@infino-ai/infino 0.7.1 → 0.8.1

package/infino/index.d.ts

@@ -157,6 +157,15 @@ export declare class Table {
      * append == one commit.
      */
     append(data: AppendData): void;
+    /**
+     * {@link append}, naming the source the rows came from. The superfiles
+     * this commit writes are keyed `data/<stem>-<uuid>.sf.parquet`, with the
+     * stem the key-safe form of `sourceName` (lowercase `[a-z0-9_]`), so a
+     * bucket listing shows where each came from. The table behaves exactly as
+     * after `append`; the name is a label on the object key. Not available on
+     * a hosted table.
+     */
+    appendNamed(data: AppendData, sourceName: string): void;
     /** Ranked BM25 search; rows as records (or an Arrow `Table`). `score` is a
      * similarity (higher is better) — opposite direction from `vectorSearch`'s
      * distance. Fuse with `hybridSearch`. */

package/infino/index.js

@@ -250,6 +250,18 @@ class Table {
         const ipc = dataToIpc(data, () => this.schema());
         guard(this.remote, () => this.inner.append(ipc));
     }
+    /**
+     * {@link append}, naming the source the rows came from. The superfiles
+     * this commit writes are keyed `data/<stem>-<uuid>.sf.parquet`, with the
+     * stem the key-safe form of `sourceName` (lowercase `[a-z0-9_]`), so a
+     * bucket listing shows where each came from. The table behaves exactly as
+     * after `append`; the name is a label on the object key. Not available on
+     * a hosted table.
+     */
+    appendNamed(data, sourceName) {
+        const ipc = dataToIpc(data, () => this.schema());
+        guard(this.remote, () => this.inner.appendNamed(ipc, sourceName));
+    }
     bm25Search(column, query, k, opts = {}) {
         const buf = guard(this.remote, () => this.inner.bm25Search(column, query, k, opts.mode, opts.stats, opts.projection));
         return decode(buf, opts.arrow);

package/infino/native.d.ts

@@ -99,6 +99,46 @@ export interface FtsOptions {
    * table and cannot be changed afterwards.
    */
   analyzer?: string
+  /**
+   * Remove this column's stopwords — the very common words whose
+   * presence says almost nothing about what a document is about.
+   * `"english"` is the only set; omit for none (the default).
+   *
+   * Applies to both sides: the words leave the index and they leave
+   * a query, and each one removed leaves a hole in the token
+   * positions, so an exact phrase still knows the words it matched
+   * were not adjacent in the text. The trade is that once a word is
+   * not indexed, no query can find it — declare it on prose, not on
+   * short identifiers. Recorded with the table and unchangeable
+   * afterwards, since it decides what is in the index: the words it
+   * removed were never written, so changing it means re-ingesting
+   * from the source text. With `stored: false` that text is never
+   * kept, so the combination is permanent.
+   */
+  stopwords?: string
+  /**
+   * Reduce this column's words to their stems, so a search for one
+   * inflection finds the others (`running`, `runs`, `run`).
+   * `"english"` is the only stemmer; omit for none (the default).
+   *
+   * Applies to both sides, like `stopwords`. The trade is precision:
+   * stemming conflates words a reader would not, and there is no way
+   * to ask for an unstemmed form on a stemmed column. Recorded with
+   * the table and unchangeable afterwards: a stem is not invertible
+   * — the index holds `run`, never the `running` it came from — so
+   * changing it means re-ingesting from the source text, and with
+   * `stored: false` that text is never kept, making the combination
+   * permanent.
+   */
+  stemmer?: string
+  /**
+   * Record token positions, which is what exact phrase queries
+   * (`'"climate policy"'`) need. Default false: positions roughly
+   * double the column's index footprint, so they are a per-column
+   * opt-in. A column without them answers a phrase query with an
+   * error naming the column, never a silent bag-of-words fallback.
+   */
+  positions?: boolean
   /**
    * Keep the raw text in the table (default true). `false` makes the
    * column index-only: searchable, but the text is never stored, so
@@ -106,6 +146,18 @@ export interface FtsOptions {
    * batches still carry it).
    */
   stored?: boolean
+  /**
+   * BM25 term-frequency saturation, `> 0`; defaults to 1.2. Recorded
+   * with the table, and the stored score bounds are built with it,
+   * so a search that does not override it pays nothing. Pass with
+   * `b` or not at all.
+   */
+  k1?: number
+  /**
+   * BM25 length normalization, in `[0, 1]`; defaults to 0.75. Pass
+   * with `k1` or not at all.
+   */
+  b?: number
 }
 /**
  * Open (or create) a catalog rooted at `uri` (local dir, `memory://`, or
@@ -128,8 +180,9 @@ export declare function builderId(): string
 export declare class IndexSpec {
   constructor()
   /**
-   * Mark `column` (a UTF-8 string column) as full-text indexed,
-   * with optional per-column `options` (analyzer, stored).
+   * Mark `column` (a UTF-8 string column) as full-text indexed, with
+   * optional per-column `options` (analyzer, stopwords, stemmer,
+   * positions, stored, k1/b).
    */
   fts(column: string, options?: FtsOptions | undefined | null): IndexSpec
   /**
@@ -181,6 +234,14 @@ export declare class Table {
    * concatenated into one commit; an empty stream is a no-op.
    */
   append(data: Buffer): void
+  /**
+   * `append`, naming the source the rows came from: the superfiles this
+   * commit writes are keyed `data/<stem>-<uuid>.sf.parquet`, with the stem
+   * the key-safe form of `sourceName` (lowercase `[a-z0-9_]`), so a bucket
+   * listing shows where each came from. The table behaves exactly as
+   * after `append`; the name is a label on the object key.
+   */
+  appendNamed(data: Buffer, sourceName: string): void
   /**
    * BM25 search over one FTS column. Returns matching rows as an Arrow
    * IPC `Buffer` (read with `tableFromIPC`). `mode` is `"or"` (default)
@@ -188,8 +249,16 @@ export declare class Table {
    * `["_id", "score"]` for just id + score, or omit for full rows.
    * `score` is a similarity (higher is better) — opposite direction
    * from `vectorSearch`'s distance. Fuse with `hybridSearch`.
+   *
+   * `k1` / `b` override the columns' declared BM25 similarity
+   * parameters for this search only — pass both or neither. The
+   * stored score bounds belong to the declared pair, so the reader
+   * corrects them for the difference: results stay exact and only
+   * pruning power is traded, and nothing is rebuilt. A pair you mean
+   * to keep belongs on the column (`IndexSpec.fts`), where the
+   * bounds are built with it and the correction disappears.
    */
-  bm25Search(column: string, query: string, k: number, mode?: string | undefined | null, stats?: string | undefined | null, projection?: Array<string> | undefined | null): Buffer
+  bm25Search(column: string, query: string, k: number, mode?: string | undefined | null, stats?: string | undefined | null, projection?: Array<string> | undefined | null, k1?: number | undefined | null, b?: number | undefined | null): Buffer
   /**
    * Vector kNN over one vector column. `query` is a `Float32Array`
    * (crosses by reference — no copy). Returns matching rows as an Arrow

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@infino-ai/infino",
-  "version": "0.7.1",
+  "version": "0.8.1",
   "description": "Fast search on object storage — SQL, full-text, and vector search.",
   "license": "Apache-2.0",
   "publishConfig": {
@@ -82,12 +82,12 @@
     "apache-arrow": "^17"
   },
   "optionalDependencies": {
-    "infx-darwin-x64": "0.7.1",
-    "infx-darwin-arm64": "0.7.1",
-    "infx-linux-x64-gnu": "0.7.1",
-    "infx-linux-arm64-gnu": "0.7.1",
-    "infx-linux-x64-musl": "0.7.1",
-    "infx-linux-arm64-musl": "0.7.1"
+    "infx-darwin-x64": "0.8.1",
+    "infx-darwin-arm64": "0.8.1",
+    "infx-linux-x64-gnu": "0.8.1",
+    "infx-linux-arm64-gnu": "0.8.1",
+    "infx-linux-x64-musl": "0.8.1",
+    "infx-linux-arm64-musl": "0.8.1"
   },
   "devDependencies": {
     "@napi-rs/cli": "^2.18.0",