@mailwoman/neural 2.0.0

package/out/classifier.d.ts

@@ -0,0 +1,40 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `NeuralAddressClassifier` ties together the tokenizer, the ONNX inference runner, and the
+ *   `@mailwoman/core` decoder. Single user-facing entrypoint: `parse(text)` returns an
+ *   `AddressTree` ready for projection into JSON / tuple / XML.
+ *
+ *   Convenience wrappers `parseJson` / `parseTuples` / `parseXml` project the tree on the way out.
+ */
+import { type AddressTree, type ComponentTag, decodeAsXml } from "@mailwoman/core/decoder";
+import { OnnxRunner } from "./onnx-runner.js";
+import { MailwomanTokenizer } from "./tokenizer.js";
+import { type ResolveWeightsOpts } from "./weights.js";
+export interface NeuralAddressClassifierConfig {
+    tokenizer: MailwomanTokenizer;
+    runner: OnnxRunner;
+    /** Label vocabulary in the order the model emits them. Defaults to Stage 1 (v0.1.0/v0.2.0). */
+    labels?: readonly string[];
+}
+export declare class NeuralAddressClassifier {
+    private readonly cfg;
+    private readonly labels;
+    constructor(cfg: NeuralAddressClassifierConfig);
+    /**
+     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and
+     * ONNX runner, and returns a ready-to-use classifier.
+     *
+     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package →
+     * throws a single actionable error.
+     */
+    static loadFromWeights(opts?: ResolveWeightsOpts): Promise<NeuralAddressClassifier>;
+    /** Tokenize → infer → argmax/softmax → decoder tree. */
+    parse(text: string): Promise<AddressTree>;
+    parseJson(text: string): Promise<Partial<Record<ComponentTag, string>>>;
+    parseTuples(text: string): Promise<Array<[ComponentTag, string]>>;
+    parseXml(text: string, opts?: Parameters<typeof decodeAsXml>[1]): Promise<string>;
+}
+//# sourceMappingURL=classifier.d.ts.map

package/out/classifier.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"classifier.d.ts","sourceRoot":"","sources":["../classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EACN,KAAK,WAAW,EAChB,KAAK,YAAY,EAKjB,WAAW,EACX,MAAM,yBAAyB,CAAA;AAEhC,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAA;AAC7C,OAAO,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AACnD,OAAO,EAAE,KAAK,kBAAkB,EAAkB,MAAM,cAAc,CAAA;AAEtE,MAAM,WAAW,6BAA6B;IAC7C,SAAS,EAAE,kBAAkB,CAAA;IAC7B,MAAM,EAAE,UAAU,CAAA;IAClB,+FAA+F;IAC/F,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;CAC1B;AAED,qBAAa,uBAAuB;IAGvB,OAAO,CAAC,QAAQ,CAAC,GAAG;IAFhC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmB;gBAEb,GAAG,EAAE,6BAA6B;IAI/D;;;;;;OAMG;WACU,eAAe,CAAC,IAAI,GAAE,kBAAuB,GAAG,OAAO,CAAC,uBAAuB,CAAC;IAS7F,wDAAwD;IAClD,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,WAAW,CAAC;IAqBzC,SAAS,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IAIvE,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IAIjE,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,UAAU,CAAC,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;CAGvF"}

package/out/classifier.js

@@ -0,0 +1,83 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `NeuralAddressClassifier` ties together the tokenizer, the ONNX inference runner, and the
+ *   `@mailwoman/core` decoder. Single user-facing entrypoint: `parse(text)` returns an
+ *   `AddressTree` ready for projection into JSON / tuple / XML.
+ *
+ *   Convenience wrappers `parseJson` / `parseTuples` / `parseXml` project the tree on the way out.
+ */
+import { buildAddressTree, decodeAsJson, decodeAsTuples, decodeAsXml, } from "@mailwoman/core/decoder";
+import { STAGE1_BIO_LABELS } from "./labels.js";
+import { OnnxRunner } from "./onnx-runner.js";
+import { MailwomanTokenizer } from "./tokenizer.js";
+import { resolveWeights } from "./weights.js";
+export class NeuralAddressClassifier {
+    cfg;
+    labels;
+    constructor(cfg) {
+        this.cfg = cfg;
+        this.labels = cfg.labels ?? STAGE1_BIO_LABELS;
+    }
+    /**
+     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and
+     * ONNX runner, and returns a ready-to-use classifier.
+     *
+     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package →
+     * throws a single actionable error.
+     */
+    static async loadFromWeights(opts = {}) {
+        const { modelPath, tokenizerPath } = resolveWeights(opts);
+        const [tokenizer, runner] = await Promise.all([
+            MailwomanTokenizer.loadFromFile(tokenizerPath),
+            OnnxRunner.create(modelPath),
+        ]);
+        return new NeuralAddressClassifier({ tokenizer, runner });
+    }
+    /** Tokenize → infer → argmax/softmax → decoder tree. */
+    async parse(text) {
+        if (text.length === 0)
+            return { raw: text, roots: [] };
+        const { pieces, ids } = this.cfg.tokenizer.encode(text);
+        const { logits } = await this.cfg.runner.infer(ids);
+        const tokens = pieces.map((p, i) => {
+            const row = logits[i];
+            const { idx, conf } = argmaxSoftmax(row);
+            return {
+                piece: p.piece,
+                start: p.start,
+                end: p.end,
+                label: (this.labels[idx] ?? "O"),
+                confidence: conf,
+            };
+        });
+        return buildAddressTree(text, tokens);
+    }
+    async parseJson(text) {
+        return decodeAsJson(await this.parse(text));
+    }
+    async parseTuples(text) {
+        return decodeAsTuples(await this.parse(text));
+    }
+    async parseXml(text, opts) {
+        return decodeAsXml(await this.parse(text), opts);
+    }
+}
+function argmaxSoftmax(row) {
+    let maxIdx = 0;
+    let maxVal = row[0];
+    for (let i = 1; i < row.length; i++) {
+        if (row[i] > maxVal) {
+            maxVal = row[i];
+            maxIdx = i;
+        }
+    }
+    let sumExp = 0;
+    for (const v of row)
+        sumExp += Math.exp(v - maxVal);
+    const conf = 1 / sumExp;
+    return { idx: maxIdx, conf };
+}
+//# sourceMappingURL=classifier.js.map

package/out/classifier.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"classifier.js","sourceRoot":"","sources":["../classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAIN,gBAAgB,EAChB,YAAY,EACZ,cAAc,EACd,WAAW,GACX,MAAM,yBAAyB,CAAA;AAChC,OAAO,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAA;AAC/C,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAA;AAC7C,OAAO,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AACnD,OAAO,EAA2B,cAAc,EAAE,MAAM,cAAc,CAAA;AAStE,MAAM,OAAO,uBAAuB;IAGN;IAFZ,MAAM,CAAmB;IAE1C,YAA6B,GAAkC;QAAlC,QAAG,GAAH,GAAG,CAA+B;QAC9D,IAAI,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,IAAI,iBAAiB,CAAA;IAC9C,CAAC;IAED;;;;;;OAMG;IACH,MAAM,CAAC,KAAK,CAAC,eAAe,CAAC,OAA2B,EAAE;QACzD,MAAM,EAAE,SAAS,EAAE,aAAa,EAAE,GAAG,cAAc,CAAC,IAAI,CAAC,CAAA;QACzD,MAAM,CAAC,SAAS,EAAE,MAAM,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;YAC7C,kBAAkB,CAAC,YAAY,CAAC,aAAa,CAAC;YAC9C,UAAU,CAAC,MAAM,CAAC,SAAS,CAAC;SAC5B,CAAC,CAAA;QACF,OAAO,IAAI,uBAAuB,CAAC,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC,CAAA;IAC1D,CAAC;IAED,wDAAwD;IACxD,KAAK,CAAC,KAAK,CAAC,IAAY;QACvB,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,KAAK,EAAE,EAAE,EAAE,CAAA;QAEtD,MAAM,EAAE,MAAM,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,CAAA;QACvD,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;QAEnD,MAAM,MAAM,GAAmB,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;YAClD,MAAM,GAAG,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACtB,MAAM,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,aAAa,CAAC,GAAG,CAAC,CAAA;YACxC,OAAO;gBACN,KAAK,EAAE,CAAC,CAAC,KAAK;gBACd,KAAK,EAAE,CAAC,CAAC,KAAK;gBACd,GAAG,EAAE,CAAC,CAAC,GAAG;gBACV,KAAK,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,GAAG,CAA0B;gBACzD,UAAU,EAAE,IAAI;aAChB,CAAA;QACF,CAAC,CAAC,CAAA;QAEF,OAAO,gBAAgB,CAAC,IAAI,EAAE,MAAM,CAAC,CAAA;IACtC,CAAC;IAED,KAAK,CAAC,SAAS,CAAC,IAAY;QAC3B,OAAO,YAAY,CAAC,MAAM,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAA;IAC5C,CAAC;IAED,KAAK,CAAC,WAAW,CAAC,IAAY;QAC7B,OAAO,cAAc,CAAC,MAAM,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAA;IAC9C,CAAC;IAED,KAAK,CAAC,QAAQ,CAAC,IAAY,EAAE,IAAwC;QACpE,OAAO,WAAW,CAAC,MAAM,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,CAAC,CAAA;IACjD,CAAC;CACD;AAED,SAAS,aAAa,CAAC,GAAa;IACnC,IAAI,MAAM,GAAG,CAAC,CAAA;IACd,IAAI,MAAM,GAAG,GAAG,CAAC,CAAC,CAAE,CAAA;IACpB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACrC,IAAI,GAAG,CAAC,CAAC,CAAE,GAAG,MAAM,EAAE,CAAC;YACtB,MAAM,GAAG,GAAG,CAAC,CAAC,CAAE,CAAA;YAChB,MAAM,GAAG,CAAC,CAAA;QACX,CAAC;IACF,CAAC;IACD,IAAI,MAAM,GAAG,CAAC,CAAA;IACd,KAAK,MAAM,CAAC,IAAI,GAAG;QAAE,MAAM,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,CAAA;IACnD,MAAM,IAAI,GAAG,CAAC,GAAG,MAAM,CAAA;IACvB,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,IAAI,EAAE,CAAA;AAC7B,CAAC"}

package/out/index.d.ts

@@ -0,0 +1,12 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ */
+export * from "./classifier.js";
+export * from "./labels.js";
+export * from "./onnx-runner.js";
+export * from "./proposal-classifier.js";
+export * from "./tokenizer.js";
+export * from "./weights.js";
+//# sourceMappingURL=index.d.ts.map

package/out/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,cAAc,gBAAgB,CAAA;AAC9B,cAAc,cAAc,CAAA"}

package/out/index.js

@@ -0,0 +1,12 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ */
+export * from "./classifier.js";
+export * from "./labels.js";
+export * from "./onnx-runner.js";
+export * from "./proposal-classifier.js";
+export * from "./tokenizer.js";
+export * from "./weights.js";
+//# sourceMappingURL=index.js.map

package/out/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,cAAc,gBAAgB,CAAA;AAC9B,cAAc,cAAc,CAAA"}

package/out/labels.d.ts

@@ -0,0 +1,20 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Mirror of `packages/corpus-python/src/mailwoman_train/labels.py::STAGE1_BIO_LABELS`.
+ *
+ *   The v0.1.0 / v0.2.0 weight packages were trained with this exact label order. Any drift here
+ *   silently corrupts downstream BIO decoding — index 5 must mean `B-locality` on both sides.
+ *
+ *   Stage 2+ models will support more labels (street, house_number, venue, …). The plan is to plumb
+ *   the label set through `model-card.json` at load time rather than hard-coding it here. Until
+ *   then this file is the source of truth on the TS side.
+ */
+import type { BioLabel } from "@mailwoman/core/decoder";
+/** Coarse component tags trained in Phase 2 Stage 1 (v0.1.0 / v0.2.0). */
+export declare const STAGE1_COARSE_TAGS: readonly ["country", "region", "locality", "dependent_locality", "postcode", "subregion", "cedex"];
+/** BIO label vocabulary for Stage 1 — O + (B-/I- per coarse tag). 1 + 14 = 15 labels. */
+export declare const STAGE1_BIO_LABELS: readonly BioLabel[];
+//# sourceMappingURL=labels.d.ts.map

package/out/labels.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"labels.d.ts","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,yBAAyB,CAAA;AAEvD,0EAA0E;AAC1E,eAAO,MAAM,kBAAkB,oGAQrB,CAAA;AAEV,yFAAyF;AACzF,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA"}

package/out/labels.js

@@ -0,0 +1,30 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Mirror of `packages/corpus-python/src/mailwoman_train/labels.py::STAGE1_BIO_LABELS`.
+ *
+ *   The v0.1.0 / v0.2.0 weight packages were trained with this exact label order. Any drift here
+ *   silently corrupts downstream BIO decoding — index 5 must mean `B-locality` on both sides.
+ *
+ *   Stage 2+ models will support more labels (street, house_number, venue, …). The plan is to plumb
+ *   the label set through `model-card.json` at load time rather than hard-coding it here. Until
+ *   then this file is the source of truth on the TS side.
+ */
+/** Coarse component tags trained in Phase 2 Stage 1 (v0.1.0 / v0.2.0). */
+export const STAGE1_COARSE_TAGS = [
+    "country",
+    "region",
+    "locality",
+    "dependent_locality",
+    "postcode",
+    "subregion",
+    "cedex",
+];
+/** BIO label vocabulary for Stage 1 — O + (B-/I- per coarse tag). 1 + 14 = 15 labels. */
+export const STAGE1_BIO_LABELS = Object.freeze([
+    "O",
+    ...STAGE1_COARSE_TAGS.flatMap((tag) => [`B-${tag}`, `I-${tag}`]),
+]);
+//# sourceMappingURL=labels.js.map

package/out/labels.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"labels.js","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAIH,0EAA0E;AAC1E,MAAM,CAAC,MAAM,kBAAkB,GAAG;IACjC,SAAS;IACT,QAAQ;IACR,UAAU;IACV,oBAAoB;IACpB,UAAU;IACV,WAAW;IACX,OAAO;CACE,CAAA;AAEV,yFAAyF;AACzF,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,kBAAkB,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACxF,CAAC,CAAA"}

package/out/onnx-runner.d.ts

@@ -0,0 +1,56 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   ONNX inference wrapper.
+ *
+ *   Loads a token-classification model exported by `packages/corpus-python/src/mailwoman_train/
+ *   export_onnx.py` (BertForTokenClassification w/ inputs `input_ids` + `attention_mask`, output
+ *   `logits` shape `[batch, sequence, num_labels]`).
+ *
+ *   Lazy-loads on first `infer()` call unless `warmup: true` is passed; the constructor itself is
+ *   cheap and synchronous.
+ */
+export interface OnnxRunnerOpts {
+    /** If true, load the model immediately in `create()`. Default false. */
+    warmup?: boolean;
+    /**
+     * Fixed sequence length the model expects. v0.1.0 / v0.2.0 quantization baked in 128 (the
+     * training-time max position) even though the fp32 export specified dynamic axes — re-quantize
+     * with a different shape to override. Inputs shorter than this are padded with id `0` and masked
+     * out via attention_mask=0; inputs longer are truncated.
+     */
+    fixedSeqLen?: number;
+}
+/** Default sequence length for v0.1.0 / v0.2.0 (BertConfig max_position_embeddings = 128). */
+export declare const DEFAULT_FIXED_SEQ_LEN = 128;
+export interface InferResult {
+    /** Logits per token per label, indexed as `logits[tokenIdx][labelIdx]`. */
+    logits: number[][];
+    /** Number of label classes (the inner-dim of the logits tensor). */
+    numLabels: number;
+}
+export declare class OnnxRunner {
+    private readonly modelPath;
+    private readonly modelBytes;
+    private session;
+    private loadPromise;
+    readonly fixedSeqLen: number;
+    private constructor();
+    /** Load by path. Reads the model lazily unless `warmup` is true. */
+    static create(modelPath: string, opts?: OnnxRunnerOpts): Promise<OnnxRunner>;
+    /** Load from an already-read byte buffer. */
+    static fromBytes(modelBytes: Uint8Array, opts?: OnnxRunnerOpts): Promise<OnnxRunner>;
+    private ensureSession;
+    /**
+     * Run inference on a single token id sequence.
+     *
+     * Pads to `fixedSeqLen` (default 128) with id 0 + mask 0; truncates if longer. Output is sliced
+     * back to the actual input length.
+     *
+     * @param tokenIds The id sequence produced by the tokenizer (no special tokens added).
+     */
+    infer(tokenIds: number[]): Promise<InferResult>;
+}
+//# sourceMappingURL=onnx-runner.d.ts.map

package/out/onnx-runner.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-runner.d.ts","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAKH,MAAM,WAAW,cAAc;IAC9B,wEAAwE;IACxE,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB;;;;;OAKG;IACH,WAAW,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,8FAA8F;AAC9F,eAAO,MAAM,qBAAqB,MAAM,CAAA;AAExC,MAAM,WAAW,WAAW;IAC3B,2EAA2E;IAC3E,MAAM,EAAE,MAAM,EAAE,EAAE,CAAA;IAClB,oEAAoE;IACpE,SAAS,EAAE,MAAM,CAAA;CACjB;AAED,qBAAa,UAAU;IAMrB,OAAO,CAAC,QAAQ,CAAC,SAAS;IAC1B,OAAO,CAAC,QAAQ,CAAC,UAAU;IAN5B,OAAO,CAAC,OAAO,CAAoC;IACnD,OAAO,CAAC,WAAW,CAA6C;IAChE,SAAgB,WAAW,EAAE,MAAM,CAAA;IAEnC,OAAO;IAQP,oEAAoE;WACvD,MAAM,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;IAMtF,6CAA6C;WAChC,SAAS,CAAC,UAAU,EAAE,UAAU,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;YAMhF,aAAa;IAgB3B;;;;;;;OAOG;IACG,KAAK,CAAC,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,WAAW,CAAC;CA8BrD"}

package/out/onnx-runner.js

@@ -0,0 +1,98 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   ONNX inference wrapper.
+ *
+ *   Loads a token-classification model exported by `packages/corpus-python/src/mailwoman_train/
+ *   export_onnx.py` (BertForTokenClassification w/ inputs `input_ids` + `attention_mask`, output
+ *   `logits` shape `[batch, sequence, num_labels]`).
+ *
+ *   Lazy-loads on first `infer()` call unless `warmup: true` is passed; the constructor itself is
+ *   cheap and synchronous.
+ */
+import { promises as fs } from "node:fs";
+import ort from "onnxruntime-node";
+/** Default sequence length for v0.1.0 / v0.2.0 (BertConfig max_position_embeddings = 128). */
+export const DEFAULT_FIXED_SEQ_LEN = 128;
+export class OnnxRunner {
+    modelPath;
+    modelBytes;
+    session = null;
+    loadPromise = null;
+    fixedSeqLen;
+    constructor(modelPath, modelBytes, opts) {
+        this.modelPath = modelPath;
+        this.modelBytes = modelBytes;
+        this.fixedSeqLen = opts.fixedSeqLen ?? DEFAULT_FIXED_SEQ_LEN;
+    }
+    /** Load by path. Reads the model lazily unless `warmup` is true. */
+    static async create(modelPath, opts = {}) {
+        const runner = new OnnxRunner(modelPath, null, opts);
+        if (opts.warmup)
+            await runner.ensureSession();
+        return runner;
+    }
+    /** Load from an already-read byte buffer. */
+    static async fromBytes(modelBytes, opts = {}) {
+        const runner = new OnnxRunner("(bytes)", modelBytes, opts);
+        if (opts.warmup)
+            await runner.ensureSession();
+        return runner;
+    }
+    async ensureSession() {
+        if (this.session)
+            return this.session;
+        if (!this.loadPromise) {
+            this.loadPromise = (async () => {
+                const bytes = this.modelBytes ?? new Uint8Array(await fs.readFile(this.modelPath));
+                const session = await ort.InferenceSession.create(bytes, {
+                    executionProviders: ["cpu"],
+                    graphOptimizationLevel: "all",
+                });
+                this.session = session;
+                return session;
+            })();
+        }
+        return this.loadPromise;
+    }
+    /**
+     * Run inference on a single token id sequence.
+     *
+     * Pads to `fixedSeqLen` (default 128) with id 0 + mask 0; truncates if longer. Output is sliced
+     * back to the actual input length.
+     *
+     * @param tokenIds The id sequence produced by the tokenizer (no special tokens added).
+     */
+    async infer(tokenIds) {
+        const session = await this.ensureSession();
+        const seqLen = Math.min(tokenIds.length, this.fixedSeqLen);
+        const padded = new BigInt64Array(this.fixedSeqLen);
+        const mask = new BigInt64Array(this.fixedSeqLen);
+        for (let i = 0; i < seqLen; i++) {
+            padded[i] = BigInt(tokenIds[i]);
+            mask[i] = 1n;
+        }
+        const feeds = {
+            input_ids: new ort.Tensor("int64", padded, [1, this.fixedSeqLen]),
+            attention_mask: new ort.Tensor("int64", mask, [1, this.fixedSeqLen]),
+        };
+        const output = await session.run(feeds);
+        const logitsTensor = output.logits;
+        if (!logitsTensor)
+            throw new Error("ONNX model did not return a `logits` output");
+        const data = logitsTensor.data;
+        const [, , numLabels] = logitsTensor.dims;
+        const logits = [];
+        for (let t = 0; t < seqLen; t++) {
+            const row = new Array(numLabels);
+            const base = t * numLabels;
+            for (let l = 0; l < numLabels; l++)
+                row[l] = data[base + l];
+            logits.push(row);
+        }
+        return { logits, numLabels };
+    }
+}
+//# sourceMappingURL=onnx-runner.js.map

package/out/onnx-runner.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-runner.js","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,QAAQ,IAAI,EAAE,EAAE,MAAM,SAAS,CAAA;AACxC,OAAO,GAAG,MAAM,kBAAkB,CAAA;AAclC,8FAA8F;AAC9F,MAAM,CAAC,MAAM,qBAAqB,GAAG,GAAG,CAAA;AASxC,MAAM,OAAO,UAAU;IAMJ;IACA;IANV,OAAO,GAAgC,IAAI,CAAA;IAC3C,WAAW,GAAyC,IAAI,CAAA;IAChD,WAAW,CAAQ;IAEnC,YACkB,SAAiB,EACjB,UAA6B,EAC9C,IAAoB;QAFH,cAAS,GAAT,SAAS,CAAQ;QACjB,eAAU,GAAV,UAAU,CAAmB;QAG9C,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,WAAW,IAAI,qBAAqB,CAAA;IAC7D,CAAC;IAED,oEAAoE;IACpE,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,SAAiB,EAAE,OAAuB,EAAE;QAC/D,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,CAAC,CAAA;QACpD,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAED,6CAA6C;IAC7C,MAAM,CAAC,KAAK,CAAC,SAAS,CAAC,UAAsB,EAAE,OAAuB,EAAE;QACvE,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,UAAU,EAAE,IAAI,CAAC,CAAA;QAC1D,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAEO,KAAK,CAAC,aAAa;QAC1B,IAAI,IAAI,CAAC,OAAO;YAAE,OAAO,IAAI,CAAC,OAAO,CAAA;QACrC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC;YACvB,IAAI,CAAC,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;gBAC9B,MAAM,KAAK,GAAG,IAAI,CAAC,UAAU,IAAI,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAA;gBAClF,MAAM,OAAO,GAAG,MAAM,GAAG,CAAC,gBAAgB,CAAC,MAAM,CAAC,KAAK,EAAE;oBACxD,kBAAkB,EAAE,CAAC,KAAK,CAAC;oBAC3B,sBAAsB,EAAE,KAAK;iBAC7B,CAAC,CAAA;gBACF,IAAI,CAAC,OAAO,GAAG,OAAO,CAAA;gBACtB,OAAO,OAAO,CAAA;YACf,CAAC,CAAC,EAAE,CAAA;QACL,CAAC;QACD,OAAO,IAAI,CAAC,WAAW,CAAA;IACxB,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,KAAK,CAAC,QAAkB;QAC7B,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,aAAa,EAAE,CAAA;QAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,CAAC,CAAA;QAC1D,MAAM,MAAM,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAClD,MAAM,IAAI,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAChD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAA;YAChC,IAAI,CAAC,CAAC,CAAC,GAAG,EAAE,CAAA;QACb,CAAC;QAED,MAAM,KAAK,GAA+B;YACzC,SAAS,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;YACjE,cAAc,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;SACpE,CAAA;QAED,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,CAAA;QACvC,MAAM,YAAY,GAAG,MAAM,CAAC,MAAM,CAAA;QAClC,IAAI,CAAC,YAAY;YAAE,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAA;QACjF,MAAM,IAAI,GAAG,YAAY,CAAC,IAAoB,CAAA;QAC9C,MAAM,CAAC,EAAE,AAAD,EAAG,SAAS,CAAC,GAAG,YAAY,CAAC,IAAyC,CAAA;QAE9E,MAAM,MAAM,GAAe,EAAE,CAAA;QAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,GAAG,GAAa,IAAI,KAAK,CAAC,SAAS,CAAC,CAAA;YAC1C,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAA;YAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE;gBAAE,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,GAAG,CAAC,CAAE,CAAA;YAC5D,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QACjB,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,CAAA;IAC7B,CAAC;CACD"}

package/out/proposal-classifier.d.ts

@@ -0,0 +1,36 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `NeuralProposalClassifier` — adapter that exposes a `NeuralAddressClassifier` as a
+ *   `ProposalClassifier` (the `@mailwoman/core/types` contract that the policy registry consumes).
+ *
+ *   Implementation: for each section, run the neural classifier on `section.body`, walk the resulting
+ *   `AddressTree`, and emit one `ClassificationProposal` per node whose tag is in the `emits` list.
+ *   Spans are rebased to the original input via `section.start + node.start` so downstream
+ *   consumers see character offsets in the caller's coordinate space — same convention as
+ *   `wrapLegacyClassifier`.
+ *
+ *   Per-section calls trade a small amount of context for the uniform `ProposalClassifier` shape.
+ *   Addresses inside a section are typically short and the model handles them well; whole-input
+ *   inference is a future optimization once the policy layer has a way to invoke a classifier "once
+ *   per parse" instead of per section.
+ */
+import type { ComponentTag, ProposalClassifier } from "@mailwoman/core/types";
+import type { NeuralAddressClassifier } from "./classifier.js";
+export interface NeuralProposalClassifierConfig {
+    /** Stable id surfaced as `source_id` on every proposal (e.g. `neural-v0.2.0-en-us`). */
+    id: string;
+    /** The underlying neural classifier instance. */
+    classifier: NeuralAddressClassifier;
+    /** Component tags this classifier may emit. Defaults to Stage 1 coarse tags. */
+    emits?: readonly ComponentTag[];
+    /** Locales this classifier is active for. `["*"]` (locale-agnostic) by default. */
+    locales?: readonly (string | "*")[];
+    /** Default penalty applied to emitted proposals. Default 0. */
+    penalty?: number;
+}
+/** Build a `ProposalClassifier` backed by a `NeuralAddressClassifier`. */
+export declare function createNeuralProposalClassifier(cfg: NeuralProposalClassifierConfig): ProposalClassifier;
+//# sourceMappingURL=proposal-classifier.d.ts.map

package/out/proposal-classifier.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"proposal-classifier.d.ts","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAIH,OAAO,KAAK,EAGX,YAAY,EACZ,kBAAkB,EAElB,MAAM,uBAAuB,CAAA;AAC9B,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,iBAAiB,CAAA;AAG9D,MAAM,WAAW,8BAA8B;IAC9C,wFAAwF;IACxF,EAAE,EAAE,MAAM,CAAA;IACV,iDAAiD;IACjD,UAAU,EAAE,uBAAuB,CAAA;IACnC,gFAAgF;IAChF,KAAK,CAAC,EAAE,SAAS,YAAY,EAAE,CAAA;IAC/B,mFAAmF;IACnF,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,GAAG,GAAG,CAAC,EAAE,CAAA;IACnC,+DAA+D;IAC/D,OAAO,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,0EAA0E;AAC1E,wBAAgB,8BAA8B,CAAC,GAAG,EAAE,8BAA8B,GAAG,kBAAkB,CA6CtG"}

package/out/proposal-classifier.js

@@ -0,0 +1,66 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `NeuralProposalClassifier` — adapter that exposes a `NeuralAddressClassifier` as a
+ *   `ProposalClassifier` (the `@mailwoman/core/types` contract that the policy registry consumes).
+ *
+ *   Implementation: for each section, run the neural classifier on `section.body`, walk the resulting
+ *   `AddressTree`, and emit one `ClassificationProposal` per node whose tag is in the `emits` list.
+ *   Spans are rebased to the original input via `section.start + node.start` so downstream
+ *   consumers see character offsets in the caller's coordinate space — same convention as
+ *   `wrapLegacyClassifier`.
+ *
+ *   Per-section calls trade a small amount of context for the uniform `ProposalClassifier` shape.
+ *   Addresses inside a section are typically short and the model handles them well; whole-input
+ *   inference is a future optimization once the policy layer has a way to invoke a classifier "once
+ *   per parse" instead of per section.
+ */
+import { STAGE1_COARSE_TAGS } from "./labels.js";
+/** Build a `ProposalClassifier` backed by a `NeuralAddressClassifier`. */
+export function createNeuralProposalClassifier(cfg) {
+    const emits = cfg.emits ?? STAGE1_COARSE_TAGS;
+    const emitsSet = new Set(emits);
+    const penalty = cfg.penalty ?? 0;
+    async function classify(section, _ctx) {
+        const tree = await cfg.classifier.parse(section.body);
+        const proposals = [];
+        const sectionOffset = section.start;
+        const visit = (node) => {
+            if (emitsSet.has(node.tag)) {
+                // Emit a structurally-Span-shaped record. We intentionally avoid `Span.from(...)` here:
+                // the tokenization module performs filesystem-bound module-init (libpostal data dir
+                // scan) which we don't want to force on every consumer of the proposal-classifier. The
+                // solver and policy registry read `start` / `end` / `body` only; if a downstream
+                // consumer needs the full Span behavior (graph membership, classifications, …), it
+                // should re-construct via Span.from(p.span.body, { start: p.span.start }).
+                const span = {
+                    start: sectionOffset + node.start,
+                    end: sectionOffset + node.end,
+                    body: node.value,
+                };
+                proposals.push({
+                    span,
+                    component: node.tag,
+                    confidence: node.confidence,
+                    source: "neural",
+                    source_id: cfg.id,
+                    penalty,
+                });
+            }
+            for (const child of node.children)
+                visit(child);
+        };
+        for (const root of tree.roots)
+            visit(root);
+        return proposals;
+    }
+    return {
+        id: cfg.id,
+        emits,
+        locales: cfg.locales ?? ["*"],
+        classify,
+    };
+}
+//# sourceMappingURL=proposal-classifier.js.map

package/out/proposal-classifier.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"proposal-classifier.js","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAYH,OAAO,EAAE,kBAAkB,EAAE,MAAM,aAAa,CAAA;AAehD,0EAA0E;AAC1E,MAAM,UAAU,8BAA8B,CAAC,GAAmC;IACjF,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,IAAI,kBAAkB,CAAA;IAC7C,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAe,KAAgC,CAAC,CAAA;IACxE,MAAM,OAAO,GAAG,GAAG,CAAC,OAAO,IAAI,CAAC,CAAA;IAEhC,KAAK,UAAU,QAAQ,CAAC,OAAgB,EAAE,IAAuB;QAChE,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,UAAU,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,CAAA;QACrD,MAAM,SAAS,GAA6B,EAAE,CAAA;QAC9C,MAAM,aAAa,GAAG,OAAO,CAAC,KAAK,CAAA;QAEnC,MAAM,KAAK,GAAG,CAAC,IAAiB,EAAQ,EAAE;YACzC,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC5B,wFAAwF;gBACxF,oFAAoF;gBACpF,uFAAuF;gBACvF,iFAAiF;gBACjF,mFAAmF;gBACnF,2EAA2E;gBAC3E,MAAM,IAAI,GAAG;oBACZ,KAAK,EAAE,aAAa,GAAG,IAAI,CAAC,KAAK;oBACjC,GAAG,EAAE,aAAa,GAAG,IAAI,CAAC,GAAG;oBAC7B,IAAI,EAAE,IAAI,CAAC,KAAK;iBACG,CAAA;gBACpB,SAAS,CAAC,IAAI,CAAC;oBACd,IAAI;oBACJ,SAAS,EAAE,IAAI,CAAC,GAAG;oBACnB,UAAU,EAAE,IAAI,CAAC,UAAU;oBAC3B,MAAM,EAAE,QAAQ;oBAChB,SAAS,EAAE,GAAG,CAAC,EAAE;oBACjB,OAAO;iBACP,CAAC,CAAA;YACH,CAAC;YACD,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,QAAQ;gBAAE,KAAK,CAAC,KAAK,CAAC,CAAA;QAChD,CAAC,CAAA;QAED,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,KAAK;YAAE,KAAK,CAAC,IAAI,CAAC,CAAA;QAC1C,OAAO,SAAS,CAAA;IACjB,CAAC;IAED,OAAO;QACN,EAAE,EAAE,GAAG,CAAC,EAAE;QACV,KAAK;QACL,OAAO,EAAE,GAAG,CAAC,OAAO,IAAI,CAAC,GAAG,CAAC;QAC7B,QAAQ;KACR,CAAA;AACF,CAAC"}

package/out/tokenizer.d.ts

@@ -0,0 +1,67 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   SentencePiece tokenizer wrapper with char-offset realignment.
+ *
+ *   `@sctg/sentencepiece-js` is a pure-JS port of the unigram SentencePiece algorithm and produces
+ *   pieces + ids but NOT offsets. The TS layer reconstructs offsets by walking the input string
+ *   alongside the emitted pieces; this gives us the `[start, end)` char ranges the BIO decoder
+ *   needs to map labels back to substrings.
+ *
+ *   Offset reconstruction algorithm:
+ *
+ *   - SentencePiece prepends `▁` (U+2581) to the first piece of each word (and to the first piece of
+ *       the input). Pieces without `▁` are continuations of the current word.
+ *   - When a piece starts with `▁`, consume any whitespace from the input before counting it.
+ *   - The piece's actual chars (the piece minus a leading `▁`) advance the input cursor by that many
+ *       code units. SentencePiece operates on Unicode codepoints, but since addresses are almost
+ *       entirely BMP characters, JS code-unit indexing is correct in practice. Surrogate- pair
+ *       codepoints would need `Array.from(s).length` accounting; deferred until the parity test
+ *       surfaces a real case.
+ *   - Byte-fallback pieces (`<0xHH>`) are not handled here. The v0.1.0 corpus and golden set are
+ *       Latin-script; the parity test will surface any unhandled cases.
+ *
+ *   The wrapper supports two load modes:
+ *
+ *   - `loadFromBase64(b64)` — for tests and Node usage where the model is read off disk and
+ *       base64-encoded before being handed to the JS port.
+ *   - `loadFromFile(path)` — convenience helper that does the read + b64 + load.
+ */
+/** SentencePiece's word-boundary marker (U+2581 LOWER ONE EIGHTH BLOCK). */
+export declare const SPACE_SENTINEL = "\u2581";
+/** A tokenized piece paired with its char-range in the original input. */
+export interface TokenizedPiece {
+    /** The piece exactly as the tokenizer emitted it (with `▁` preserved where present). */
+    piece: string;
+    /** The vocab id for this piece. */
+    id: number;
+    /** Inclusive start char offset in the original input. */
+    start: number;
+    /** Exclusive end char offset in the original input. */
+    end: number;
+}
+export interface EncodeResult {
+    pieces: TokenizedPiece[];
+    ids: number[];
+}
+export declare class MailwomanTokenizer {
+    private readonly processor;
+    private constructor();
+    /** Load from a base64-encoded `tokenizer.model`. Use for in-memory / test setups. */
+    static loadFromBase64(b64: string): Promise<MailwomanTokenizer>;
+    /** Load from a path to a `tokenizer.model` file on disk. Node-only convenience. */
+    static loadFromFile(modelPath: string): Promise<MailwomanTokenizer>;
+    /**
+     * Tokenize `text` to pieces + ids + realigned char offsets.
+     *
+     * The returned `pieces[i].piece` matches what the Python `sp.EncodeAsPieces(text)[i]` returns,
+     * and `pieces[i].id` matches `sp.EncodeAsIds(text)[i]`. The offsets are reconstructed in TS — see
+     * file header for the algorithm.
+     */
+    encode(text: string): EncodeResult;
+    /** Decode a list of ids back to a string. Delegates to the underlying processor. */
+    decode(ids: number[] | Int32Array): string;
+}
+//# sourceMappingURL=tokenizer.d.ts.map

package/out/tokenizer.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tokenizer.d.ts","sourceRoot":"","sources":["../tokenizer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAKH,4EAA4E;AAC5E,eAAO,MAAM,cAAc,WAAM,CAAA;AAEjC,0EAA0E;AAC1E,MAAM,WAAW,cAAc;IAC9B,wFAAwF;IACxF,KAAK,EAAE,MAAM,CAAA;IACb,mCAAmC;IACnC,EAAE,EAAE,MAAM,CAAA;IACV,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb,uDAAuD;IACvD,GAAG,EAAE,MAAM,CAAA;CACX;AAED,MAAM,WAAW,YAAY;IAC5B,MAAM,EAAE,cAAc,EAAE,CAAA;IACxB,GAAG,EAAE,MAAM,EAAE,CAAA;CACb;AAED,qBAAa,kBAAkB;IACV,OAAO,CAAC,QAAQ,CAAC,SAAS;IAA9C,OAAO;IAEP,qFAAqF;WACxE,cAAc,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC,kBAAkB,CAAC;IAMrE,mFAAmF;WACtE,YAAY,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,kBAAkB,CAAC;IAKzE;;;;;;OAMG;IACH,MAAM,CAAC,IAAI,EAAE,MAAM,GAAG,YAAY;IA2BlC,oFAAoF;IACpF,MAAM,CAAC,GAAG,EAAE,MAAM,EAAE,GAAG,UAAU,GAAG,MAAM;CAI1C"}

package/out/tokenizer.js

@@ -0,0 +1,86 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   SentencePiece tokenizer wrapper with char-offset realignment.
+ *
+ *   `@sctg/sentencepiece-js` is a pure-JS port of the unigram SentencePiece algorithm and produces
+ *   pieces + ids but NOT offsets. The TS layer reconstructs offsets by walking the input string
+ *   alongside the emitted pieces; this gives us the `[start, end)` char ranges the BIO decoder
+ *   needs to map labels back to substrings.
+ *
+ *   Offset reconstruction algorithm:
+ *
+ *   - SentencePiece prepends `▁` (U+2581) to the first piece of each word (and to the first piece of
+ *       the input). Pieces without `▁` are continuations of the current word.
+ *   - When a piece starts with `▁`, consume any whitespace from the input before counting it.
+ *   - The piece's actual chars (the piece minus a leading `▁`) advance the input cursor by that many
+ *       code units. SentencePiece operates on Unicode codepoints, but since addresses are almost
+ *       entirely BMP characters, JS code-unit indexing is correct in practice. Surrogate- pair
+ *       codepoints would need `Array.from(s).length` accounting; deferred until the parity test
+ *       surfaces a real case.
+ *   - Byte-fallback pieces (`<0xHH>`) are not handled here. The v0.1.0 corpus and golden set are
+ *       Latin-script; the parity test will surface any unhandled cases.
+ *
+ *   The wrapper supports two load modes:
+ *
+ *   - `loadFromBase64(b64)` — for tests and Node usage where the model is read off disk and
+ *       base64-encoded before being handed to the JS port.
+ *   - `loadFromFile(path)` — convenience helper that does the read + b64 + load.
+ */
+import { SentencePieceProcessor } from "@sctg/sentencepiece-js";
+import { promises as fs } from "node:fs";
+/** SentencePiece's word-boundary marker (U+2581 LOWER ONE EIGHTH BLOCK). */
+export const SPACE_SENTINEL = "▁";
+export class MailwomanTokenizer {
+    processor;
+    constructor(processor) {
+        this.processor = processor;
+    }
+    /** Load from a base64-encoded `tokenizer.model`. Use for in-memory / test setups. */
+    static async loadFromBase64(b64) {
+        const processor = new SentencePieceProcessor();
+        await processor.loadFromB64StringModel(b64);
+        return new MailwomanTokenizer(processor);
+    }
+    /** Load from a path to a `tokenizer.model` file on disk. Node-only convenience. */
+    static async loadFromFile(modelPath) {
+        const buf = await fs.readFile(modelPath);
+        return MailwomanTokenizer.loadFromBase64(buf.toString("base64"));
+    }
+    /**
+     * Tokenize `text` to pieces + ids + realigned char offsets.
+     *
+     * The returned `pieces[i].piece` matches what the Python `sp.EncodeAsPieces(text)[i]` returns,
+     * and `pieces[i].id` matches `sp.EncodeAsIds(text)[i]`. The offsets are reconstructed in TS — see
+     * file header for the algorithm.
+     */
+    encode(text) {
+        const pieces = this.processor.encodePieces(text);
+        const ids = this.processor.encodeIds(text);
+        const tokenized = [];
+        let cursor = 0;
+        for (let i = 0; i < pieces.length; i++) {
+            const piece = pieces[i];
+            const id = ids[i] ?? -1;
+            const hasSentinel = piece.startsWith(SPACE_SENTINEL);
+            const literal = hasSentinel ? piece.slice(SPACE_SENTINEL.length) : piece;
+            if (hasSentinel) {
+                while (cursor < text.length && /\s/.test(text[cursor]))
+                    cursor++;
+            }
+            const start = cursor;
+            cursor += literal.length;
+            const end = cursor;
+            tokenized.push({ piece, id, start, end });
+        }
+        return { pieces: tokenized, ids };
+    }
+    /** Decode a list of ids back to a string. Delegates to the underlying processor. */
+    decode(ids) {
+        const arr = ids instanceof Int32Array ? ids : Int32Array.from(ids);
+        return this.processor.decodeIds(arr);
+    }
+}
+//# sourceMappingURL=tokenizer.js.map

package/out/tokenizer.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tokenizer.js","sourceRoot":"","sources":["../tokenizer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAE,sBAAsB,EAAE,MAAM,wBAAwB,CAAA;AAC/D,OAAO,EAAE,QAAQ,IAAI,EAAE,EAAE,MAAM,SAAS,CAAA;AAExC,4EAA4E;AAC5E,MAAM,CAAC,MAAM,cAAc,GAAG,GAAG,CAAA;AAmBjC,MAAM,OAAO,kBAAkB;IACO;IAArC,YAAqC,SAAiC;QAAjC,cAAS,GAAT,SAAS,CAAwB;IAAG,CAAC;IAE1E,qFAAqF;IACrF,MAAM,CAAC,KAAK,CAAC,cAAc,CAAC,GAAW;QACtC,MAAM,SAAS,GAAG,IAAI,sBAAsB,EAAE,CAAA;QAC9C,MAAM,SAAS,CAAC,sBAAsB,CAAC,GAAG,CAAC,CAAA;QAC3C,OAAO,IAAI,kBAAkB,CAAC,SAAS,CAAC,CAAA;IACzC,CAAC;IAED,mFAAmF;IACnF,MAAM,CAAC,KAAK,CAAC,YAAY,CAAC,SAAiB;QAC1C,MAAM,GAAG,GAAG,MAAM,EAAE,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAA;QACxC,OAAO,kBAAkB,CAAC,cAAc,CAAC,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAA;IACjE,CAAC;IAED;;;;;;OAMG;IACH,MAAM,CAAC,IAAY;QAClB,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,IAAI,CAAC,CAAA;QAChD,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,IAAI,CAAC,CAAA;QAE1C,MAAM,SAAS,GAAqB,EAAE,CAAA;QACtC,IAAI,MAAM,GAAG,CAAC,CAAA;QAEd,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACxC,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACxB,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAA;YACvB,MAAM,WAAW,GAAG,KAAK,CAAC,UAAU,CAAC,cAAc,CAAC,CAAA;YACpD,MAAM,OAAO,GAAG,WAAW,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,KAAK,CAAA;YAExE,IAAI,WAAW,EAAE,CAAC;gBACjB,OAAO,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,CAAE,CAAC;oBAAE,MAAM,EAAE,CAAA;YAClE,CAAC;YAED,MAAM,KAAK,GAAG,MAAM,CAAA;YACpB,MAAM,IAAI,OAAO,CAAC,MAAM,CAAA;YACxB,MAAM,GAAG,GAAG,MAAM,CAAA;YAElB,SAAS,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,EAAE,EAAE,KAAK,EAAE,GAAG,EAAE,CAAC,CAAA;QAC1C,CAAC;QAED,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,GAAG,EAAE,CAAA;IAClC,CAAC;IAED,oFAAoF;IACpF,MAAM,CAAC,GAA0B;QAChC,MAAM,GAAG,GAAG,GAAG,YAAY,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QAClE,OAAO,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,GAAG,CAAW,CAAA;IAC/C,CAAC;CACD"}

package/out/vitest.config.d.ts

@@ -0,0 +1,11 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Per-package vitest config that resolves @mailwoman/core subpath imports to source. Mirrors the
+ *   layout used in core/vitest.config.ts.
+ */
+declare const _default: import("vite").UserConfig;
+export default _default;
+//# sourceMappingURL=vitest.config.d.ts.map

package/out/vitest.config.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"vitest.config.d.ts","sourceRoot":"","sources":["../vitest.config.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;;AAUH,wBAiBE"}

package/out/vitest.config.js

@@ -0,0 +1,32 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Per-package vitest config that resolves @mailwoman/core subpath imports to source. Mirrors the
+ *   layout used in core/vitest.config.ts.
+ */
+/// <reference types="vitest/config" />
+import { resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+import { defineConfig } from "vite";
+const here = fileURLToPath(new URL(".", import.meta.url));
+export default defineConfig({
+    resolve: {
+        alias: [
+            // Sub-subpath alias for the pure proposal-pipeline module — avoids dragging in
+            // AddressParser → classification → tokenization → libpostal init cascade.
+            {
+                find: "@mailwoman/core/parser/proposal-pipeline",
+                replacement: resolve(here, "../core/parser/proposal-pipeline.ts"),
+            },
+            { find: /^@mailwoman\/core\/(.+)$/, replacement: resolve(here, "../core/$1/index.ts") },
+            { find: /^@mailwoman\/core$/, replacement: resolve(here, "../core/index.ts") },
+        ],
+    },
+    test: {
+        isolate: false,
+        exclude: ["**/node_modules/**", "**/out/**", "**/dist/**"],
+    },
+});
+//# sourceMappingURL=vitest.config.js.map

package/out/vitest.config.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"vitest.config.js","sourceRoot":"","sources":["../vitest.config.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,uCAAuC;AAEvC,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAA;AACnC,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAA;AACxC,OAAO,EAAE,YAAY,EAAE,MAAM,MAAM,CAAA;AAEnC,MAAM,IAAI,GAAG,aAAa,CAAC,IAAI,GAAG,CAAC,GAAG,EAAE,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAA;AAEzD,eAAe,YAAY,CAAC;IAC3B,OAAO,EAAE;QACR,KAAK,EAAE;YACN,+EAA+E;YAC/E,0EAA0E;YAC1E;gBACC,IAAI,EAAE,0CAA0C;gBAChD,WAAW,EAAE,OAAO,CAAC,IAAI,EAAE,qCAAqC,CAAC;aACjE;YACD,EAAE,IAAI,EAAE,0BAA0B,EAAE,WAAW,EAAE,OAAO,CAAC,IAAI,EAAE,qBAAqB,CAAC,EAAE;YACvF,EAAE,IAAI,EAAE,oBAAoB,EAAE,WAAW,EAAE,OAAO,CAAC,IAAI,EAAE,kBAAkB,CAAC,EAAE;SAC9E;KACD;IACD,IAAI,EAAE;QACL,OAAO,EAAE,KAAK;QACd,OAAO,EAAE,CAAC,oBAAoB,EAAE,WAAW,EAAE,YAAY,CAAC;KAC1D;CACD,CAAC,CAAA"}

package/out/weights.d.ts

@@ -0,0 +1,39 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Weight-package resolution.
+ *
+ *   The `@mailwoman/neural-weights-<locale>` packages ship the `model.onnx` + `tokenizer.model` files
+ *   declared in their `files` array. At install time npm bundles those files alongside the
+ *   package.json; at runtime we locate them by resolving the package.json then walking sideways.
+ *
+ *   Local development gotcha: the weights packages in the monorepo carry only metadata (package.json
+ *
+ *   - README.md + model-card.json). The actual binary files are produced by Phase 2 training and copied
+ *       in at publish time. To run the neural classifier locally without publishing, either:
+ *
+ *   1. Pass explicit `modelPath` + `tokenizerPath` to `loadFromWeights`, or
+ *   2. Symlink the dev model files into the weights package directory — see
+ *        `scripts/link-dev-weights.sh` in each weights package.
+ *
+ *   The resolver checks for both files and throws a single actionable error when neither is findable,
+ *   naming all the paths it tried.
+ */
+export interface ResolveWeightsOpts {
+    /** BCP-47-ish locale tag, e.g. "en-us" or "fr-fr". Used to pick the weights package. */
+    locale?: string;
+    /** Explicit model.onnx path; takes precedence over package auto-resolve. */
+    modelPath?: string;
+    /** Explicit tokenizer.model path; takes precedence over package auto-resolve. */
+    tokenizerPath?: string;
+}
+export interface ResolvedWeights {
+    modelPath: string;
+    tokenizerPath: string;
+    /** "explicit" if both paths came from opts; "package:<name>" if resolved via require.resolve. */
+    source: string;
+}
+export declare function resolveWeights(opts: ResolveWeightsOpts): ResolvedWeights;
+//# sourceMappingURL=weights.d.ts.map

package/out/weights.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"weights.d.ts","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAQH,MAAM,WAAW,kBAAkB;IAClC,wFAAwF;IACxF,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,iFAAiF;IACjF,aAAa,CAAC,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,eAAe;IAC/B,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB,iGAAiG;IACjG,MAAM,EAAE,MAAM,CAAA;CACd;AAED,wBAAgB,cAAc,CAAC,IAAI,EAAE,kBAAkB,GAAG,eAAe,CAoCxE"}

package/out/weights.js

@@ -0,0 +1,59 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Weight-package resolution.
+ *
+ *   The `@mailwoman/neural-weights-<locale>` packages ship the `model.onnx` + `tokenizer.model` files
+ *   declared in their `files` array. At install time npm bundles those files alongside the
+ *   package.json; at runtime we locate them by resolving the package.json then walking sideways.
+ *
+ *   Local development gotcha: the weights packages in the monorepo carry only metadata (package.json
+ *
+ *   - README.md + model-card.json). The actual binary files are produced by Phase 2 training and copied
+ *       in at publish time. To run the neural classifier locally without publishing, either:
+ *
+ *   1. Pass explicit `modelPath` + `tokenizerPath` to `loadFromWeights`, or
+ *   2. Symlink the dev model files into the weights package directory — see
+ *        `scripts/link-dev-weights.sh` in each weights package.
+ *
+ *   The resolver checks for both files and throws a single actionable error when neither is findable,
+ *   naming all the paths it tried.
+ */
+import { existsSync } from "node:fs";
+import { createRequire } from "node:module";
+import { dirname, resolve } from "node:path";
+const req = createRequire(import.meta.url);
+export function resolveWeights(opts) {
+    const tried = [];
+    if (opts.modelPath && opts.tokenizerPath) {
+        if (!existsSync(opts.modelPath))
+            throw new Error(`Explicit modelPath does not exist: ${opts.modelPath}`);
+        if (!existsSync(opts.tokenizerPath))
+            throw new Error(`Explicit tokenizerPath does not exist: ${opts.tokenizerPath}`);
+        return { modelPath: opts.modelPath, tokenizerPath: opts.tokenizerPath, source: "explicit" };
+    }
+    const locale = opts.locale ?? "en-us";
+    const packageName = `@mailwoman/neural-weights-${locale}`;
+    let packageDir;
+    try {
+        const pkgJsonPath = req.resolve(`${packageName}/package.json`);
+        packageDir = dirname(pkgJsonPath);
+    }
+    catch {
+        throw new Error(`Could not resolve ${packageName}. Install it via: npm install ${packageName}\n` +
+            `Or pass --model + --tokenizer with explicit paths.`);
+    }
+    const modelPath = opts.modelPath ?? resolve(packageDir, "model.onnx");
+    const tokenizerPath = opts.tokenizerPath ?? resolve(packageDir, "tokenizer.model");
+    tried.push(modelPath, tokenizerPath);
+    if (!existsSync(modelPath) || !existsSync(tokenizerPath)) {
+        throw new Error(`Weights package ${packageName} resolved at ${packageDir} but is missing model files.\n` +
+            `Tried:\n  ${tried.join("\n  ")}\n` +
+            `Run \`scripts/link-dev-weights.sh\` inside the package to symlink dev weights, ` +
+            `or pass --model + --tokenizer with explicit paths.`);
+    }
+    return { modelPath, tokenizerPath, source: `package:${packageName}` };
+}
+//# sourceMappingURL=weights.js.map

package/out/weights.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"weights.js","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAA;AACpC,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAA;AAC3C,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAA;AAE5C,MAAM,GAAG,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;AAkB1C,MAAM,UAAU,cAAc,CAAC,IAAwB;IACtD,MAAM,KAAK,GAAa,EAAE,CAAA;IAE1B,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;QAC1C,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,IAAI,CAAC,SAAS,EAAE,CAAC,CAAA;QACxG,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,0CAA0C,IAAI,CAAC,aAAa,EAAE,CAAC,CAAA;QACpH,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,SAAS,EAAE,aAAa,EAAE,IAAI,CAAC,aAAa,EAAE,MAAM,EAAE,UAAU,EAAE,CAAA;IAC5F,CAAC;IAED,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,OAAO,CAAA;IACrC,MAAM,WAAW,GAAG,6BAA6B,MAAM,EAAE,CAAA;IACzD,IAAI,UAAkB,CAAA;IACtB,IAAI,CAAC;QACJ,MAAM,WAAW,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,WAAW,eAAe,CAAC,CAAA;QAC9D,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,CAAA;IAClC,CAAC;IAAC,MAAM,CAAC;QACR,MAAM,IAAI,KAAK,CACd,qBAAqB,WAAW,iCAAiC,WAAW,IAAI;YAC/E,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,UAAU,EAAE,YAAY,CAAC,CAAA;IACrE,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,OAAO,CAAC,UAAU,EAAE,iBAAiB,CAAC,CAAA;IAClF,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,CAAA;IAEpC,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC,EAAE,CAAC;QAC1D,MAAM,IAAI,KAAK,CACd,mBAAmB,WAAW,gBAAgB,UAAU,gCAAgC;YACvF,aAAa,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI;YACnC,iFAAiF;YACjF,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,MAAM,EAAE,WAAW,WAAW,EAAE,EAAE,CAAA;AACtE,CAAC"}

package/package.json

@@ -0,0 +1,31 @@
+{
+ "name": "@mailwoman/neural",
+ "version": "2.0.0",
+ "description": "Mailwoman neural classifier runtime: SentencePiece tokenizer + ONNX inference + decoder wiring.",
+ "license": "AGPL-3.0-only",
+ "repository": {
+  "type": "git",
+  "url": "https://github.com/sister-software/mailwoman.git",
+  "directory": "neural"
+ },
+ "type": "module",
+ "exports": {
+  "./package.json": "./package.json",
+  ".": "./out/index.js",
+  "./tokenizer": "./out/tokenizer.js"
+ },
+ "dependencies": {
+  "@mailwoman/core": "workspace:*",
+  "@sctg/sentencepiece-js": "^1.3.3",
+  "onnxruntime-node": "^1.26.0"
+ },
+ "files": [
+  "out/**/*.js",
+  "out/**/*.js.map",
+  "out/**/*.d.ts",
+  "out/**/*.d.ts.map"
+ ],
+ "publishConfig": {
+  "access": "public"
+ }
+}