Class PgVectorProvider

packages/providers/src/vector/pgvector.ts:75–837

 * - For very large single-user datasets (>100K vectors), consider partial HNSW indexes
 */
export class PgVectorProvider implements IVectorProvider {
  private prisma: PrismaClient;
  private dimensions: number;
  private infrastructureInitialized = false;

  private readonly INDEX_CONFIGS = [
    {
      table: "statement_embeddings",
      name: "statement_embeddings_vector_idx",
    },
    {
      table: "episode_embeddings",
      name: "episode_embeddings_vector_idx",
    },
    {
      table: "entity_embeddings",
      name: "entity_embeddings_vector_idx",
    },
    {
      table: "compacted_session_embeddings",
      name: "compacted_session_embeddings_vector_idx",
    },
    {
      table: "label_embeddings",
      name: "label_embeddings_vector_idx",
    },
    {
      table: "voice_aspect_embeddings",
      name: "voice_aspect_embeddings_vector_idx",
    },
  ] as const;

  constructor(config: PgVectorConfig) {
    this.prisma = config.prisma;
    this.dimensions = config.dimensions
      ?? (process.env.EMBEDDING_MODEL_SIZE
        ? parseInt(process.env.EMBEDDING_MODEL_SIZE, 10)
        : 1024);
  }

  /**
   * Initialize HNSW indexes for all embedding tables
   * This method is idempotent and safe to call multiple times
   *
   * @returns true if initialization succeeded or was already done, false on failure
   */
  async initializeInfrastructure(): Promise<boolean> {
    if (this.infrastructureInitialized) {
      console.log("[PgVector] Infrastructure already initialized, skipping...");
      return true;
    }

    try {
      console.log(`[PgVector] Initializing HNSW indexes with dimension ${this.dimensions}...`);

      // Create indexes if they don't exist
      for (const { table, name } of this.INDEX_CONFIGS) {
        await this.ensureIndex(table, name);