@mastra/fastembed 1.3.5-alpha.2 → 1.3.5

package/dist/index.js

@@ -18352,6 +18352,7 @@ let SparseEmbeddingModel = /* @__PURE__ */ function(SparseEmbeddingModel) {
 	return SparseEmbeddingModel;
 }({});
 function normalize(v) {
+	if (v.some((val) => !Number.isFinite(val))) throw new Error("Model produced a non-finite dense embedding value");
 	const norm = Math.sqrt(v.reduce((acc, val) => acc + val * val, 0));
 	const epsilon = 1e-12;
 	return v.map((val) => val / Math.max(norm, epsilon));
@@ -18366,7 +18367,7 @@ function getEmbeddings(data, dimensions, masks) {
 		for (let tokenIndex = 0; tokenIndex < sequenceLength; tokenIndex++) {
 			if (masks[batchIndex][tokenIndex] === 0n) continue;
 			tokenCount++;
-			for (let dimension = 0; dimension < hiddenSize; dimension++) embedding[dimension] = embedding[dimension] + data[start + tokenIndex * hiddenSize + dimension];
+			for (let dimension = 0; dimension < hiddenSize; dimension++) embedding[dimension] += data[start + tokenIndex * hiddenSize + dimension];
 		}
 		return embedding.map((value) => value / Math.max(tokenCount, 1));
 	});
@@ -18414,6 +18415,11 @@ function assertValidBatchSize(batchSize) {
 function assertValidMaxLength(maxLength) {
 	if (!Number.isSafeInteger(maxLength) || maxLength <= 0) throw new RangeError(`maxLength must be a positive safe integer, received ${String(maxLength)}`);
 }
+function getRequiredOutput(outputs, outputName) {
+	const tensor = outputs[outputName];
+	if (!tensor) throw new Error(`Expected model output "${outputName}" was not found. Available outputs: ${Object.keys(outputs).join(", ") || "(none)"}`);
+	return tensor;
+}
 function getOutputDimensions(outputName, tensor) {
 	if (tensor.dims.length !== 3) throw new Error(`Expected model output "${outputName}" to have 3 dimensions, received [${tensor.dims.join(", ")}]`);
 	return tensor.dims;
@@ -18527,7 +18533,7 @@ var FlagEmbedding = class FlagEmbedding extends Embedding {
 				token_type_ids: new ort.Tensor("int64", typeIdsArray.flat(), [batchTexts.length, maxLength])
 			};
 			if (this.model === "fast-multilingual-e5-large") delete inputs.token_type_ids;
-			const lastHiddenState = (await this.session.run(inputs)).last_hidden_state;
+			const lastHiddenState = getRequiredOutput(await this.session.run(inputs), "last_hidden_state");
 			yield getEmbeddings(lastHiddenState.data, getOutputDimensions("last_hidden_state", lastHiddenState), this.model === "fast-all-MiniLM-L6-v2" || this.model === "fast-multilingual-e5-large" ? maskArray : void 0).map(normalize);
 		}
 	}
@@ -18669,14 +18675,16 @@ var SparseTextEmbedding = class SparseTextEmbedding extends SparseEmbedding {
 				input_mask: new ort.Tensor("int64", maskArray.flat().map(BigInt), [batchTexts.length, maxLength]),
 				segment_ids: new ort.Tensor("int64", typeIdsArray.flat(), [batchTexts.length, maxLength])
 			};
-			const outputTensor = (await this.session.run(inputs)).output;
+			const outputTensor = getRequiredOutput(await this.session.run(inputs), "output");
 			const logits = outputTensor.cpuData;
 			const [currentBatchSize, seqLen, currentVocabSize] = getOutputDimensions("output", outputTensor);
 			const sparseVectors = [];
 			for (let batchIdx = 0; batchIdx < currentBatchSize; batchIdx++) {
 				const valuesArr = new Float32Array(currentVocabSize).fill(0);
 				for (let seqIdx = 0; seqIdx < seqLen; seqIdx++) if (maskArray[batchIdx][seqIdx] > 0) for (let vocabIdx = 0; vocabIdx < currentVocabSize; vocabIdx++) {
-					const logitValue = logits[batchIdx * seqLen * currentVocabSize + seqIdx * currentVocabSize + vocabIdx];
+					const logitIdx = batchIdx * seqLen * currentVocabSize + seqIdx * currentVocabSize + vocabIdx;
+					const logitValue = logits[logitIdx];
+					if (!Number.isFinite(logitValue)) throw new Error(`Model produced a non-finite sparse embedding value at index ${logitIdx}`);
 					const logValue = Math.log(1 + Math.max(0, logitValue));
 					valuesArr[vocabIdx] = Math.max(valuesArr[vocabIdx], logValue);
 				}