A lightweight, header-only C library for storing and searching high-dimensional vector embeddings with optional multithreading support.

• Header-only implementation (single file: vdb.h)
• Multiple distance metrics (cosine, euclidean, dot product)
• Optional thread-safe operations via #define VDB_MULTITHREADED
• Save/load database to/from disk
• Custom memory allocators support
• No dependencies (except pthreads for multithreading)
• Python bindings (refer to vdb.py)

/*test.c*/
#include "vdb.h"

int main(void) {
  vdb_database *db = vdb_create(128, VDB_METRIC_COSINE);

  float embedding[128] = { /* ... */ };
  vdb_add_vector(db, embedding, "vec1", NULL);

  float query[128] = { /* ... */ };
  vdb_result_set *results = vdb_search(db, query, 5);

  vdb_free_result_set(results);
  vdb_destroy(db);
  return 0;
}

Include vdb.h and compile with either approach, pthreads is not necessarily available which is why this is behind a flag.

Single-threaded:

gcc -O2 test.c -o test -lm

Multi-threaded:

gcc -O2 -DVDB_MULTITHREADED test.c -o test -lpthread -lm

vdb_database *vdb_create(size_t dimensions, vdb_metric metric) Creates a new vector database.

void vdb_destroy(vdb_database *db) Frees all resources associated with the database.

size_t vdb_count(const vdb_database *db) Returns the number of vectors in the database.

size_t vdb_dimensions(const vdb_database *db) Returns the dimensionality of vectors.

vdb_error vdb_add_vector(vdb_database *db, const float *data, const char *id, void *metadata) Adds a vector to the database with optional ID and metadata.

vdb_error vdb_remove_vector(vdb_database *db, size_t index) Removes a vector at the specified index.

**vdb_error vdb_get_vector(const vdb_database \*db, size_t index, float **out_data, char **out_id, void **out_metadata)** Retrieves a vector and its metadata.

vdb_result_set *vdb_search(const vdb_database *db, const float *query, size_t k) Performs k-nearest neighbor search. Returns NULL on error.

void vdb_free_result_set(vdb_result_set *result_set) Frees search results.

vdb_error vdb_save(const vdb_database *db, const char *filename) Saves the database to disk.

vdb_database *vdb_load(const char *filename) Loads a database from disk.

• VDB_METRIC_COSINE - Cosine distance (1 - cosine similarity)
• VDB_METRIC_EUCLIDEAN - Euclidean (L2) distance
• VDB_METRIC_DOT_PRODUCT - Negative dot product

VDB_OK = 0
VDB_ERROR_NULL_POINTER = -1
VDB_ERROR_INVALID_DIMENSIONS = -2
VDB_ERROR_OUT_OF_MEMORY = -3
VDB_ERROR_NOT_FOUND = -4
VDB_ERROR_INVALID_INDEX = -5
VDB_ERROR_THREAD_FAILURE = -6

Define before including vdb.h:

#define VDB_MALLOC my_malloc
#define VDB_FREE my_free
#define VDB_REALLOC my_realloc
#include "vdb.h"

When compiled with VDB_MULTITHREADED, all operations are thread-safe using read-write locks:

• Multiple threads can search simultaneously
• Add/remove operations are exclusive
• No external locking required

vdb uses a binary format with magic number 0x56444230:

• Header: magic (4 bytes), dimensions, count, metric
• Vectors: float array + ID length + ID string (for each vector)
• Metadata is not persisted

Apache v2.0 License