@superlinked/sie-lancedb 0.1.10

package/dist/index.d.cts

@@ -0,0 +1,141 @@
+import * as apache_arrow from 'apache-arrow';
+import { SIEClient } from '@superlinked/sie-sdk';
+
+/**
+ * Configuration options for SIEEmbeddingFunction.
+ */
+interface SIEEmbeddingFunctionOptions {
+    /** URL of the SIE server. @default "http://localhost:8080" */
+    baseUrl?: string;
+    /** Model name/ID to use for encoding. @default "BAAI/bge-m3" */
+    model?: string;
+    /** Optional pre-configured SIEClient instance. */
+    client?: SIEClient;
+    /** Instruction prefix for instruction-tuned models (e.g., E5). */
+    instruction?: string;
+    /** Output data type: "float32" (default), "float16", "int8", "binary". */
+    outputDtype?: string;
+    /** Target GPU type for routing (e.g., "l4", "a100-80gb"). */
+    gpu?: string;
+    /** Request timeout in milliseconds. @default 180000 */
+    timeout?: number;
+}
+/**
+ * Dense text embeddings via SIE for LanceDB.
+ *
+ * Implements LanceDB's EmbeddingFunction interface. Embeddings are computed
+ * automatically when using `sourceField()` / `vectorField()` schema helpers.
+ *
+ * Use `ndims()` or pass `{ dims }` to `vectorField()` for schema definition.
+ * `ndims()` queries the `/v1/models` metadata API (lightweight, no model loading).
+ *
+ * @example
+ * ```typescript
+ * import { SIEEmbeddingFunction } from "@superlinked/sie-lancedb";
+ * import { LanceSchema } from "@lancedb/lancedb/embedding";
+ *
+ * const func = new SIEEmbeddingFunction({ model: "BAAI/bge-m3" });
+ * const schema = LanceSchema({
+ *   text: func.sourceField(),
+ *   vector: func.vectorField({ dims: 1024 }),
+ * });
+ * ```
+ */
+declare class SIEEmbeddingFunction {
+    private readonly model;
+    private readonly baseUrl;
+    private readonly instruction;
+    private readonly outputDtype;
+    private readonly clientOptions;
+    private _client;
+    private _ndims;
+    constructor(options?: SIEEmbeddingFunctionOptions);
+    private get client();
+    /**
+     * Return embedding dimensionality from /v1/models metadata.
+     *
+     * Queries the SIE server's model config (lightweight GET, no model
+     * loading or inference). Cached after first call.
+     */
+    ndims(): Promise<number>;
+    /**
+     * Generate dense embeddings for a list of texts.
+     *
+     * @param texts - Texts to embed.
+     * @returns Array of embedding vectors.
+     */
+    generateEmbeddings(texts: string[]): Promise<number[][]>;
+    /**
+     * Embed documents (no isQuery flag).
+     */
+    embedDocuments(texts: string[]): Promise<number[][]>;
+    /**
+     * Embed a single query (passes isQuery: true for asymmetric models).
+     */
+    embedQuery(text: string): Promise<number[]>;
+    private extractDense;
+    close(): Promise<void>;
+}
+/**
+ * Configuration options for SIEReranker.
+ */
+interface SIERerankerOptions {
+    /** URL of the SIE server. @default "http://localhost:8080" */
+    baseUrl?: string;
+    /** Reranker model name/ID. @default "jinaai/jina-reranker-v2-base-multilingual" */
+    model?: string;
+    /** Name of the text column to score. @default "text" */
+    column?: string;
+    /** Optional pre-configured SIEClient instance. */
+    client?: SIEClient;
+    /** Target GPU type for routing. */
+    gpu?: string;
+    /** Request timeout in milliseconds. @default 180000 */
+    timeout?: number;
+}
+/**
+ * Cross-encoder reranker using SIE for LanceDB hybrid search.
+ *
+ * Implements LanceDB's Reranker interface. Plugs into hybrid search
+ * pipelines via `.rerank()`.
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-lancedb";
+ *
+ * const reranker = new SIEReranker({
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ * });
+ *
+ * const results = await table
+ *   .search("query", { queryType: "hybrid" })
+ *   .rerank(reranker)
+ *   .limit(10)
+ *   .toArray();
+ * ```
+ */
+declare class SIEReranker {
+    private readonly model;
+    private readonly column;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    private _client;
+    constructor(options?: SIERerankerOptions);
+    private get client();
+    /**
+     * Rerank hybrid search results (vector + FTS).
+     *
+     * This is the method LanceDB calls during `.rerank()`. It scores all
+     * rows against the query using SIE's cross-encoder and returns a
+     * RecordBatch with `_relevance_score` added.
+     */
+    rerankHybrid(query: string, vecResults: apache_arrow.RecordBatch, ftsResults: apache_arrow.RecordBatch): Promise<apache_arrow.RecordBatch>;
+    /**
+     * Merge vector and FTS result batches, deduplicating by _rowid.
+     * Rows from vecResults take priority for duplicate _rowid values.
+     */
+    private mergeResults;
+    close(): Promise<void>;
+}
+
+export { SIEEmbeddingFunction, type SIEEmbeddingFunctionOptions, SIEReranker, type SIERerankerOptions };

package/dist/index.d.ts

@@ -0,0 +1,141 @@
+import * as apache_arrow from 'apache-arrow';
+import { SIEClient } from '@superlinked/sie-sdk';
+
+/**
+ * Configuration options for SIEEmbeddingFunction.
+ */
+interface SIEEmbeddingFunctionOptions {
+    /** URL of the SIE server. @default "http://localhost:8080" */
+    baseUrl?: string;
+    /** Model name/ID to use for encoding. @default "BAAI/bge-m3" */
+    model?: string;
+    /** Optional pre-configured SIEClient instance. */
+    client?: SIEClient;
+    /** Instruction prefix for instruction-tuned models (e.g., E5). */
+    instruction?: string;
+    /** Output data type: "float32" (default), "float16", "int8", "binary". */
+    outputDtype?: string;
+    /** Target GPU type for routing (e.g., "l4", "a100-80gb"). */
+    gpu?: string;
+    /** Request timeout in milliseconds. @default 180000 */
+    timeout?: number;
+}
+/**
+ * Dense text embeddings via SIE for LanceDB.
+ *
+ * Implements LanceDB's EmbeddingFunction interface. Embeddings are computed
+ * automatically when using `sourceField()` / `vectorField()` schema helpers.
+ *
+ * Use `ndims()` or pass `{ dims }` to `vectorField()` for schema definition.
+ * `ndims()` queries the `/v1/models` metadata API (lightweight, no model loading).
+ *
+ * @example
+ * ```typescript
+ * import { SIEEmbeddingFunction } from "@superlinked/sie-lancedb";
+ * import { LanceSchema } from "@lancedb/lancedb/embedding";
+ *
+ * const func = new SIEEmbeddingFunction({ model: "BAAI/bge-m3" });
+ * const schema = LanceSchema({
+ *   text: func.sourceField(),
+ *   vector: func.vectorField({ dims: 1024 }),
+ * });
+ * ```
+ */
+declare class SIEEmbeddingFunction {
+    private readonly model;
+    private readonly baseUrl;
+    private readonly instruction;
+    private readonly outputDtype;
+    private readonly clientOptions;
+    private _client;
+    private _ndims;
+    constructor(options?: SIEEmbeddingFunctionOptions);
+    private get client();
+    /**
+     * Return embedding dimensionality from /v1/models metadata.
+     *
+     * Queries the SIE server's model config (lightweight GET, no model
+     * loading or inference). Cached after first call.
+     */
+    ndims(): Promise<number>;
+    /**
+     * Generate dense embeddings for a list of texts.
+     *
+     * @param texts - Texts to embed.
+     * @returns Array of embedding vectors.
+     */
+    generateEmbeddings(texts: string[]): Promise<number[][]>;
+    /**
+     * Embed documents (no isQuery flag).
+     */
+    embedDocuments(texts: string[]): Promise<number[][]>;
+    /**
+     * Embed a single query (passes isQuery: true for asymmetric models).
+     */
+    embedQuery(text: string): Promise<number[]>;
+    private extractDense;
+    close(): Promise<void>;
+}
+/**
+ * Configuration options for SIEReranker.
+ */
+interface SIERerankerOptions {
+    /** URL of the SIE server. @default "http://localhost:8080" */
+    baseUrl?: string;
+    /** Reranker model name/ID. @default "jinaai/jina-reranker-v2-base-multilingual" */
+    model?: string;
+    /** Name of the text column to score. @default "text" */
+    column?: string;
+    /** Optional pre-configured SIEClient instance. */
+    client?: SIEClient;
+    /** Target GPU type for routing. */
+    gpu?: string;
+    /** Request timeout in milliseconds. @default 180000 */
+    timeout?: number;
+}
+/**
+ * Cross-encoder reranker using SIE for LanceDB hybrid search.
+ *
+ * Implements LanceDB's Reranker interface. Plugs into hybrid search
+ * pipelines via `.rerank()`.
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-lancedb";
+ *
+ * const reranker = new SIEReranker({
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ * });
+ *
+ * const results = await table
+ *   .search("query", { queryType: "hybrid" })
+ *   .rerank(reranker)
+ *   .limit(10)
+ *   .toArray();
+ * ```
+ */
+declare class SIEReranker {
+    private readonly model;
+    private readonly column;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    private _client;
+    constructor(options?: SIERerankerOptions);
+    private get client();
+    /**
+     * Rerank hybrid search results (vector + FTS).
+     *
+     * This is the method LanceDB calls during `.rerank()`. It scores all
+     * rows against the query using SIE's cross-encoder and returns a
+     * RecordBatch with `_relevance_score` added.
+     */
+    rerankHybrid(query: string, vecResults: apache_arrow.RecordBatch, ftsResults: apache_arrow.RecordBatch): Promise<apache_arrow.RecordBatch>;
+    /**
+     * Merge vector and FTS result batches, deduplicating by _rowid.
+     * Rows from vecResults take priority for duplicate _rowid values.
+     */
+    private mergeResults;
+    close(): Promise<void>;
+}
+
+export { SIEEmbeddingFunction, type SIEEmbeddingFunctionOptions, SIEReranker, type SIERerankerOptions };