warpvector 0.1.1 → 0.1.2

package/dist/index.d.mts

@@ -533,6 +533,7 @@ declare class MlpAdapter implements WarpAdapter {
     private layerDimsPtr;
     private activationsPtr;
     private bufferPtr;
+    private bufBPtr;
     constructor(layers: MlpLayer[]);
     /**
      * WASMの初期化と、MLP構造をWASMメモリに書き込む準備を行います。
@@ -585,6 +586,10 @@ declare class WhiteningAdapter implements WarpAdapter {
     private count;
     private learningRate;
     private numComponents;
+    private componentsPtr;
+    private xResidualPtr;
+    private isWasmReady;
+    init(): Promise<void>;
     /**
      * 新しい WhiteningAdapter を作成します。
      * @param dim ベクトルの次元数
@@ -837,6 +842,12 @@ interface QuantizationConfig {
      * 量子化するベクトルの次元数
      */
     dim: number;
+    /**
+     * 動的キャリブレーション（スケール算出）を有効にするか。
+     * trueの場合、ベクトルの絶対値の最大値 max(abs(v)) を算出し、
+     * scale = 127 / max(abs(v)) を用いて量子化します。
+     */
+    dynamic?: boolean;
 }
 /**
  * QuantizationAdapter は、Float32のベクトルを Int8 や Binary に圧縮し、
@@ -845,6 +856,7 @@ interface QuantizationConfig {
 declare class QuantizationAdapter implements WarpAdapter {
     private type;
     private dim;
+    private dynamic;
     constructor(config: QuantizationConfig);
     tune(vector: number[] | Float32Array): Int8Array | Uint8Array;
     /**
@@ -854,6 +866,7 @@ declare class QuantizationAdapter implements WarpAdapter {
     static hammingDistance(a: Uint8Array, b: Uint8Array): number;
     /**
      * Int8量子化された2つのベクトル間のドット積（内積）を計算します。
+     * 動的スケーリングが埋め込まれている場合はスケールを戻して計算します。
      */
     static int8DotProduct(a: Int8Array, b: Int8Array): number;
     exportState(): string;

package/dist/index.d.ts

@@ -533,6 +533,7 @@ declare class MlpAdapter implements WarpAdapter {
     private layerDimsPtr;
     private activationsPtr;
     private bufferPtr;
+    private bufBPtr;
     constructor(layers: MlpLayer[]);
     /**
      * WASMの初期化と、MLP構造をWASMメモリに書き込む準備を行います。
@@ -585,6 +586,10 @@ declare class WhiteningAdapter implements WarpAdapter {
     private count;
     private learningRate;
     private numComponents;
+    private componentsPtr;
+    private xResidualPtr;
+    private isWasmReady;
+    init(): Promise<void>;
     /**
      * 新しい WhiteningAdapter を作成します。
      * @param dim ベクトルの次元数
@@ -837,6 +842,12 @@ interface QuantizationConfig {
      * 量子化するベクトルの次元数
      */
     dim: number;
+    /**
+     * 動的キャリブレーション（スケール算出）を有効にするか。
+     * trueの場合、ベクトルの絶対値の最大値 max(abs(v)) を算出し、
+     * scale = 127 / max(abs(v)) を用いて量子化します。
+     */
+    dynamic?: boolean;
 }
 /**
  * QuantizationAdapter は、Float32のベクトルを Int8 や Binary に圧縮し、
@@ -845,6 +856,7 @@ interface QuantizationConfig {
 declare class QuantizationAdapter implements WarpAdapter {
     private type;
     private dim;
+    private dynamic;
     constructor(config: QuantizationConfig);
     tune(vector: number[] | Float32Array): Int8Array | Uint8Array;
     /**
@@ -854,6 +866,7 @@ declare class QuantizationAdapter implements WarpAdapter {
     static hammingDistance(a: Uint8Array, b: Uint8Array): number;
     /**
      * Int8量子化された2つのベクトル間のドット積（内積）を計算します。
+     * 動的スケーリングが埋め込まれている場合はスケールを戻して計算します。
      */
     static int8DotProduct(a: Int8Array, b: Int8Array): number;
     exportState(): string;

package/dist/index.js

@@ -829,7 +829,7 @@ function applyAffine(matrix, bias, vector, result, inDim, outDim = inDim) {
 }
 
 // src/wasm/wasm-binary.ts
-var wasmBase64 = "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";
+var wasmBase64 = "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";
 
 // src/wasm/wasm-loader.ts
 var wasmInstance = null;
@@ -867,6 +867,22 @@ function ensureWasmMemory(requiredBytes) {
     return false;
   }
 }
+var globalOffset = 65536;
+function allocateWasmMemory(bytes) {
+  const ptr = globalOffset;
+  globalOffset += bytes;
+  if (globalOffset % 4 !== 0) {
+    globalOffset += 4 - globalOffset % 4;
+  }
+  ensureWasmMemory(globalOffset);
+  return ptr;
+}
+function getWasmAllocatorOffset() {
+  return globalOffset;
+}
+function setWasmAllocatorOffset(offset) {
+  globalOffset = offset;
+}
 function writeFloat32ArrayToWasm(memory, data, byteOffset) {
   const f32 = new Float32Array(memory.buffer);
   const floatOffset = byteOffset / 4;
@@ -1706,6 +1722,7 @@ var MlpAdapter = class _MlpAdapter {
   layerDimsPtr = 0;
   activationsPtr = 0;
   bufferPtr = 0;
+  bufBPtr = 0;
   constructor(layers) {
     if (layers.length === 0) {
       throw new Error("MlpAdapter requires at least one layer.");
@@ -1756,21 +1773,13 @@ var MlpAdapter = class _MlpAdapter {
     }
     this.inputDim = layerDims[0];
     this.outputDim = layerDims[layerDims.length - 1];
-    let offset = 65536;
-    this.inputPtr = offset;
-    offset += this.inputDim * 4;
-    this.outputPtr = offset;
-    offset += this.outputDim * 4;
-    this.layerDimsPtr = offset;
-    offset += (this.numLayers + 1) * 4;
-    this.activationsPtr = offset;
-    offset += this.numLayers * 4;
-    if (offset % 4 !== 0) offset += 4 - offset % 4;
-    this.weightsPtr = offset;
-    offset += totalWeights * 4;
-    this.bufferPtr = offset;
-    offset += maxDim * 8 + 8192;
-    ensureWasmMemory(offset);
+    this.inputPtr = allocateWasmMemory(this.inputDim * 4);
+    this.outputPtr = allocateWasmMemory(this.outputDim * 4);
+    this.layerDimsPtr = allocateWasmMemory((this.numLayers + 1) * 4);
+    this.activationsPtr = allocateWasmMemory(this.numLayers * 4);
+    this.weightsPtr = allocateWasmMemory(totalWeights * 4);
+    this.bufferPtr = allocateWasmMemory(maxDim * 4);
+    this.bufBPtr = allocateWasmMemory(maxDim * 4);
     const memoryBuffer = memory.buffer;
     const f32 = new Float32Array(memoryBuffer);
     const i32 = new Int32Array(memoryBuffer);
@@ -1830,7 +1839,8 @@ var MlpAdapter = class _MlpAdapter {
       this.layerDimsPtr,
       this.activationsPtr,
       this.numLayers,
-      this.bufferPtr
+      this.bufferPtr,
+      this.bufBPtr
     );
     const result = new Float32Array(this.outputDim);
     const outF32 = new Float32Array(memory.buffer);
@@ -1884,6 +1894,19 @@ var WhiteningAdapter = class _WhiteningAdapter {
   count = 0;
   learningRate;
   numComponents;
+  componentsPtr = 0;
+  xResidualPtr = 0;
+  isWasmReady = false;
+  async init() {
+    const instance = await initWasm();
+    if (instance) {
+      const componentsSize = this.numComponents * this.dim * 4;
+      const xResidualSize = this.dim * 4;
+      this.componentsPtr = allocateWasmMemory(componentsSize);
+      this.xResidualPtr = allocateWasmMemory(xResidualSize);
+      this.isWasmReady = true;
+    }
+  }
   /**
    * 新しい WhiteningAdapter を作成します。
    * @param dim ベクトルの次元数
@@ -1927,13 +1950,17 @@ var WhiteningAdapter = class _WhiteningAdapter {
     }
     const x_residual = new Float32Array(x);
     const instance = getWasmInstance();
-    const componentsSize = this.numComponents * this.dim * 4;
-    const xResidualSize = this.dim * 4;
-    const requiredBytes = componentsSize + xResidualSize;
-    if (instance && instance.exports.sangerUpdateWasm && ensureWasmMemory(requiredBytes)) {
+    if (instance && instance.exports.sangerUpdateWasm) {
+      if (!this.isWasmReady) {
+        const componentsSize = this.numComponents * this.dim * 4;
+        const xResidualSize = this.dim * 4;
+        this.componentsPtr = allocateWasmMemory(componentsSize);
+        this.xResidualPtr = allocateWasmMemory(xResidualSize);
+        this.isWasmReady = true;
+      }
       const memory = instance.exports.memory;
-      const componentsPtr = 0;
-      const xResidualPtr = componentsSize;
+      const componentsPtr = this.componentsPtr;
+      const xResidualPtr = this.xResidualPtr;
       const f32 = new Float32Array(memory.buffer);
       for (let k = 0; k < this.numComponents; k++) {
         const comp = this.components[k];
@@ -2048,19 +2075,22 @@ var ColbertAdapter = class {
     const { memory, colbertMaxSimWasm } = this.wasm.exports;
     const queryBytes = queryTokens.byteLength;
     const docBytes = documentTokens.byteLength;
-    const totalBytes = queryBytes + docBytes;
-    ensureWasmMemory(totalBytes);
-    const queryPtr = 0;
-    const docPtr = queryBytes;
-    writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
-    writeFloat32ArrayToWasm(memory, documentTokens, docPtr);
-    return colbertMaxSimWasm(
-      queryPtr,
-      docPtr,
-      numQueryTokens,
-      numDocTokens,
-      dim
-    );
+    const initialOffset = getWasmAllocatorOffset();
+    try {
+      const queryPtr = allocateWasmMemory(queryBytes);
+      const docPtr = allocateWasmMemory(docBytes);
+      writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
+      writeFloat32ArrayToWasm(memory, documentTokens, docPtr);
+      return colbertMaxSimWasm(
+        queryPtr,
+        docPtr,
+        numQueryTokens,
+        numDocTokens,
+        dim
+      );
+    } finally {
+      setWasmAllocatorOffset(initialOffset);
+    }
   }
   /**
    * クエリと複数のドキュメント間の MaxSim スコアを計算し、スコアの降順にソートして返します。
@@ -2084,27 +2114,30 @@ var ColbertAdapter = class {
     }
     const queryBytes = queryTokens.byteLength;
     const maxDocBytes = maxDocLen * Float32Array.BYTES_PER_ELEMENT;
-    const totalBytes = queryBytes + maxDocBytes;
-    ensureWasmMemory(totalBytes);
-    const queryPtr = 0;
-    const docPtr = queryBytes;
-    writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
-    const results = documentTokensArray.map((doc, index) => {
-      const numDocTokens = doc.length / dim;
-      if (numDocTokens % 1 !== 0) {
-        throw new Error(`Invalid documentTokens length at index ${index}`);
-      }
-      writeFloat32ArrayToWasm(memory, doc, docPtr);
-      const score = colbertMaxSimWasm(
-        queryPtr,
-        docPtr,
-        numQueryTokens,
-        numDocTokens,
-        dim
-      );
-      return { index, score };
-    });
-    return results.sort((a, b) => b.score - a.score);
+    const initialOffset = getWasmAllocatorOffset();
+    try {
+      const queryPtr = allocateWasmMemory(queryBytes);
+      const docPtr = allocateWasmMemory(maxDocBytes);
+      writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
+      const results = documentTokensArray.map((doc, index) => {
+        const numDocTokens = doc.length / dim;
+        if (numDocTokens % 1 !== 0) {
+          throw new Error(`Invalid documentTokens length at index ${index}`);
+        }
+        writeFloat32ArrayToWasm(memory, doc, docPtr);
+        const score = colbertMaxSimWasm(
+          queryPtr,
+          docPtr,
+          numQueryTokens,
+          numDocTokens,
+          dim
+        );
+        return { index, score };
+      });
+      return results.sort((a, b) => b.score - a.score);
+    } finally {
+      setWasmAllocatorOffset(initialOffset);
+    }
   }
 };
 
@@ -2124,6 +2157,13 @@ var VectorDBAdapter = class {
    * @returns {string} `'[0.1, 0.2, 0.3]'` のような文字列表現
    */
   static toPgvector(vector) {
+    if (vector instanceof Uint8Array) {
+      let bitString = "";
+      for (let i = 0; i < vector.length; i++) {
+        bitString += vector[i].toString(2).padStart(8, "0");
+      }
+      return bitString;
+    }
     return `[${Array.from(vector).join(", ")}]`;
   }
   /**
@@ -2187,18 +2227,37 @@ var withWarpVector = (config) => {
            */
           async searchByVector(args) {
             const context = import_extension.Prisma.getExtensionContext(this);
-            const tableName = context.$name;
+            const tableName = context.$name || "document";
+            if (!/^[a-zA-Z0-9_]+$/.test(tableName)) {
+              throw new Error(`Invalid table name: ${tableName}`);
+            }
+            if (!/^[a-zA-Z0-9_]+$/.test(vectorField)) {
+              throw new Error(`Invalid vector field name: ${vectorField}`);
+            }
+            if (!["<->", "<#>", "<=>"].includes(distanceOp)) {
+              throw new Error(`Invalid distance operator: ${distanceOp}`);
+            }
+            let whereClause = "";
+            if (args.where) {
+              const trimmedWhere = args.where.trim();
+              if (/;|--|\/\*/.test(trimmedWhere)) {
+                throw new Error("Potential SQL injection detected in where clause.");
+              }
+              whereClause = `WHERE ${trimmedWhere}`;
+            }
+            const limit = args.topK ?? 10;
+            if (typeof limit !== "number" || isNaN(limit) || limit < 0) {
+              throw new Error("Invalid topK value.");
+            }
             const tunedVector = config.adapter.tune(args.vector);
             const pgVectorStr = VectorDBAdapter.toPgvector(tunedVector);
-            const limit = args.topK ?? 10;
-            const whereClause = args.where ? `WHERE ${args.where}` : "";
             const sql = `
               SELECT *
               FROM "${tableName}"
               ${whereClause}
               ORDER BY "${vectorField}" ${distanceOp} '${pgVectorStr}'::vector
               LIMIT ${limit};
-            `;
+            `.trim();
             return client.$queryRawUnsafe(sql);
           }
         }
@@ -2570,9 +2629,11 @@ var WarpLlamaIndexEmbeddings = class {
 var QuantizationAdapter = class _QuantizationAdapter {
   type;
   dim;
+  dynamic;
   constructor(config) {
     this.type = config.type;
     this.dim = config.dim;
+    this.dynamic = config.dynamic ?? false;
     if (this.type === "binary" && this.dim % 8 !== 0) {
       throw new Error(
         `Binary quantization requires dimension to be a multiple of 8. Got ${this.dim}`
@@ -2582,14 +2643,33 @@ var QuantizationAdapter = class _QuantizationAdapter {
   tune(vector) {
     assertDimension(vector, this.dim, "QuantizationAdapter.tune");
     if (this.type === "int8") {
-      const result = new Int8Array(this.dim);
-      for (let i = 0; i < this.dim; i++) {
-        let val = Math.round(vector[i] * 127);
-        if (val > 127) val = 127;
-        if (val < -128) val = -128;
-        result[i] = val;
+      if (this.dynamic) {
+        let maxVal = 1e-8;
+        for (let i = 0; i < this.dim; i++) {
+          const abs = Math.abs(vector[i]);
+          if (abs > maxVal) maxVal = abs;
+        }
+        const scale = 127 / maxVal;
+        const result = new Int8Array(this.dim + 4);
+        for (let i = 0; i < this.dim; i++) {
+          let val = Math.round(vector[i] * scale);
+          if (val > 127) val = 127;
+          if (val < -128) val = -128;
+          result[i] = val;
+        }
+        const view = new DataView(result.buffer, result.byteOffset, result.byteLength);
+        view.setFloat32(this.dim, maxVal, true);
+        return result;
+      } else {
+        const result = new Int8Array(this.dim);
+        for (let i = 0; i < this.dim; i++) {
+          let val = Math.round(vector[i] * 127);
+          if (val > 127) val = 127;
+          if (val < -128) val = -128;
+          result[i] = val;
+        }
+        return result;
       }
-      return result;
     } else if (this.type === "binary") {
       const bytesLength = this.dim / 8;
       const result = new Uint8Array(bytesLength);
@@ -2622,17 +2702,45 @@ var QuantizationAdapter = class _QuantizationAdapter {
   }
   /**
    * Int8量子化された2つのベクトル間のドット積（内積）を計算します。
+   * 動的スケーリングが埋め込まれている場合はスケールを戻して計算します。
    */
   static int8DotProduct(a, b) {
     if (a.length !== b.length) throw new Error("Length mismatch");
-    let dot = 0;
-    for (let i = 0; i < a.length; i++) {
-      dot += a[i] * b[i];
+    const hasScaleBytes = a.length > 4;
+    let isDynamic = false;
+    let maxA = 1;
+    let maxB = 1;
+    if (hasScaleBytes) {
+      const dim = a.length - 4;
+      try {
+        const viewA = new DataView(a.buffer, a.byteOffset, a.byteLength);
+        const viewB = new DataView(b.buffer, b.byteOffset, b.byteLength);
+        maxA = viewA.getFloat32(dim, true);
+        maxB = viewB.getFloat32(dim, true);
+        if (!isNaN(maxA) && !isNaN(maxB) && maxA > 0 && maxA < 1e3 && maxB > 0 && maxB < 1e3) {
+          isDynamic = true;
+        }
+      } catch (e) {
+        isDynamic = false;
+      }
+    }
+    if (isDynamic) {
+      const dim = a.length - 4;
+      let dot = 0;
+      for (let i = 0; i < dim; i++) {
+        dot += a[i] * b[i];
+      }
+      return dot * (maxA / 127) * (maxB / 127);
+    } else {
+      let dot = 0;
+      for (let i = 0; i < a.length; i++) {
+        dot += a[i] * b[i];
+      }
+      return dot;
     }
-    return dot;
   }
   exportState() {
-    return JSON.stringify({ type: this.type, dim: this.dim });
+    return JSON.stringify({ type: this.type, dim: this.dim, dynamic: this.dynamic });
   }
   static importState(stateJson) {
     const config = JSON.parse(stateJson);
@@ -2667,22 +2775,78 @@ var AbstractAdamTrainer = class {
     const beta1 = 0.9;
     const beta2 = 0.999;
     const epsilon = 1e-8;
-    for (let i = 0; i < tDim; i++) {
-      const bGrad = outputGradients[i];
-      mBias[i] = beta1 * mBias[i] + (1 - beta1) * bGrad;
-      vBias[i] = beta2 * vBias[i] + (1 - beta2) * (bGrad * bGrad);
-      const mHatB = mBias[i] / (1 - Math.pow(beta1, t));
-      const vHatB = vBias[i] / (1 - Math.pow(beta2, t));
-      bias[i] -= lr * mHatB / (Math.sqrt(vHatB) + epsilon);
-      const rowOffset = i * sDim;
-      for (let j = 0; j < sDim; j++) {
-        const wIdx = rowOffset + j;
-        const wGrad = bGrad * input[j] + reg * matrix[wIdx];
-        mMatrix[wIdx] = beta1 * mMatrix[wIdx] + (1 - beta1) * wGrad;
-        vMatrix[wIdx] = beta2 * vMatrix[wIdx] + (1 - beta2) * (wGrad * wGrad);
-        const mHatW = mMatrix[wIdx] / (1 - Math.pow(beta1, t));
-        const vHatW = vMatrix[wIdx] / (1 - Math.pow(beta2, t));
-        matrix[wIdx] -= lr * mHatW / (Math.sqrt(vHatW) + epsilon);
+    const instance = getWasmInstance();
+    if (instance && instance.exports.adamUpdateWasm) {
+      const memory = instance.exports.memory;
+      const matrixBytes = matrix.byteLength;
+      const biasBytes = bias.byteLength;
+      const inputBytes = sDim * 4;
+      const gradBytes = tDim * 4;
+      const initialOffset = getWasmAllocatorOffset();
+      try {
+        const matrixPtr = allocateWasmMemory(matrixBytes);
+        const biasPtr = allocateWasmMemory(biasBytes);
+        const mMatrixPtr = allocateWasmMemory(matrixBytes);
+        const vMatrixPtr = allocateWasmMemory(matrixBytes);
+        const mBiasPtr = allocateWasmMemory(biasBytes);
+        const vBiasPtr = allocateWasmMemory(biasBytes);
+        const inputPtr = allocateWasmMemory(inputBytes);
+        const gradPtr = allocateWasmMemory(gradBytes);
+        writeFloat32ArrayToWasm(memory, matrix, matrixPtr);
+        writeFloat32ArrayToWasm(memory, bias, biasPtr);
+        writeFloat32ArrayToWasm(memory, mMatrix, mMatrixPtr);
+        writeFloat32ArrayToWasm(memory, vMatrix, vMatrixPtr);
+        writeFloat32ArrayToWasm(memory, mBias, mBiasPtr);
+        writeFloat32ArrayToWasm(memory, vBias, vBiasPtr);
+        writeFloat32ArrayToWasm(memory, input, inputPtr);
+        writeFloat32ArrayToWasm(memory, outputGradients, gradPtr);
+        const adamUpdateWasm = instance.exports.adamUpdateWasm;
+        adamUpdateWasm(
+          matrixPtr,
+          biasPtr,
+          mMatrixPtr,
+          vMatrixPtr,
+          mBiasPtr,
+          vBiasPtr,
+          inputPtr,
+          gradPtr,
+          lr,
+          reg,
+          beta1,
+          beta2,
+          epsilon,
+          t,
+          sDim,
+          tDim
+        );
+        const f32 = new Float32Array(memory.buffer);
+        matrix.set(f32.subarray(matrixPtr / 4, matrixPtr / 4 + matrix.length));
+        bias.set(f32.subarray(biasPtr / 4, biasPtr / 4 + bias.length));
+        mMatrix.set(f32.subarray(mMatrixPtr / 4, mMatrixPtr / 4 + mMatrix.length));
+        vMatrix.set(f32.subarray(vMatrixPtr / 4, vMatrixPtr / 4 + vMatrix.length));
+        mBias.set(f32.subarray(mBiasPtr / 4, mBiasPtr / 4 + mBias.length));
+        vBias.set(f32.subarray(vBiasPtr / 4, vBiasPtr / 4 + vBias.length));
+      } finally {
+        setWasmAllocatorOffset(initialOffset);
+      }
+    } else {
+      for (let i = 0; i < tDim; i++) {
+        const bGrad = outputGradients[i];
+        mBias[i] = beta1 * mBias[i] + (1 - beta1) * bGrad;
+        vBias[i] = beta2 * vBias[i] + (1 - beta2) * (bGrad * bGrad);
+        const mHatB = mBias[i] / (1 - Math.pow(beta1, t));
+        const vHatB = vBias[i] / (1 - Math.pow(beta2, t));
+        bias[i] -= lr * mHatB / (Math.sqrt(vHatB) + epsilon);
+        const rowOffset = i * sDim;
+        for (let j = 0; j < sDim; j++) {
+          const wIdx = rowOffset + j;
+          const wGrad = bGrad * input[j] + reg * matrix[wIdx];
+          mMatrix[wIdx] = beta1 * mMatrix[wIdx] + (1 - beta1) * wGrad;
+          vMatrix[wIdx] = beta2 * vMatrix[wIdx] + (1 - beta2) * (wGrad * wGrad);
+          const mHatW = mMatrix[wIdx] / (1 - Math.pow(beta1, t));
+          const vHatW = vMatrix[wIdx] / (1 - Math.pow(beta2, t));
+          matrix[wIdx] -= lr * mHatW / (Math.sqrt(vHatW) + epsilon);
+        }
       }
     }
   }

package/dist/index.mjs

@@ -788,7 +788,7 @@ function applyAffine(matrix, bias, vector, result, inDim, outDim = inDim) {
 }
 
 // src/wasm/wasm-binary.ts
-var wasmBase64 = "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";
+var wasmBase64 = "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";
 
 // src/wasm/wasm-loader.ts
 var wasmInstance = null;
@@ -826,6 +826,22 @@ function ensureWasmMemory(requiredBytes) {
     return false;
   }
 }
+var globalOffset = 65536;
+function allocateWasmMemory(bytes) {
+  const ptr = globalOffset;
+  globalOffset += bytes;
+  if (globalOffset % 4 !== 0) {
+    globalOffset += 4 - globalOffset % 4;
+  }
+  ensureWasmMemory(globalOffset);
+  return ptr;
+}
+function getWasmAllocatorOffset() {
+  return globalOffset;
+}
+function setWasmAllocatorOffset(offset) {
+  globalOffset = offset;
+}
 function writeFloat32ArrayToWasm(memory, data, byteOffset) {
   const f32 = new Float32Array(memory.buffer);
   const floatOffset = byteOffset / 4;
@@ -1665,6 +1681,7 @@ var MlpAdapter = class _MlpAdapter {
   layerDimsPtr = 0;
   activationsPtr = 0;
   bufferPtr = 0;
+  bufBPtr = 0;
   constructor(layers) {
     if (layers.length === 0) {
       throw new Error("MlpAdapter requires at least one layer.");
@@ -1715,21 +1732,13 @@ var MlpAdapter = class _MlpAdapter {
     }
     this.inputDim = layerDims[0];
     this.outputDim = layerDims[layerDims.length - 1];
-    let offset = 65536;
-    this.inputPtr = offset;
-    offset += this.inputDim * 4;
-    this.outputPtr = offset;
-    offset += this.outputDim * 4;
-    this.layerDimsPtr = offset;
-    offset += (this.numLayers + 1) * 4;
-    this.activationsPtr = offset;
-    offset += this.numLayers * 4;
-    if (offset % 4 !== 0) offset += 4 - offset % 4;
-    this.weightsPtr = offset;
-    offset += totalWeights * 4;
-    this.bufferPtr = offset;
-    offset += maxDim * 8 + 8192;
-    ensureWasmMemory(offset);
+    this.inputPtr = allocateWasmMemory(this.inputDim * 4);
+    this.outputPtr = allocateWasmMemory(this.outputDim * 4);
+    this.layerDimsPtr = allocateWasmMemory((this.numLayers + 1) * 4);
+    this.activationsPtr = allocateWasmMemory(this.numLayers * 4);
+    this.weightsPtr = allocateWasmMemory(totalWeights * 4);
+    this.bufferPtr = allocateWasmMemory(maxDim * 4);
+    this.bufBPtr = allocateWasmMemory(maxDim * 4);
     const memoryBuffer = memory.buffer;
     const f32 = new Float32Array(memoryBuffer);
     const i32 = new Int32Array(memoryBuffer);
@@ -1789,7 +1798,8 @@ var MlpAdapter = class _MlpAdapter {
       this.layerDimsPtr,
       this.activationsPtr,
       this.numLayers,
-      this.bufferPtr
+      this.bufferPtr,
+      this.bufBPtr
     );
     const result = new Float32Array(this.outputDim);
     const outF32 = new Float32Array(memory.buffer);
@@ -1843,6 +1853,19 @@ var WhiteningAdapter = class _WhiteningAdapter {
   count = 0;
   learningRate;
   numComponents;
+  componentsPtr = 0;
+  xResidualPtr = 0;
+  isWasmReady = false;
+  async init() {
+    const instance = await initWasm();
+    if (instance) {
+      const componentsSize = this.numComponents * this.dim * 4;
+      const xResidualSize = this.dim * 4;
+      this.componentsPtr = allocateWasmMemory(componentsSize);
+      this.xResidualPtr = allocateWasmMemory(xResidualSize);
+      this.isWasmReady = true;
+    }
+  }
   /**
    * 新しい WhiteningAdapter を作成します。
    * @param dim ベクトルの次元数
@@ -1886,13 +1909,17 @@ var WhiteningAdapter = class _WhiteningAdapter {
     }
     const x_residual = new Float32Array(x);
     const instance = getWasmInstance();
-    const componentsSize = this.numComponents * this.dim * 4;
-    const xResidualSize = this.dim * 4;
-    const requiredBytes = componentsSize + xResidualSize;
-    if (instance && instance.exports.sangerUpdateWasm && ensureWasmMemory(requiredBytes)) {
+    if (instance && instance.exports.sangerUpdateWasm) {
+      if (!this.isWasmReady) {
+        const componentsSize = this.numComponents * this.dim * 4;
+        const xResidualSize = this.dim * 4;
+        this.componentsPtr = allocateWasmMemory(componentsSize);
+        this.xResidualPtr = allocateWasmMemory(xResidualSize);
+        this.isWasmReady = true;
+      }
       const memory = instance.exports.memory;
-      const componentsPtr = 0;
-      const xResidualPtr = componentsSize;
+      const componentsPtr = this.componentsPtr;
+      const xResidualPtr = this.xResidualPtr;
       const f32 = new Float32Array(memory.buffer);
       for (let k = 0; k < this.numComponents; k++) {
         const comp = this.components[k];
@@ -2007,19 +2034,22 @@ var ColbertAdapter = class {
     const { memory, colbertMaxSimWasm } = this.wasm.exports;
     const queryBytes = queryTokens.byteLength;
     const docBytes = documentTokens.byteLength;
-    const totalBytes = queryBytes + docBytes;
-    ensureWasmMemory(totalBytes);
-    const queryPtr = 0;
-    const docPtr = queryBytes;
-    writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
-    writeFloat32ArrayToWasm(memory, documentTokens, docPtr);
-    return colbertMaxSimWasm(
-      queryPtr,
-      docPtr,
-      numQueryTokens,
-      numDocTokens,
-      dim
-    );
+    const initialOffset = getWasmAllocatorOffset();
+    try {
+      const queryPtr = allocateWasmMemory(queryBytes);
+      const docPtr = allocateWasmMemory(docBytes);
+      writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
+      writeFloat32ArrayToWasm(memory, documentTokens, docPtr);
+      return colbertMaxSimWasm(
+        queryPtr,
+        docPtr,
+        numQueryTokens,
+        numDocTokens,
+        dim
+      );
+    } finally {
+      setWasmAllocatorOffset(initialOffset);
+    }
   }
   /**
    * クエリと複数のドキュメント間の MaxSim スコアを計算し、スコアの降順にソートして返します。
@@ -2043,27 +2073,30 @@ var ColbertAdapter = class {
     }
     const queryBytes = queryTokens.byteLength;
     const maxDocBytes = maxDocLen * Float32Array.BYTES_PER_ELEMENT;
-    const totalBytes = queryBytes + maxDocBytes;
-    ensureWasmMemory(totalBytes);
-    const queryPtr = 0;
-    const docPtr = queryBytes;
-    writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
-    const results = documentTokensArray.map((doc, index) => {
-      const numDocTokens = doc.length / dim;
-      if (numDocTokens % 1 !== 0) {
-        throw new Error(`Invalid documentTokens length at index ${index}`);
-      }
-      writeFloat32ArrayToWasm(memory, doc, docPtr);
-      const score = colbertMaxSimWasm(
-        queryPtr,
-        docPtr,
-        numQueryTokens,
-        numDocTokens,
-        dim
-      );
-      return { index, score };
-    });
-    return results.sort((a, b) => b.score - a.score);
+    const initialOffset = getWasmAllocatorOffset();
+    try {
+      const queryPtr = allocateWasmMemory(queryBytes);
+      const docPtr = allocateWasmMemory(maxDocBytes);
+      writeFloat32ArrayToWasm(memory, queryTokens, queryPtr);
+      const results = documentTokensArray.map((doc, index) => {
+        const numDocTokens = doc.length / dim;
+        if (numDocTokens % 1 !== 0) {
+          throw new Error(`Invalid documentTokens length at index ${index}`);
+        }
+        writeFloat32ArrayToWasm(memory, doc, docPtr);
+        const score = colbertMaxSimWasm(
+          queryPtr,
+          docPtr,
+          numQueryTokens,
+          numDocTokens,
+          dim
+        );
+        return { index, score };
+      });
+      return results.sort((a, b) => b.score - a.score);
+    } finally {
+      setWasmAllocatorOffset(initialOffset);
+    }
   }
 };
 
@@ -2083,6 +2116,13 @@ var VectorDBAdapter = class {
    * @returns {string} `'[0.1, 0.2, 0.3]'` のような文字列表現
    */
   static toPgvector(vector) {
+    if (vector instanceof Uint8Array) {
+      let bitString = "";
+      for (let i = 0; i < vector.length; i++) {
+        bitString += vector[i].toString(2).padStart(8, "0");
+      }
+      return bitString;
+    }
     return `[${Array.from(vector).join(", ")}]`;
   }
   /**
@@ -2146,18 +2186,37 @@ var withWarpVector = (config) => {
            */
           async searchByVector(args) {
             const context = Prisma.getExtensionContext(this);
-            const tableName = context.$name;
+            const tableName = context.$name || "document";
+            if (!/^[a-zA-Z0-9_]+$/.test(tableName)) {
+              throw new Error(`Invalid table name: ${tableName}`);
+            }
+            if (!/^[a-zA-Z0-9_]+$/.test(vectorField)) {
+              throw new Error(`Invalid vector field name: ${vectorField}`);
+            }
+            if (!["<->", "<#>", "<=>"].includes(distanceOp)) {
+              throw new Error(`Invalid distance operator: ${distanceOp}`);
+            }
+            let whereClause = "";
+            if (args.where) {
+              const trimmedWhere = args.where.trim();
+              if (/;|--|\/\*/.test(trimmedWhere)) {
+                throw new Error("Potential SQL injection detected in where clause.");
+              }
+              whereClause = `WHERE ${trimmedWhere}`;
+            }
+            const limit = args.topK ?? 10;
+            if (typeof limit !== "number" || isNaN(limit) || limit < 0) {
+              throw new Error("Invalid topK value.");
+            }
             const tunedVector = config.adapter.tune(args.vector);
             const pgVectorStr = VectorDBAdapter.toPgvector(tunedVector);
-            const limit = args.topK ?? 10;
-            const whereClause = args.where ? `WHERE ${args.where}` : "";
             const sql = `
               SELECT *
               FROM "${tableName}"
               ${whereClause}
               ORDER BY "${vectorField}" ${distanceOp} '${pgVectorStr}'::vector
               LIMIT ${limit};
-            `;
+            `.trim();
             return client.$queryRawUnsafe(sql);
           }
         }
@@ -2529,9 +2588,11 @@ var WarpLlamaIndexEmbeddings = class {
 var QuantizationAdapter = class _QuantizationAdapter {
   type;
   dim;
+  dynamic;
   constructor(config) {
     this.type = config.type;
     this.dim = config.dim;
+    this.dynamic = config.dynamic ?? false;
     if (this.type === "binary" && this.dim % 8 !== 0) {
       throw new Error(
         `Binary quantization requires dimension to be a multiple of 8. Got ${this.dim}`
@@ -2541,14 +2602,33 @@ var QuantizationAdapter = class _QuantizationAdapter {
   tune(vector) {
     assertDimension(vector, this.dim, "QuantizationAdapter.tune");
     if (this.type === "int8") {
-      const result = new Int8Array(this.dim);
-      for (let i = 0; i < this.dim; i++) {
-        let val = Math.round(vector[i] * 127);
-        if (val > 127) val = 127;
-        if (val < -128) val = -128;
-        result[i] = val;
+      if (this.dynamic) {
+        let maxVal = 1e-8;
+        for (let i = 0; i < this.dim; i++) {
+          const abs = Math.abs(vector[i]);
+          if (abs > maxVal) maxVal = abs;
+        }
+        const scale = 127 / maxVal;
+        const result = new Int8Array(this.dim + 4);
+        for (let i = 0; i < this.dim; i++) {
+          let val = Math.round(vector[i] * scale);
+          if (val > 127) val = 127;
+          if (val < -128) val = -128;
+          result[i] = val;
+        }
+        const view = new DataView(result.buffer, result.byteOffset, result.byteLength);
+        view.setFloat32(this.dim, maxVal, true);
+        return result;
+      } else {
+        const result = new Int8Array(this.dim);
+        for (let i = 0; i < this.dim; i++) {
+          let val = Math.round(vector[i] * 127);
+          if (val > 127) val = 127;
+          if (val < -128) val = -128;
+          result[i] = val;
+        }
+        return result;
       }
-      return result;
     } else if (this.type === "binary") {
       const bytesLength = this.dim / 8;
       const result = new Uint8Array(bytesLength);
@@ -2581,17 +2661,45 @@ var QuantizationAdapter = class _QuantizationAdapter {
   }
   /**
    * Int8量子化された2つのベクトル間のドット積（内積）を計算します。
+   * 動的スケーリングが埋め込まれている場合はスケールを戻して計算します。
    */
   static int8DotProduct(a, b) {
     if (a.length !== b.length) throw new Error("Length mismatch");
-    let dot = 0;
-    for (let i = 0; i < a.length; i++) {
-      dot += a[i] * b[i];
+    const hasScaleBytes = a.length > 4;
+    let isDynamic = false;
+    let maxA = 1;
+    let maxB = 1;
+    if (hasScaleBytes) {
+      const dim = a.length - 4;
+      try {
+        const viewA = new DataView(a.buffer, a.byteOffset, a.byteLength);
+        const viewB = new DataView(b.buffer, b.byteOffset, b.byteLength);
+        maxA = viewA.getFloat32(dim, true);
+        maxB = viewB.getFloat32(dim, true);
+        if (!isNaN(maxA) && !isNaN(maxB) && maxA > 0 && maxA < 1e3 && maxB > 0 && maxB < 1e3) {
+          isDynamic = true;
+        }
+      } catch (e) {
+        isDynamic = false;
+      }
+    }
+    if (isDynamic) {
+      const dim = a.length - 4;
+      let dot = 0;
+      for (let i = 0; i < dim; i++) {
+        dot += a[i] * b[i];
+      }
+      return dot * (maxA / 127) * (maxB / 127);
+    } else {
+      let dot = 0;
+      for (let i = 0; i < a.length; i++) {
+        dot += a[i] * b[i];
+      }
+      return dot;
     }
-    return dot;
   }
   exportState() {
-    return JSON.stringify({ type: this.type, dim: this.dim });
+    return JSON.stringify({ type: this.type, dim: this.dim, dynamic: this.dynamic });
   }
   static importState(stateJson) {
     const config = JSON.parse(stateJson);
@@ -2626,22 +2734,78 @@ var AbstractAdamTrainer = class {
     const beta1 = 0.9;
     const beta2 = 0.999;
     const epsilon = 1e-8;
-    for (let i = 0; i < tDim; i++) {
-      const bGrad = outputGradients[i];
-      mBias[i] = beta1 * mBias[i] + (1 - beta1) * bGrad;
-      vBias[i] = beta2 * vBias[i] + (1 - beta2) * (bGrad * bGrad);
-      const mHatB = mBias[i] / (1 - Math.pow(beta1, t));
-      const vHatB = vBias[i] / (1 - Math.pow(beta2, t));
-      bias[i] -= lr * mHatB / (Math.sqrt(vHatB) + epsilon);
-      const rowOffset = i * sDim;
-      for (let j = 0; j < sDim; j++) {
-        const wIdx = rowOffset + j;
-        const wGrad = bGrad * input[j] + reg * matrix[wIdx];
-        mMatrix[wIdx] = beta1 * mMatrix[wIdx] + (1 - beta1) * wGrad;
-        vMatrix[wIdx] = beta2 * vMatrix[wIdx] + (1 - beta2) * (wGrad * wGrad);
-        const mHatW = mMatrix[wIdx] / (1 - Math.pow(beta1, t));
-        const vHatW = vMatrix[wIdx] / (1 - Math.pow(beta2, t));
-        matrix[wIdx] -= lr * mHatW / (Math.sqrt(vHatW) + epsilon);
+    const instance = getWasmInstance();
+    if (instance && instance.exports.adamUpdateWasm) {
+      const memory = instance.exports.memory;
+      const matrixBytes = matrix.byteLength;
+      const biasBytes = bias.byteLength;
+      const inputBytes = sDim * 4;
+      const gradBytes = tDim * 4;
+      const initialOffset = getWasmAllocatorOffset();
+      try {
+        const matrixPtr = allocateWasmMemory(matrixBytes);
+        const biasPtr = allocateWasmMemory(biasBytes);
+        const mMatrixPtr = allocateWasmMemory(matrixBytes);
+        const vMatrixPtr = allocateWasmMemory(matrixBytes);
+        const mBiasPtr = allocateWasmMemory(biasBytes);
+        const vBiasPtr = allocateWasmMemory(biasBytes);
+        const inputPtr = allocateWasmMemory(inputBytes);
+        const gradPtr = allocateWasmMemory(gradBytes);
+        writeFloat32ArrayToWasm(memory, matrix, matrixPtr);
+        writeFloat32ArrayToWasm(memory, bias, biasPtr);
+        writeFloat32ArrayToWasm(memory, mMatrix, mMatrixPtr);
+        writeFloat32ArrayToWasm(memory, vMatrix, vMatrixPtr);
+        writeFloat32ArrayToWasm(memory, mBias, mBiasPtr);
+        writeFloat32ArrayToWasm(memory, vBias, vBiasPtr);
+        writeFloat32ArrayToWasm(memory, input, inputPtr);
+        writeFloat32ArrayToWasm(memory, outputGradients, gradPtr);
+        const adamUpdateWasm = instance.exports.adamUpdateWasm;
+        adamUpdateWasm(
+          matrixPtr,
+          biasPtr,
+          mMatrixPtr,
+          vMatrixPtr,
+          mBiasPtr,
+          vBiasPtr,
+          inputPtr,
+          gradPtr,
+          lr,
+          reg,
+          beta1,
+          beta2,
+          epsilon,
+          t,
+          sDim,
+          tDim
+        );
+        const f32 = new Float32Array(memory.buffer);
+        matrix.set(f32.subarray(matrixPtr / 4, matrixPtr / 4 + matrix.length));
+        bias.set(f32.subarray(biasPtr / 4, biasPtr / 4 + bias.length));
+        mMatrix.set(f32.subarray(mMatrixPtr / 4, mMatrixPtr / 4 + mMatrix.length));
+        vMatrix.set(f32.subarray(vMatrixPtr / 4, vMatrixPtr / 4 + vMatrix.length));
+        mBias.set(f32.subarray(mBiasPtr / 4, mBiasPtr / 4 + mBias.length));
+        vBias.set(f32.subarray(vBiasPtr / 4, vBiasPtr / 4 + vBias.length));
+      } finally {
+        setWasmAllocatorOffset(initialOffset);
+      }
+    } else {
+      for (let i = 0; i < tDim; i++) {
+        const bGrad = outputGradients[i];
+        mBias[i] = beta1 * mBias[i] + (1 - beta1) * bGrad;
+        vBias[i] = beta2 * vBias[i] + (1 - beta2) * (bGrad * bGrad);
+        const mHatB = mBias[i] / (1 - Math.pow(beta1, t));
+        const vHatB = vBias[i] / (1 - Math.pow(beta2, t));
+        bias[i] -= lr * mHatB / (Math.sqrt(vHatB) + epsilon);
+        const rowOffset = i * sDim;
+        for (let j = 0; j < sDim; j++) {
+          const wIdx = rowOffset + j;
+          const wGrad = bGrad * input[j] + reg * matrix[wIdx];
+          mMatrix[wIdx] = beta1 * mMatrix[wIdx] + (1 - beta1) * wGrad;
+          vMatrix[wIdx] = beta2 * vMatrix[wIdx] + (1 - beta2) * (wGrad * wGrad);
+          const mHatW = mMatrix[wIdx] / (1 - Math.pow(beta1, t));
+          const vHatW = vMatrix[wIdx] / (1 - Math.pow(beta2, t));
+          matrix[wIdx] -= lr * mHatW / (Math.sqrt(vHatW) + epsilon);
+        }
       }
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "warpvector",
-  "version": "0.1.1",
+  "version": "0.1.2",
   "description": "A lightweight, zero-dependency TypeScript middleware for dynamic vector space transformation (affine transformation).",
   "main": "dist/index.js",
   "module": "dist/index.mjs",