@autorag/librarian 2.13.4 → 2.13.6

package/dist/parser/spreadsheet-render.d.ts

@@ -0,0 +1,53 @@
+/**
+ * Renders decoded spreadsheet grids as retrieval-ready text.
+ *
+ * Every spreadsheet row becomes one line that names its sheet, cell range and table caption and
+ * pairs each value with its column label:
+ *
+ *   [예산총괄!A4:F4 · 2026년 연구개발비 편성표 (단위: 천원)] 구분: 직접비 | 비목: 인건비 | 편성액: 742,000
+ *
+ * Chunkers split mirror text between lines, so a value never reaches a chunk without its labels.
+ * A row too long for one retrieval chunk is split into several lines that each repeat the row key.
+ *
+ * Structure is inferred per block of consecutive non-empty rows:
+ * - leading rows holding a single value (titles such as "… (단위: 천원)") become the block caption,
+ *   and a title block directly above a table captions that table;
+ * - the first remaining row is the header only when the evidence is strong: a two-level header under
+ *   horizontally merged group cells, or a text header above at least three data rows with a column
+ *   whose values are mostly numbers, dates or formulas. Otherwise the rows are kept as written, so a
+ *   form-style label/value block is never relabelled with invented column names;
+ * - a block without its own header that continues the previous table's columns reuses that header;
+ * - a one-column vertical merge repeats its value on every covered row; horizontal merges and
+ *   blocks merged across rows and columns (side notes) are shown once.
+ */
+export type CellKind = "text" | "number" | "date" | "boolean" | "error" | "formula";
+export interface SheetCell {
+    /** Display text as the spreadsheet shows it; may contain line breaks. */
+    readonly text: string;
+    readonly kind: CellKind;
+    readonly comment?: string;
+}
+/** A merged range, 0-based and inclusive. */
+export interface SheetMerge {
+    readonly top: number;
+    readonly left: number;
+    readonly bottom: number;
+    readonly right: number;
+}
+export interface SheetGrid {
+    readonly name: string;
+    readonly hidden?: boolean;
+    /** Row-major cells. Rows and cells may be missing. */
+    readonly rows: readonly (readonly (SheetCell | undefined)[] | undefined)[];
+    readonly merges?: readonly SheetMerge[];
+}
+/**
+ * Upper bound on a rendered line, in MinSync's `estimate_tokens_upper` units. It keeps several lines
+ * per chunk under both MinSync's 4096-character recursive chunks and the 2048-token embedder budget.
+ */
+export declare const MAX_LINE_TOKENS = 768;
+/** MinSync's `estimate_tokens_upper`: half-tokens 1 per ASCII char, 2 per 2-byte, 4 per wider char. */
+export declare function estimateLineTokens(text: string): number;
+/** Kind of a value that arrived as text: numbers and dates typed as strings still count as data. */
+export declare function classifyText(text: string): CellKind;
+export declare function renderWorkbook(sheets: readonly SheetGrid[]): string;

package/dist/parser/spreadsheet.d.ts

@@ -0,0 +1,27 @@
+import { type ParseInput, type ParseOutput, Parser } from "./types.js";
+/** Workbook extensions SheetJS decodes for the spreadsheet parser. */
+export declare const SPREADSHEET_EXTENSIONS: readonly [".xlsx", ".xlsm", ".xlsb", ".xls", ".ods"];
+export interface SpreadsheetParserOptions {
+    /** Rows read per sheet, header included. Default 100,000. */
+    readonly maxRowsPerSheet?: number;
+}
+/**
+ * Parses Excel (`.xlsx`, `.xlsm`, `.xlsb`, legacy `.xls`) and OpenDocument (`.ods`) workbooks with
+ * SheetJS, then renders every row as one self-describing line (see `spreadsheet-render.ts`).
+ *
+ * Cell text is what the spreadsheet shows: cached formula results with their number formats applied
+ * (`742,000`, `32.1%`, dates). A format that scales or hides the stored number (`#,##0,"천원"`) keeps
+ * the stored value next to it, Korean-numeral formats (`[DBNum1]`) fall back to the plain number, and a
+ * formula saved without a result is shown as its formula (`=SUM(B2:B3)`).
+ *
+ * A web export saved under a workbook extension (an HTML table, common for bank and ERP downloads) is read
+ * as a sheet with its cell text kept as written, so `2023-06-01` and `00123` are not turned into numbers.
+ */
+export declare class SpreadsheetParser extends Parser {
+    #private;
+    readonly name = "spreadsheet";
+    readonly extensions: readonly [".xlsx", ".xlsm", ".xlsb", ".xls", ".ods"];
+    readonly version: string;
+    constructor(options?: SpreadsheetParserOptions);
+    parse(input: ParseInput): Promise<ParseOutput>;
+}

package/dist/profiles/loop-guard.d.ts

@@ -17,7 +17,13 @@ export declare function includedExtensions(include: readonly string[]): readonly
  * The automation runner appends it to the profile's own prompt; a profile cannot
  * leave it out. An agent that may write files could otherwise change a file the
  * rule watches, which fires the rule again and starts another run of the agent:
- * an endless loop. The section names the rule's folders, file types, and events,
- * and forbids every change that would fire the rule.
+ * an endless loop.
+ *
+ * The section states the one invariant (no change of this run may fire the rule
+ * again) and gives the agent the rule's exact decision procedure, so it can check
+ * its own plan instead of obeying a list of banned operations. A rename, a move or
+ * a write is allowed whenever the events it causes would not be picked up. The
+ * agent refuses only when the task cannot be done without a loop, and then
+ * proposes a softer variant.
  */
 export declare function loopGuardPrompt(rule: LoopGuardRule): string;

package/dist/profiles/retrieval-tools.d.ts

@@ -19,6 +19,7 @@ export declare function hasRetrievalTools(options: RetrievalToolOptions | undefi
  * throwaway memory file with the memory embedder off, so nothing the profile
  * searches is read from or recorded into the user's real `memory.json`. Only
  * the search tools are handed out; the backing agent's other surfaces (web,
- * Everything, FSearch, Jev, duplicate scan) stay off.
+ * Everything, FSearch, duplicate scan) stay off. Its decision model is built
+ * lazily and never asked: the tools search without the two-phase pipeline.
  */
 export declare function createRetrievalToolkit(config: CliConfig, options: RetrievalToolOptions): Promise<RetrievalToolkit>;