@workglow/huggingface-transformers 0.2.35 → 0.2.37

package/dist/ai.js

@@ -736,6 +736,7 @@ var HFT_TEXT_CLASSIFICATION = ["text.classification"];
 var HFT_TEXT_LANGUAGE_DETECTION = [
   "text.language-detection"
 ];
+var HFT_TEXT_RERANKING = ["text.reranking"];
 var HFT_TEXT_FILL_MASK = ["text.fill-mask"];
 var HFT_TEXT_NER = ["text.ner"];
 var HFT_IMAGE_CLASSIFICATION = ["image.classification"];
@@ -764,6 +765,7 @@ var HFT_CAPABILITY_SETS = [
   HFT_TEXT_EMBEDDING,
   HFT_TEXT_CLASSIFICATION,
   HFT_TEXT_LANGUAGE_DETECTION,
+  HFT_TEXT_RERANKING,
   HFT_TEXT_FILL_MASK,
   HFT_TEXT_NER,
   HFT_IMAGE_CLASSIFICATION,
@@ -985,21 +987,6 @@ var HFT_BackgroundRemoval = async (input, model, signal, emit) => {
 // src/ai/common/HFT_Chat.ts
 init_HFT_Pipeline();
 
-// src/ai/common/HFT_ToolCalling.ts
-init_HFT_Pipeline();
-import {
-  buildToolDescription,
-  filterValidToolCalls,
-  toTextFlatMessages
-} from "@workglow/ai/worker";
-import {
-  adaptParserResult,
-  forcedToolSelection,
-  getAvailableParsers,
-  getGenerationPrefix,
-  parseToolCalls
-} from "@workglow/ai/provider-utils";
-
 // src/ai/common/HFT_Streaming.ts
 function createStreamingTextStreamer(tokenizer, onText, textStreamer) {
   return new textStreamer(tokenizer, {
@@ -1023,6 +1010,19 @@ function createTextStreamer(tokenizer, updateProgress, textStreamer) {
 }
 
 // src/ai/common/HFT_ToolCalling.ts
+init_HFT_Pipeline();
+import {
+  adaptParserResult,
+  forcedToolSelection,
+  getAvailableParsers,
+  getGenerationPrefix,
+  parseToolCalls
+} from "@workglow/ai/provider-utils";
+import {
+  buildToolDescription,
+  filterValidToolCalls,
+  toTextFlatMessages
+} from "@workglow/ai/worker";
 function getModelTextCandidates(model) {
   return [model.model_id, model.title, model.description, model.provider_config.model_path].filter((value) => typeof value === "string" && value.length > 0).map((value) => value.toLowerCase());
 }
@@ -1442,8 +1442,8 @@ var HFT_ImageClassification = async (input, model, signal, emit) => {
 
 // src/ai/common/HFT_ImageEmbedding.ts
 init_HFT_Pipeline();
-import { getLogger as getLogger3 } from "@workglow/util/worker";
 import { imageValueToBlob as imageValueToBlob3 } from "@workglow/ai/provider-utils";
+import { getLogger as getLogger3 } from "@workglow/util/worker";
 var HFT_ImageEmbedding = async (input, model, signal, emit) => {
   const logger = getLogger3();
   const timerLabel = `hft:ImageEmbedding:${model?.provider_config.model_path}`;
@@ -1475,6 +1475,29 @@ var HFT_ImageEmbedding = async (input, model, signal, emit) => {
 // src/ai/common/HFT_ImageSegmentation.ts
 init_HFT_Pipeline();
 import { imageValueToBlob as imageValueToBlob4 } from "@workglow/ai/provider-utils";
+import { imageValueFromBuffer } from "@workglow/util/media";
+function rawImageToImageValue(image) {
+  const raw = image;
+  const pixelCount = raw.width * raw.height;
+  const rgba = new Uint8Array(pixelCount * 4);
+  for (let pixel = 0;pixel < pixelCount; pixel++) {
+    const sourceOffset = pixel * raw.channels;
+    const targetOffset = pixel * 4;
+    if (raw.channels === 1) {
+      const value = raw.data[sourceOffset] ?? 0;
+      rgba[targetOffset] = value;
+      rgba[targetOffset + 1] = value;
+      rgba[targetOffset + 2] = value;
+      rgba[targetOffset + 3] = 255;
+    } else {
+      rgba[targetOffset] = raw.data[sourceOffset] ?? 0;
+      rgba[targetOffset + 1] = raw.data[sourceOffset + 1] ?? raw.data[sourceOffset] ?? 0;
+      rgba[targetOffset + 2] = raw.data[sourceOffset + 2] ?? raw.data[sourceOffset] ?? 0;
+      rgba[targetOffset + 3] = raw.channels >= 4 ? raw.data[sourceOffset + 3] ?? 255 : 255;
+    }
+  }
+  return imageValueFromBuffer(Buffer.from(rgba), "raw-rgba", raw.width, raw.height);
+}
 var HFT_ImageSegmentation = async (input, model, signal, emit) => {
   const segmenter = await getPipeline(model, emit, {}, signal);
   const imageArg = await imageValueToBlob4(input.image);
@@ -1483,11 +1506,13 @@ var HFT_ImageSegmentation = async (input, model, signal, emit) => {
     mask_threshold: input.maskThreshold
   });
   const masks = Array.isArray(result) ? result : [result];
-  const processedMasks = await Promise.all(masks.map(async (mask) => ({
-    label: mask.label || "",
-    score: mask.score || 0,
-    mask: {}
-  })));
+  const processedMasks = masks.map((mask) => {
+    return {
+      label: mask.label || "",
+      score: mask.score || 0,
+      mask: rawImageToImageValue(mask.mask)
+    };
+  });
   emit({
     type: "finish",
     data: {
@@ -1614,7 +1639,7 @@ var HFT_ModelInfo = async (input, model, _signal, emit) => {
 };
 
 // src/ai/common/HFT_ModelSearch.ts
-import { searchHfModels, mapHfModelResult } from "@workglow/ai/provider-utils";
+import { mapHfModelResult, searchHfModels } from "@workglow/ai/provider-utils";
 var HFT_ModelSearch = async (input, _model, signal, emit) => {
   const entries = await searchHfModels(input.query?.trim() ?? "", { filter: "onnx" }, ["siblings"], signal);
   const results = entries.map((entry) => {
@@ -1959,6 +1984,54 @@ var HFT_TextQuestionAnswer = async (input, model, signal, emit) => {
   emit({ type: "finish", data: { text: answerText } });
 };
 
+// src/ai/common/HFT_TextReranker.ts
+init_HFT_Pipeline();
+import { KbRerankerOutputError } from "@workglow/ai";
+import { getLogger as getLogger6 } from "@workglow/util/worker";
+function isScored(v) {
+  return typeof v === "object" && v !== null && typeof v.score === "number";
+}
+function truncateShape(value) {
+  try {
+    const json = JSON.stringify(value);
+    if (typeof json !== "string")
+      return String(value);
+    return json.length > 200 ? `${json.slice(0, 200)}…` : json;
+  } catch {
+    return String(value);
+  }
+}
+function validateAndExtractRerankerScores(rawResults, modelPath) {
+  if (!Array.isArray(rawResults)) {
+    throw new KbRerankerOutputError(`HFT_TextReranker: unexpected pipeline output shape for model ${modelPath}`, truncateShape(rawResults));
+  }
+  const scores = new Array(rawResults.length);
+  for (let i = 0;i < rawResults.length; i++) {
+    const entry = rawResults[i];
+    const candidate = Array.isArray(entry) ? entry[0] : entry;
+    if (!isScored(candidate)) {
+      throw new KbRerankerOutputError(`HFT_TextReranker: unexpected pipeline output shape for model ${modelPath}`, truncateShape(entry));
+    }
+    scores[i] = candidate.score;
+  }
+  return scores;
+}
+var HFT_TextReranker = async (input, model, signal, emit) => {
+  const logger = getLogger6();
+  const modelPath = model?.provider_config.model_path;
+  const timerLabel = `hft:TextReranker:${modelPath}`;
+  logger.time(timerLabel, { docs: input.documents.length });
+  const reranker = await getPipeline(model, emit, {}, signal);
+  const pairs = input.documents.map((doc) => ({ text: input.query, text_pair: doc }));
+  const callable = reranker;
+  const rawResults = await callable(pairs, { top_k: 1 });
+  const scores = validateAndExtractRerankerScores(rawResults, modelPath);
+  const indices = scores.map((score, idx) => ({ score, idx })).sort((a, b) => b.score - a.score).map((p) => p.idx);
+  const limited = typeof input.topK === "number" ? indices.slice(0, input.topK) : indices;
+  logger.timeEnd(timerLabel, { docs: input.documents.length });
+  emit({ type: "finish", data: { scores, indices: limited } });
+};
+
 // src/ai/common/HFT_TextRewriter.ts
 init_HFT_Pipeline();
 var HFT_TextRewriter = async (input, model, signal, emit) => {
@@ -2034,6 +2107,7 @@ var HFT_RUN_FNS = [
   { serves: HFT_TEXT_EMBEDDING, runFn: HFT_TextEmbedding },
   { serves: HFT_TEXT_CLASSIFICATION, runFn: HFT_TextClassification },
   { serves: HFT_TEXT_LANGUAGE_DETECTION, runFn: HFT_TextLanguageDetection },
+  { serves: HFT_TEXT_RERANKING, runFn: HFT_TextReranker },
   { serves: HFT_TEXT_FILL_MASK, runFn: HFT_TextFillMask },
   { serves: HFT_TEXT_NER, runFn: HFT_TextNamedEntityRecognition },
   { serves: HFT_IMAGE_CLASSIFICATION, runFn: HFT_ImageClassification },
@@ -2076,4 +2150,4 @@ export {
   HF_TRANSFORMERS_ONNX
 };
 
-//# debugId=F0665A2512C7736664756E2164756E21
+//# debugId=F1F15E5800BCD03B64756E2164756E21