sonamu 0.7.11 → 0.7.13

package/src/vector/vector-search.ts

@@ -1,261 +0,0 @@
-import type { Knex } from "knex";
-import pgvector from "pgvector/knex";
-import { DEFAULT_VECTOR_CONFIG } from "./config";
-import { Embedding } from "./embedding";
-import type {
-  EmbeddingItem,
-  EmbeddingProvider,
-  HybridSearchOptions,
-  HybridSearchResult,
-  ProgressCallback,
-  VectorConfig,
-  VectorSearchOptions,
-  VectorSearchResult,
-} from "./types";
-
-/**
- * 벡터 검색
- * pgvector를 활용한 벡터 검색 및 하이브리드 검색 지원
- */
-export class VectorSearch<T = Record<string, unknown>> {
-  private db: Knex;
-  private config: VectorConfig;
-  private embedding: Embedding;
-  private tableName: string;
-
-  constructor(db: Knex, tableName: string, config: Partial<VectorConfig> = {}) {
-    this.db = db;
-    this.tableName = tableName;
-    this.config = {
-      voyage: { ...DEFAULT_VECTOR_CONFIG.voyage, ...config.voyage },
-      openai: { ...DEFAULT_VECTOR_CONFIG.openai, ...config.openai },
-      chunking: { ...DEFAULT_VECTOR_CONFIG.chunking, ...config.chunking },
-      search: { ...DEFAULT_VECTOR_CONFIG.search, ...config.search },
-      pgvector: { ...DEFAULT_VECTOR_CONFIG.pgvector, ...config.pgvector },
-    };
-    this.embedding = new Embedding(config);
-  }
-
-  /**
-   * 단일 항목에 임베딩 저장
-   */
-  async saveEmbedding(
-    id: number,
-    text: string,
-    provider: EmbeddingProvider,
-    embeddingColumn: string = "content_embedding",
-  ): Promise<void> {
-    const { embedding } = await this.embedding.embedOne(text, provider, "document");
-
-    await this.db(this.tableName)
-      .where("id", id)
-      .update({
-        [embeddingColumn]: pgvector.toSql(embedding),
-      });
-  }
-
-  /**
-   * 여러 항목에 임베딩 일괄 저장
-   */
-  async saveEmbeddingsBatch(
-    items: EmbeddingItem[],
-    provider: EmbeddingProvider,
-    embeddingColumn: string = "content_embedding",
-    onProgress?: ProgressCallback,
-  ): Promise<void> {
-    const texts = items.map((item) => item.text);
-    const embeddings = await this.embedding.embed(texts, provider, "document", onProgress);
-
-    await this.db.transaction(async (trx) => {
-      for (let i = 0; i < items.length; i++) {
-        await trx(this.tableName)
-          .where("id", items[i].id)
-          .update({
-            [embeddingColumn]: pgvector.toSql(embeddings[i].embedding),
-          });
-      }
-    });
-  }
-
-  /**
-   * 벡터 검색 (코사인 유사도)
-   */
-  async search(
-    query: string,
-    provider: EmbeddingProvider,
-    options: VectorSearchOptions = {},
-  ): Promise<VectorSearchResult<T>[]> {
-    const {
-      embeddingColumn = "content_embedding",
-      limit = this.config.search.defaultLimit,
-      threshold = this.config.search.similarityThreshold,
-      where,
-    } = options;
-
-    // 쿼리 임베딩 (input_type: 'query' 중요!)
-    const { embedding } = await this.embedding.embedOne(query, provider, "query");
-
-    // pgvector 세션 설정
-    if (this.config.pgvector.iterativeScan) {
-      await this.db.raw("SET hnsw.iterative_scan = relaxed_order");
-    }
-    await this.db.raw(`SET hnsw.ef_search = ${this.config.pgvector.efSearch}`);
-
-    // 코사인 유사도 = 1 - 코사인 거리
-    const vectorStr = pgvector.toSql(embedding);
-    let queryBuilder = this.db(this.tableName)
-      .select("*")
-      .select(this.db.raw(`1 - (${embeddingColumn} <=> ?::vector) AS similarity`, [vectorStr]))
-      .whereNotNull(embeddingColumn)
-      .orderByRaw(`${embeddingColumn} <=> ?::vector`, [vectorStr])
-      .limit(limit);
-
-    if (where) {
-      queryBuilder = queryBuilder.whereRaw(where);
-    }
-
-    const rows = await queryBuilder;
-
-    return rows
-      .filter((row: { similarity: number }) => row.similarity >= threshold)
-      .map((row: T & { similarity: number }) => ({
-        id: (row as unknown as { id: number }).id,
-        similarity: parseFloat(String(row.similarity)),
-        data: row as T,
-      }));
-  }
-
-  /**
-   * 하이브리드 검색 (Vector + FTS)
-   */
-  async hybridSearch(
-    query: string,
-    provider: EmbeddingProvider,
-    options: HybridSearchOptions = {},
-  ): Promise<HybridSearchResult<T>[]> {
-    const {
-      embeddingColumn = "content_embedding",
-      ftsColumn = "content_tsv",
-      limit = this.config.search.defaultLimit,
-      vectorWeight = this.config.search.vectorWeight,
-      ftsWeight = this.config.search.ftsWeight,
-    } = options;
-
-    const { embedding } = await this.embedding.embedOne(query, provider, "query");
-    const vectorStr = pgvector.toSql(embedding);
-
-    // pgvector 세션 설정
-    if (this.config.pgvector.iterativeScan) {
-      await this.db.raw("SET hnsw.iterative_scan = relaxed_order");
-    }
-    await this.db.raw(`SET hnsw.ef_search = ${this.config.pgvector.efSearch}`);
-
-    const sql = `
-      WITH vector_search AS (
-        SELECT
-          id,
-          ROW_NUMBER() OVER (ORDER BY ${embeddingColumn} <=> ?::vector) AS rank
-        FROM ${this.tableName}
-        WHERE ${embeddingColumn} IS NOT NULL
-        ORDER BY ${embeddingColumn} <=> ?::vector
-        LIMIT 50
-      ),
-      fts_search AS (
-        SELECT
-          id,
-          ROW_NUMBER() OVER (ORDER BY ts_rank(${ftsColumn}, query) DESC) AS rank
-        FROM ${this.tableName}, plainto_tsquery('simple', ?) query
-        WHERE ${ftsColumn} @@ query
-        LIMIT 50
-      ),
-      combined AS (
-        SELECT
-          COALESCE(v.id, f.id) AS id,
-          COALESCE(1.0 / (60 + v.rank), 0) AS vector_score,
-          COALESCE(1.0 / (60 + f.rank), 0) AS fts_score
-        FROM vector_search v
-        FULL OUTER JOIN fts_search f ON v.id = f.id
-      )
-      SELECT
-        t.*,
-        c.vector_score,
-        c.fts_score,
-        (c.vector_score * ? + c.fts_score * ?) AS similarity
-      FROM combined c
-      JOIN ${this.tableName} t ON c.id = t.id
-      ORDER BY similarity DESC
-      LIMIT ?
-    `;
-
-    const { rows } = await this.db.raw(sql, [
-      vectorStr,
-      vectorStr,
-      query,
-      vectorWeight,
-      ftsWeight,
-      limit,
-    ]);
-
-    return rows.map(
-      (
-        row: T & {
-          similarity: number;
-          vector_score: number;
-          fts_score: number;
-        },
-      ) => ({
-        id: (row as unknown as { id: number }).id,
-        similarity: parseFloat(String(row.similarity)),
-        vectorScore: parseFloat(String(row.vector_score)),
-        ftsScore: parseFloat(String(row.fts_score)),
-        data: row as T,
-      }),
-    );
-  }
-
-  /**
-   * 임베딩 현황 조회
-   */
-  async getEmbeddingStatus(embeddingColumn: string = "content_embedding"): Promise<{
-    total: number;
-    withEmbedding: number;
-    withoutEmbedding: number;
-  }> {
-    const result = await this.db(this.tableName)
-      .count("* as total")
-      .count(`${embeddingColumn} as with_embedding`)
-      .first();
-
-    const total = parseInt(String(result?.total ?? 0), 10);
-    const withEmbedding = parseInt(String(result?.with_embedding ?? 0), 10);
-
-    return {
-      total,
-      withEmbedding,
-      withoutEmbedding: total - withEmbedding,
-    };
-  }
-
-  /**
-   * 임베딩이 없는 항목 ID 조회
-   */
-  async getItemsWithoutEmbedding(
-    embeddingColumn: string = "content_embedding",
-    limit: number = 100,
-  ): Promise<number[]> {
-    const rows = await this.db(this.tableName)
-      .select("id")
-      .whereNull(embeddingColumn)
-      .orderBy("id")
-      .limit(limit);
-
-    return rows.map((row: { id: number }) => row.id);
-  }
-
-  /**
-   * Embedding 인스턴스 반환 (고급 사용)
-   */
-  getEmbedding(): Embedding {
-    return this.embedding;
-  }
-}