Requiring a minimum number of links for the check to be meaningful.
*
* This check is useful when we want to just update a node that already
* exists in the graph. Often the new vector is a learned embedding generated
* by some model, and the embedding represents some document that perhaps
* changed just slightly compared to the past, so the new embedding will
* be very nearby. We need to find a way do determine if the current node
* neighbors (practically speaking its location in the grapb) are good
* enough even with the new vector.
*
* XXX: this function needs improvements: successive updates to the same
* node with more and more distant vectors will make the node drift away
* from its neighbors. One of the additional metrics used could be
* neighbor-to-neighbor distance, that represents a more absolute check
* of fit for the new vector. */
int hnsw_should_reuse_node(HNSW *index, hnswNode *node, int is_normalized, const float *new_vector) {
/* Step 1: Not enough links? Advice to avoid reuse. */
const uint32_t min_links_for_reuse = 4;
uint32_t layer0_connections = node->layers[0].num_links;
if (layer0_connections < min_links_for_reuse) return 0;
/* Step2: get all current distances and run our heuristic. */
float *old_distances = hmalloc(sizeof(float) * layer0_connections);
if (!old_distances) return 0;
// Temporary node with the new vector, to simplify the next logic.
hnswNode tmp_node;
if (hnsw_init_tmp_node(index,&tmp_node,is_normalized,new_vector) == 0) {
hfree(old_distances);
return 0;
}
/* Get old dinstances and sort them to access the 25% worst
* (bigger) ones. */
for (uint32_t i = 0; i < layer0_connections; i++) {
old_distances[i] = hnsw_distance(index, node, node->layers[0].links[i]);
}
qsort(old_distances, layer0_connections, sizeof(float),
(int (*)(const void*, const void*))(&compare_floats));
uint32_t count = (layer0_connections+3)/4; // 25% approx to larger int.
if (count > layer0_connections) count = layer0_connections; // Futureproof.
float worst_avg = 0;
// Compute average of 25% worst dinstances.
for (uint32_t i = 0; i < count; i++) worst_avg += old_distances[i];
worst_avg /= count;
hfree(old_distances);
// Count how many new distances stay below the threshold.
uint32_t good_distances = 0;
for (uint32_t i = 0; i < layer0_connections; i++) {
float new_dist = hnsw_distance(index, &tmp_node, node->layers[0].links[i]);
if (new_dist <= worst_avg) good_distances++;
}
hnsw_free_tmp_node(&tmp_node,new_vector);
/* At least 50% of the nodes should pass our quality test, for the
* node to be reused. */
return good_distances >= layer0_connections/2;
}
/* ============================= Serialization ==============================
*
* TO SERIALIZE
* ============
*
* To serialize on disk, you need to persist the vector dimension, number
* of elements, and the quantization type index->quant_type. These are
* global values for the whole index.
*
* Then, to serialize each node:
*
* call hnsw_serialize_node() with each node you find in the linked list
* of nodes, starting at index->head (each node has a next pointer).
* The function will return an hnswSerNode structure, you will need
* to store the following on disk (for each node):
*
* - The sernode->vector data, that is sernode->vector_size bytes.
* - The sernode->params array, that points to an array of uint64_t
* integers. There are sernode->params_count total items. These
* parameters contain everything there is to need about your node: how
* many levels it has, its ID, the list of neighbors for each level (as node
* IDs), and so forth.
*
* You need to to save your own node->value in some way as well, but it already
* belongs to the user of the API, since, for this library, it's just a pointer,
* so the user should know how to serialized its private data.
*
* RELOADING FROM DISK / NET
* =========================
*
* When reloading nodes, you first load the index vector dimension and
* quantization type, and create the index with:
*
* HNSW *hnsw_new(uint32_t vector_dim, uint32_t quant_type);
*
* Then you load back, for each node (you stored how many nodes you had)
* the vector and the params array / count.
* You also load the value associated with your node.
*
* At this point you add back the loaded elements into the index with:
*
* hnsw_insert_serialized(HNSW *index, void *vector, uint64_t params,
* uint32_t params_len, void *value);
*
* Once you added all the nodes back, you need to resolve the pointers
* (since so far they are added just with the node IDs as reference), so
* you call:
*
* hnsw_deserialize_index(index);
*
* The index is now ready to be used like if it has been always in memory.
*
* DESIGN NOTES
* ============
*
* Why this API does not just give you a binary blob to save? Because in
* many systems (and in Redis itself) to save integers / floats can have
* more interesting encodings that just storing a 64 bit value. Many vector
* indexes will be small, and their IDs will be small numbers, so the storage
* system can exploit that and use less disk space, less network bandwidth
* and so forth.
*
* How is the data stored in these arrays of numbers? Oh well, we have
* things that are obviously numbers like node ID, number of levels for the
* node and so forth. Also each of our nodes have an unique incremental ID,
* so we can store a node set of links in terms of linked node IDs. This
* data is put directly in the loaded node pointer space! We just cast the
* integer to the pointer (so THIS IS NOT SAFE for 32 bit systems). Then
* we want to translate such IDs into pointers. To do that, we build an
* hash table, then scan all the nodes again and fix all the links converting
* the ID to the pointer. */
/* Return the serialized node information as specified in the top comment
* above. Note that the returned information is true as long as the node
* provided is not deleted or modified, so this function should be called
* when there are no concurrent writes.
*
* The function hnsw_serialize_node() should be called in order to
* free the result of this function. */
hnswSerNode *hnsw_serialize_node(HNSW *index, hnswNode *node) {
/* The first step is calculating the number of uint64_t parameters
* that we need in order to serialize the node. */
uint32_t num_params = 0;
num_params += 2; // node ID, number of layers.
for (uint32_t i = 0; i <= node->level; i++) {
num_params += 2; // max_links and num_links info for this layer.
num_params += node->layers[i].num_links; // The IDs of linked nodes.
}
/* We use another 64bit value to store two floats that are about
* the vector: l2 and quantization range (that is only used if the
* vector is quantized). */
num_params++;
/* Allocate the return object and the parameters array. */
hnswSerNode *sn = hmalloc(sizeof(hnswSerNode));
if (sn == NULL) return NULL;
sn->params = hmalloc(sizeof(uint64_t)*num_params);
if (sn->params == NULL) {
hfree(sn);
return NULL;
}
/* Fill data. */
sn->params_count = num_params;
sn->vector = node->vector;
sn->vector_size = hnsw_quants_bytes(index);
uint32_t param_idx = 0;
sn->params[param_idx++] = node->id;
sn->params[param_idx++] = node->level;
for (uint32_t i = 0; i <= node->level; i++) {
sn->params[param_idx++] = node->layers[i].num_links;
sn->params[param_idx++] = node->layers[i].max_links;
for (uint32_t j = 0; j < node->layers[i].num_links; j++) {
sn->params[param_idx++] = node->layers[i].links[j]->id;
}
}
uint64_t l2_and_range = 0;
unsigned char *aux = (unsigned char*)&l2_and_range;
memcpy(aux,&node->l2,sizeof(float));
memcpy(aux+4,&node->quants_range,sizeof(float));
sn->params[param_idx++] = l2_and_range;
/* Better safe than sorry: */
assert(param_idx == num_params);
return sn;
}
/* This is needed in order to free hnsw_serialize_node() returned
* structure. */
void hnsw_free_serialized_node(hnswSerNode *sn) {
hfree(sn->params);
hfree(sn);
}
/* Load a serialized node. See the top comment in this section of code
* for the documentation about how to use this.
*
* The function returns NULL both on out of memory and if the remaining
* parameters length does not match the number of links or other items
* to load. */
hnswNode *hnsw_insert_serialized(HNSW *index, void *vector, uint64_t *params, uint32_t params_len, void *value)
{
if (params_len < 2) return NULL;
uint64_t id = params[0];
uint32_t level = params[1];
/* Keep track of maximum ID seen while loading. */
if (id >= index->last_id) index->last_id = id;
/* Create node, passing vector data directly based on quantization type. */
hnswNode *node;
if (index->quant_type != HNSW_QUANT_NONE) {
node = hnsw_node_new(index, id, NULL, vector, 0, level, 0);
} else {
node = hnsw_node_new(index, id, vector, NULL, 0, level, 0);
}
if (!node) return NULL;
/* Load params array into the node. */
uint32_t param_idx = 2;
for (uint32_t i = 0; i <= level; i++) {
/* Sanity check. */
if (param_idx + 2 > params_len) {
hnsw_node_free(node);
return NULL;
}
uint32_t num_links = params[param_idx++];
uint32_t max_links = params[param_idx++];
/* If max_links is larger than current allocation, reallocate.
* It could happen in select_neighbors() that we over-allocate the
* node under very unlikely to happen conditions. */
if (max_links > node->layers[i].max_links) {
hnswNode **new_links = hrealloc(node->layers[i].links,
sizeof(hnswNode*) * max_links);
if (!new_links) {
hnsw_node_free(node);
return NULL;
}
node->layers[i].links = new_links;
node->layers[i].max_links = max_links;
}
node->layers[i].num_links = num_links;
/* Sanity check. */
if (param_idx + num_links > params_len) {
hnsw_node_free(node);
return NULL;
}
/* Fill links for this layer with the IDs. Note that this
* is going to not work in 32 bit systems. Deleting / adding-back
* nodes can produce IDs larger than 2^32-1 even if we can't never
* fit more than 2^32 nodes in a 32 bit system. */
for (uint32_t j = 0; j < num_links; j++)
node->layers[i].links[j] = (hnswNode*)params[param_idx++];
}
/* Get l2 and quantization range. */
if (param_idx >= params_len) {
hnsw_node_free(node);
return NULL;
}
uint64_t l2_and_range = params[param_idx];
unsigned char *aux = (unsigned char*)&l2_and_range;
memcpy(&node->l2, aux, sizeof(float));
memcpy(&node->quants_range, aux+4, sizeof(float));
node->value = value;
hnsw_add_node(index, node);
/* Keep track of higher node level and set the entry point to the
* greatest level node seen so far: thanks to this check we don't
* need to remember what our entry point was during serialization. */
if (index->enter_point == NULL || level > index->max_level) {
index->max_level = level;
index->enter_point = node;
}
return node;
}
/* Integer hashing, used by hnsw_deserialize_index().
* MurmurHash3's 64-bit finalizer function. */
uint64_t hnsw_hash_node_id(uint64_t id) {
id ^= id >> 33;
id *= 0xff51afd7ed558ccd;
id ^= id >> 33;
id *= 0xc4ceb9fe1a85ec53;
id ^= id >> 33;
return id;
}
/* Fix pointers of neighbors nodes: after loading the serialized nodes, the
* neighbors links are just IDs (casted to pointers), instead of the actual
* pointers. We need to resolve IDs into pointers.
*
* Return 0 on error (out of memory or some ID that can't be resolved), 1 on
* success. */
int hnsw_deserialize_index(HNSW *index) {
/* We will use simple linear probing, so over-allocating is a good
* idea: anyway this flat array of pointers will consume a fraction
* of the memory of the loaded index. */
uint64_t min_size = index->node_count*2;
uint64_t table_size = 1;
while(table_size < min_size) table_size <<= 1;
hnswNode **table = hmalloc(sizeof(hnswNode*) * table_size);
if (table == NULL) return 0;
memset(table,0,sizeof(hnswNode*) * table_size);
/* First pass: populate the ID -> pointer hash table. */
hnswNode *node = index->head;
while(node) {
uint64_t bucket = hnsw_hash_node_id(node->id) & (table_size-1);
for (uint64_t j = 0; j < table_size; j++) {
if (table[bucket] == NULL) {
table[bucket] = node;
break;
}
bucket = (bucket+1) & (table_size-1);
}
node = node->next;
}
/* Second pass: fix pointers of all the neighbors links. */
node = index->head; // Rewind.
while(node) {
for (uint32_t i = 0; i <= node->level; i++) {
for (uint32_t j = 0; j < node->layers[i].num_links; j++) {
uint64_t linked_id = (uint64_t) node->layers[i].links[j];
uint64_t bucket = hnsw_hash_node_id(linked_id) & (table_size-1);
hnswNode *neighbor = NULL;
for (uint64_t k = 0; k < table_size; k++) {
if (table[bucket] && table[bucket]->id == linked_id) {
neighbor = table[bucket];
break;
}
bucket = (bucket+1) & (table_size-1);
}
if (neighbor == NULL) {
/* Unresolved link! Either a bug in this code
* or broken serialization data. */
hfree(table);
return 0;
}
node->layers[i].links[j] = neighbor;
}
}
node = node->next;
}
hfree(table);
return 1;
}
/* ================================ Iterator ================================ */
/* Get a cursor that can be used as argument of hnsw_cursor_next() to iterate
* all the elements that remain there from the start to the end of the
* iteration, excluding newly added elements.
*
* The function returns NULL on out of memory. */
hnswCursor *hnsw_cursor_init(HNSW *index) {
hnswCursor *cursor = hmalloc(sizeof(*cursor));
if (cursor == NULL) return NULL;
cursor->next = index->cursors;
cursor->current = index->head;
index->cursors = cursor;
return cursor;
}
/* Free the cursor. Can be called both at the end of the iteration, when
* hnsw_cursor_next() returned NULL, or before. */
void hnsw_cursor_free(HNSW *index, hnswCursor *cursor) {
hnswCursor *x = index->cursors;
hnswCursor *prev = NULL;
while(x) {
if (x == cursor) {
if (prev)
prev->next = cursor->next;
else
index->cursors = cursor->next;
hfree(cursor);
return;
}
x = x->next;
}
}
/* Return the next element of the HNSW. See hnsw_cursor_init() for
* the guarantees of the function. */
hnswNode *hnsw_cursor_next(HNSW *index, hnswCursor *cursor) {
(void)index; // Unused but future proof to have.
hnswNode *ret = cursor->current;
if (ret) cursor->current = ret->next;
return ret;
}
/* Called by hnsw_unlink_node() if there is at least an active cursor.
* Will scan the cursors to see if any cursor is going to yeld this
* one, and in this case, updates the current element to the next. */
void hnsw_cursor_element_deleted(HNSW *index, hnswNode *deleted) {
hnswCursor *x = index->cursors;
while(x) {
if (x->current == deleted) x->current = deleted->next;
x = x->next;
}
}
/* ============================ Debugging stuff ============================= */
/* Show stats about nodes connections. */
void hnsw_print_stats(HNSW *index) {
if (!index || !index->head) {
printf("Empty index or NULL pointer passed\n");
return;
}
long long total_links = 0;
int min_links = -1; // We'll set this to first node's count.
int isolated_nodes = 0;
uint32_t node_count = 0;
// Iterate through all nodes using the linked list.
hnswNode *current = index->head;
while (current) {
// Count total links for this node across all layers.
int node_total_links = 0;
for (uint32_t layer = 0; layer <= current->level; layer++)
node_total_links += current->layers[layer].num_links;
// Update statistics.
total_links += node_total_links;
// Initialize or update minimum links.
if (min_links == -1 || node_total_links < min_links) {
min_links = node_total_links;
}
// Check if node is isolated (no links at all).
if (node_total_links == 0) isolated_nodes++;
node_count++;
current = current->next;
}
// Print statistics
printf("HNSW Graph Statistics:\n");
printf("----------------------\n");
printf("Total nodes: %u\n", node_count);
if (node_count > 0) {
printf("Average links per node: %.2f\n",
(float)total_links / node_count);
printf("Minimum links in a single node: %d\n", min_links);
printf("Number of isolated nodes: %d (%.1f%%)\n",
isolated_nodes,
(float)isolated_nodes * 100 / node_count);
}
}
/* Validate graph connectivity and link reciprocity. Takes pointers to store results:
* - connected_nodes: will contain number of reachable nodes from entry point.
* - reciprocal_links: will contain 1 if all links are reciprocal, 0 otherwise.
* Returns 0 on success, -1 on error (NULL parameters and such).
*/
int hnsw_validate_graph(HNSW *index, uint64_t *connected_nodes, int *reciprocal_links) {
if (!index || !connected_nodes || !reciprocal_links) return -1;
if (!index->enter_point) {
*connected_nodes = 0;
*reciprocal_links = 1; // Empty graph is valid.
return 0;
}
// Initialize connectivity check.
index->current_epoch[0]++;
*connected_nodes = 0;
*reciprocal_links = 1;
// Initialize node stack.
uint64_t stack_size = index->node_count;
hnswNode **stack = hmalloc(sizeof(hnswNode*) * stack_size);
if (!stack) return -1;
uint64_t stack_top = 0;
// Start from entry point.
index->enter_point->visited_epoch[0] = index->current_epoch[0];
(*connected_nodes)++;
stack[stack_top++] = index->enter_point;
// Process all reachable nodes.
while (stack_top > 0) {
hnswNode *current = stack[--stack_top];
// Explore all neighbors at each level.
for (uint32_t level = 0; level <= current->level; level++) {
for (uint64_t i = 0; i < current->layers[level].num_links; i++) {
hnswNode *neighbor = current->layers[level].links[i];
// Check reciprocity.
int found_backlink = 0;
for (uint64_t j = 0; j < neighbor->layers[level].num_links; j++) {
if (neighbor->layers[level].links[j] == current) {
found_backlink = 1;
break;
}
}
if (!found_backlink) {
*reciprocal_links = 0;
}
// If we haven't visited this neighbor yet.
if (neighbor->visited_epoch[0] != index->current_epoch[0]) {
neighbor->visited_epoch[0] = index->current_epoch[0];
(*connected_nodes)++;
if (stack_top < stack_size) {
stack[stack_top++] = neighbor;
} else {
// This should never happen in a valid graph.
hfree(stack);
return -1;
}
}
}
}
}
hfree(stack);
// Now scan for unreachable nodes and print debug info.
printf("\nUnreachable nodes debug information:\n");
printf("=====================================\n");
hnswNode *current = index->head;
while (current) {
if (current->visited_epoch[0] != index->current_epoch[0]) {
printf("\nUnreachable node found:\n");
printf("- Node pointer: %p\n", (void*)current);
printf("- Node ID: %llu\n", (unsigned long long)current->id);
printf("- Node level: %u\n", current->level);
// Print info about all its links at each level.
for (uint32_t level = 0; level <= current->level; level++) {
printf(" Level %u links (%u):\n", level,
current->layers[level].num_links);
for (uint64_t i = 0; i < current->layers[level].num_links; i++) {
hnswNode *neighbor = current->layers[level].links[i];
// Check reciprocity for this specific link
int found_backlink = 0;
for (uint64_t j = 0; j < neighbor->layers[level].num_links; j++) {
if (neighbor->layers[level].links[j] == current) {
found_backlink = 1;
break;
}
}
printf(" - Link %llu: pointer=%p, id=%llu, visited=%s,recpr=%s\n",
(unsigned long long)i, (void*)neighbor,
(unsigned long long)neighbor->id,
neighbor->visited_epoch[0] == index->current_epoch[0] ?
"yes" : "no",
found_backlink ? "yes" : "no");
}
}
}
current = current->next;
}
printf("Total connected nodes: %llu\n", (unsigned long long)*connected_nodes);
printf("All links are bi-directiona? %s\n", (*reciprocal_links)?"yes":"no");
return 0;
}
/* Test graph recall ability by verifying each node can be found searching
* for its own vector. This helps validate that the majority of nodes are
* properly connected and easily reachable in the graph structure. Every
* unreachable node is reported.
*
* Normally only a small percentage of nodes will be not reachable when
* visited. This is expected and part of the statistical properties
* of HNSW. This happens especially with entries that have an ambiguous
* meaning in the represented space, and are across two or multiple clusters
* of items.
*
* The function works by:
* 1. Iterating through all nodes in the linked list
* 2. Using each node's vector to perform a search with specified EF
* 3. Verifying the node can find itself as nearest neighbor
* 4. Collecting and reporting statistics about reachability
*
* This is just a debugging function that reports stuff in the standard
* output, part of the implementation because this kind of functions
* provide some visiblity on what happens inside the HNSW.
*/
void hnsw_test_graph_recall(HNSW *index, int test_ef, int verbose) {
// Stats
uint32_t total_nodes = 0;
uint32_t unreachable_nodes = 0;
uint32_t perfectly_reachable = 0; // Node finds itself as first result
// For storing search results
hnswNode **neighbors = hmalloc(sizeof(hnswNode*) * test_ef);
float *distances = hmalloc(sizeof(float) * test_ef);
float *test_vector = hmalloc(sizeof(float) * index->vector_dim);
if (!neighbors || !distances || !test_vector) {
hfree(neighbors);
hfree(distances);
hfree(test_vector);
return;
}
// Get a read slot for searching (even if it's highly unlikely that
// this test will be run threaded...).
int slot = hnsw_acquire_read_slot(index);
if (slot < 0) {
hfree(neighbors);
hfree(distances);
return;
}
printf("\nTesting graph recall\n");
printf("====================\n");
// Process one node at a time using the linked list
hnswNode *current = index->head;
while (current) {
total_nodes++;
// If using quantization, we need to reconstruct the normalized vector
if (index->quant_type == HNSW_QUANT_Q8) {
int8_t *quants = current->vector;
// Reconstruct normalized vector from quantized data
for (uint32_t j = 0; j < index->vector_dim; j++) {
test_vector[j] = (quants[j] * current->quants_range) / 127;
}
} else if (index->quant_type == HNSW_QUANT_NONE) {
memcpy(test_vector,current->vector,sizeof(float)*index->vector_dim);
} else {
assert(0 && "Quantization type not supported.");
}
// Search using the node's own vector with high ef
int found = hnsw_search(index, test_vector, test_ef, neighbors,
distances, slot, 1);
if (found == 0) continue; // Empty HNSW?
// Look for the node itself in the results
int found_self = 0;
int self_position = -1;
for (int i = 0; i < found; i++) {
if (neighbors[i] == current) {
found_self = 1;
self_position = i;
break;
}
}
if (!found_self || self_position != 0) {
unreachable_nodes++;
if (verbose) {
if (!found_self)
printf("\nNode %s cannot find itself:\n", (char*)current->value);
else
printf("\nNode %s is not top result:\n", (char*)current->value);
printf("- Node ID: %llu\n", (unsigned long long)current->id);
printf("- Node level: %u\n", current->level);
printf("- Found %d neighbors but self not among them\n", found);
printf("- Closest neighbor distance: %f\n", distances[0]);
printf("- Neighbors: ");
for (uint32_t i = 0; i < current->layers[0].num_links; i++) {
printf("%s ", (char*)current->layers[0].links[i]->value);
}
printf("\n");
printf("\nFound instead: ");
for (int j = 0; j < found && j < 10; j++) {
printf("%s ", (char*)neighbors[j]->value);
}
printf("\n");
}
} else {
perfectly_reachable++;
}
current = current->next;
}
// Release read slot
hnsw_release_read_slot(index, slot);
// Free resources
hfree(neighbors);
hfree(distances);
hfree(test_vector);
// Print final statistics
printf("Total nodes tested: %u\n", total_nodes);
printf("Perfectly reachable nodes: %u (%.1f%%)\n",
perfectly_reachable,
total_nodes ? (float)perfectly_reachable * 100 / total_nodes : 0);
printf("Unreachable/suboptimal nodes: %u (%.1f%%)\n",
unreachable_nodes,
total_nodes ? (float)unreachable_nodes * 100 / total_nodes : 0);
}