expo-vector-search 0.1.0 → 0.3.0

package/README.md

@@ -3,17 +3,29 @@
 `expo-vector-search` is a high-performance, on-device **vector search engine** module for Expo and React Native.
 
 > [!NOTE]
-> This module is currently **optimized and validated for Android**. The C++ JSI layer is fully functional, but the build configuration and native runtime have been primarily verified on Android devices. iOS support is on the roadmap.
+> This module is **cross-platform** (Android & iOS). The C++ JSI core and build configurations have been validated on production devices (Galaxy S23 FE and iPhone 12).
 
-## Performance (Real-world Benchmarks)
-Results obtained on a **Samsung Galaxy S23 FE**:
-- **Search (1000 items, 128 dims)**: **0.08ms** per query.
-- **Batch Ingest (1000 items)**: **74.44ms** total time.
-- **Memory Optimization (10k items, 384 dims)**:
-    - F32 Footprint: 36,964.94 KB.
-    - Int8 Footprint: 20,580.94 KB.
+## Performance (Release Benchmarks)
 
-![Performance Lab Benchmarks on S23 FE](../../assets/images/perf_lab.jpg)
+Benchmark results obtained using **Release builds** on physical devices (1,000 vectors, 128 dimensions for search/ingestion; 10,000 vectors, 384 dimensions for memory optimization).
+
+### JS vs. Native Engine Race
+| Platform | JavaScript (Runtime Loop) | Expo Vector Search (Native) | Speedup |
+| :--- | :--- | :--- | :--- |
+| **Android** (S23 FE) | 6.20 ms | 0.15 ms | **~41x** |
+| **iOS** (iPhone 12) | 12.06 ms | 0.10 ms | **~120x** |
+
+### Bulk Ingestion (1,000 items)
+| Platform | Individual `.add` | Batch `.addBatch` |
+| :--- | :--- | :--- |
+| **Android** (S23 FE) | 79.87 ms | 76.70 ms |
+| **iOS** (iPhone 12) | 107.94 ms | 102.59 ms |
+
+### Memory Optimization (10,000 items, 384 dims)
+| Platform | Full Precision (F32) | Quantized (Int8) | Savings |
+| :--- | :--- | :--- | :--- |
+| **Android** (S23 FE) | 36,943.84 KB | 20,559.84 KB | **~44%** |
+| **iOS** (iPhone 12) | 36,943.97 KB | 20,559.97 KB | **~44%** |
 
 ## Key Features
 
@@ -22,16 +34,43 @@ Results obtained on a **Samsung Galaxy S23 FE**:
 - **Production-Grade Quantization**: Support for Int8 quantization to reduce memory footprint by up to 4x with minimal impact on search accuracy.
 - **Disk Persistence**: Built-in methods to save and load vector indices to/from the local file system.
 - **Memory Safety**: Strict buffer alignment checks and type validation to ensure native stability.
+- **Extended Distance Metrics**: Support for Cosine, Euclidean (L2), Inner Product, Hamming (binary), and Jaccard distances.
 - **Explicit Memory Management**: Provision for deterministic resource release via the `delete()` method.
 
 ## Installation
 
-This module contains custom native code. You must use development builds to use this module.
+This module contains custom native code (C++/JSI). **It does not work in [Expo Go](https://expo.dev/go).** You must generate and build the native project.
 
+### 1. Install the package
 ```bash
 npx expo install expo-vector-search
 ```
 
+### 2. Generate Native Folders
+To access the `ios` and `android` directories and include the C++ engine, run:
+```bash
+npx expo prebuild
+```
+> [!NOTE]
+> This command generates the native projects based on your `app.json`. If you already have `ios` and `android` folders (Bare Workflow), you can skip this step.
+
+### 3. Build and Run
+You must compile the native code to use the module. Use the following commands to build and launch the app:
+```bash
+npx expo run:android
+# or
+npx expo run:ios
+```
+
+---
+
+### Compatibility
+
+| Environment | Supported | Requirement |
+| :--- | :--- | :--- |
+| **Expo Go** | ❌ No | Requires custom native engine |
+| **Bare Workflow** | ✅ Yes | Standard `ios`/`android` folders |
+
 ## Architecture
 
 The module is designed for performance-critical applications where latency and battery efficiency are paramount.
@@ -43,6 +82,17 @@ The module is designed for performance-critical applications where latency and b
 
 ## API Reference
 
+### useVectorSearch (React Hook)
+
+A wrapper hook that manages the lifecycle of a `VectorIndex`, handling creation and cleanup automatically.
+
+```typescript
+const { index, search, add } = useVectorSearch(384, {
+  quantization: 'i8',
+  metric: 'cos'
+});
+```
+
 ### VectorIndex
 
 The primary class for managing a vector collection.
@@ -51,6 +101,7 @@ The primary class for managing a vector collection.
 Initializes a new vector index.
 - `dimensions`: The dimensionality of the vectors (e.g., 128, 384, 768).
 - `options.quantization`: Scaling mode (`'f32'` or `'i8'`). Use `'i8'` for significant memory savings.
+- `options.metric`: Distance metric calculation (`'cos'`, `'l2sq'`, `'ip'`, `'hamming'`, `'jaccard'`). Default is `'cos'`.
 
 #### `add(key: number, vector: Float32Array): void`
 Inserts a vector into the index.
@@ -62,12 +113,22 @@ High-performance batch insertion. Significantly reduces JSI overhead by processi
 - `keys`: An `Int32Array` of unique identifiers.
 - `vectors`: A single `Float32Array` containing all vectors concatenated (must match `keys.length * dimensions`).
 
-#### `search(vector: Float32Array, count: number): SearchResult[]`
+#### `search(vector: Float32Array, count: number, options?: SearchOptions): SearchResult[]`
 Performs an ANN search.
 - `vector`: The query embedding.
 - `count`: Number of nearest neighbors to retrieve.
+- `options.allowedKeys`: Optional array of keys to restrict the search to (filtering).
 - **Returns**: An array of `SearchResult` objects `{ key: number, distance: number }`.
 
+#### `remove(key: number): void`
+Removes a vector from the index.
+- `key`: The unique numeric identifier of the vector to remove.
+
+#### `update(key: number, vector: Float32Array): void`
+Updates an existing vector in the index (upsert operation).
+- `key`: The unique numeric identifier.
+- `vector`: The new vector data.
+
 #### `save(path: string): void`
 Serializes the current state of the index to a specified file path.
 
@@ -77,6 +138,18 @@ Deserializes an index from a file path.
 #### `delete(): void`
 Manually releases native memory resources. The index instance becomes unusable after this call.
 
+#### `loadVectorsFromFile(path: string): number`
+Loads raw vectors directly from a binary file into the index.
+- `path`: Absolute path to the binary file containing packed floats.
+- **Returns**: The number of vectors successfully loaded.
+- **Note**: This is significantly faster than parsing JSON/Base64 in JavaScript and adding vectors loop by loop.
+
+#### `getItemVector(key: number): Float32Array | undefined`
+Retrieves the vector associated with a specific key.
+- `key`: The unique numeric identifier.
+- **Returns**: A `Float32Array` copy of the vector, or `undefined` if the key does not exist.
+- **Use Case**: Allows you to store vectors ONLY in native memory (saving JS RAM) and fetch them only when needed (e.g., for "Find Similar" queries).
+
 #### `dimensions: number` (readonly)
 Returns the dimensionality of the index.
 
@@ -129,5 +202,24 @@ This module performs strict validation on input buffers.
 - **Alignment**: `Float32Array` buffers must be 4-byte aligned for safe native access.
 - **Type Safety**: Input vectors are validated against the index's defined dimensions to prevent out-of-bounds memory operations.
 
+## Known Limitations & Roadmap
+
+### Performance Considerations (F32 vs Int8)
+While **Int8 Quantization** provides significant memory savings (~44% total index reduction and ~75% raw vector reduction), it currently involves a computational overhead during the ingestion process on Android:
+- **Full Precision (F32)**: ~9,284 ms per 10k vectors (Standard).
+- **Quantized (Int8)**: ~34,608 ms per 10k vectors (Slower due to real-time conversion).
+
+This performance gap is a known characteristic of the current version. The Int8 path requires a conversion from `float` to `int8` for every dimension, which is not yet fully vectorized in the Android build.
+
+### Future Roadmap
+- [x] **Dynamic CRUD Support**: Implemented `remove(key)` and `update(key, vector)`.
+- [x] **Metadata Filtering**: Support for `allowedKeys` filtering during search.
+- [ ] **Architecture-Specific SIMD**: Enable NEON/SVE optimizations for Android builds.
+- [ ] **Hybrid Search**: Integration with a keywords-based engine for hybrid results.
+- [ ] **Background Indexing**: True multithreaded ingestion to avoid JS bridge/thread locks.
+- [x] **Extended Distance Metrics**: Support for L2, IP, Hamming, and Jaccard.
+- [ ] **USearch Upgrade**: Migration to `v2.23.0+` for enhanced performance.
+- [ ] **Incremental Persistence**: Local storage optimizations for large datasets.
+
 ## License
 MIT

package/android/CMakeLists.txt

@@ -2,12 +2,12 @@ cmake_minimum_required(VERSION 3.10.2)
 project(ExpoVectorSearch)
 
 # 1. React Native / Expo Configuration (Default)
-set(CMAKE_VERBOSE_MAKEFILE ON)
+set(CMAKE_VERBOSE_MAKEFILE OFF)
 set(CMAKE_CXX_STANDARD 17)
 # Disable external FP16 library requirement (use native or fallback)
 add_compile_definitions(USEARCH_USE_FP16LIB=0)
 
-# 2. Baixar e Configurar o USearch
+# 2. Download and Configure USearch
 include(FetchContent)
 
 FetchContent_Declare(
@@ -16,10 +16,10 @@ FetchContent_Declare(
   GIT_TAG        v2.9.0 # Use a recent stable version
 )
 
-# Disponibiliza o usearch para o projeto
+# Make usearch available to the project
 FetchContent_MakeAvailable(usearch)
 
-# 3. Definir sua biblioteca compartilhada
+# 3. Define the shared library
 add_library(
   expo-vector-search
   SHARED
@@ -33,9 +33,10 @@ target_include_directories(
   expo-vector-search
   PUBLIC
   ${usearch_SOURCE_DIR}/include
+  ${CMAKE_CURRENT_SOURCE_DIR}/../cpp
 )
 
-# 5. Linkar bibliotecas (Log do Android e bibliotecas do JS Engine)
+# 5. Link libraries (Android Log and JS Engine libraries)
 # Important: React Native uses Prefab, so we need to find the ReactAndroid package
 find_package(ReactAndroid REQUIRED CONFIG)
 

package/android/build.gradle

@@ -1,7 +1,7 @@
 apply plugin: 'com.android.library'
 
 group = 'expo.modules.vectorsearch'
-version = '0.1.0'
+version = '0.2.0'
 
 def expoModulesCorePlugin = new File(project(":expo-modules-core").projectDir.absolutePath, "ExpoModulesCorePlugin.gradle")
 apply from: expoModulesCorePlugin
@@ -35,8 +35,8 @@ android {
   }
   
   defaultConfig {
-    versionCode 1
-    versionName "0.1.0"
+    versionCode 2
+    versionName "0.2.0"
 
 
     externalNativeBuild {

package/android/src/main/cpp/ExpoVectorSearchModule.cpp

@@ -1,306 +1,12 @@
+#include "ExpoVectorSearch.h"
 #include <jni.h>
-#include <string>
-#include <vector>
-#include <memory>
-#include <android/log.h>
-
-#define TAG "ExpoVectorSearch"
-#define LOGD(...) __android_log_print(ANDROID_LOG_DEBUG, TAG, __VA_ARGS__)
-#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, TAG, __VA_ARGS__)
-
-// Include the correct USearch header (Dense Index)
-// Polyfill for aligned_alloc on Android API < 28
-#if defined(__ANDROID__) && __ANDROID_API__ < 28
-#include <stdlib.h>
-extern "C" void* aligned_alloc(size_t alignment, size_t size) {
-    void* ptr = nullptr;
-    if (posix_memalign(&ptr, alignment, size) != 0) {
-        return nullptr;
-    }
-    return ptr;
-}
-#endif
-
-#include <usearch/index_dense.hpp>
-
-// JSI Headers
 #include <jsi/jsi.h>
 
-using namespace facebook;
-using namespace unum::usearch;
-
-// Helper function to get raw pointer from a Float32Array
-// Returns {float_pointer, element_count}
-std::pair<const float*, size_t> getRawVector(jsi::Runtime &runtime, const jsi::Value &val) {
-    if (!val.isObject()) {
-        throw jsi::JSError(runtime, "Invalid argument: Expected a Float32Array.");
-    }
-    jsi::Object obj = val.asObject(runtime);
-
-    // Check if it has the 'buffer' property (indicates it's a TypedArray)
-    if (!obj.hasProperty(runtime, "buffer")) {
-        throw jsi::JSError(runtime, "Invalid argument: Object must have a 'buffer' (Float32Array).");
-    }
-
-    // Access the ArrayBuffer
-    auto bufferValue = obj.getProperty(runtime, "buffer");
-    if (!bufferValue.isObject() || !bufferValue.asObject(runtime).isArrayBuffer(runtime)) {
-        throw jsi::JSError(runtime, "Internal failure: 'buffer' is not a valid ArrayBuffer.");
-    }
-    auto arrayBuffer = bufferValue.asObject(runtime).getArrayBuffer(runtime);
-
-    // Google-grade Hardening: Buffer length validation
-    if (arrayBuffer.size(runtime) == 0) {
-        throw jsi::JSError(runtime, "Invalid argument: Float32Array is empty.");
-    }
-
-    // Read offset and length to support Views (sub-arrays)
-    size_t byteOffset = 0;
-    if (obj.hasProperty(runtime, "byteOffset")) {
-        byteOffset = static_cast<size_t>(obj.getProperty(runtime, "byteOffset").asNumber());
-    }
-    
-    size_t byteLength = arrayBuffer.size(runtime); // Default full size
-    if (obj.hasProperty(runtime, "byteLength")) {
-         byteLength = static_cast<size_t>(obj.getProperty(runtime, "byteLength").asNumber());
-    }
-
-    // ZERO-COPY access to the memory pointer
-    uint8_t* rawBytes = arrayBuffer.data(runtime) + byteOffset;
-
-    // Google-grade Hardening: Memory Alignment Check
-    // Float32 needs 4-byte alignment. reinterpret_cast on unaligned memory is undefined behavior.
-    if (reinterpret_cast<uintptr_t>(rawBytes) % sizeof(float) != 0) {
-        throw jsi::JSError(runtime, "Memory Alignment Error: Float32Array buffer is not 4-byte aligned.");
-    }
-    
-    // Cast to float*
-    const float* floatPtr = reinterpret_cast<const float*>(rawBytes);
-    size_t count = byteLength / sizeof(float);
-
-    return {floatPtr, count};
-}
-
-// Helper to normalize and sanitize paths (strips file:// and prevents traversal)
-std::string normalizePath(jsi::Runtime &runtime, std::string path) {
-    // Strips file:// prefix if present
-    if (path.compare(0, 7, "file://") == 0) {
-        path = path.substr(7);
-    }
-    
-    // Security Audit: Basic prevention of path traversal
-    if (path.find("..") != std::string::npos) {
-        throw jsi::JSError(runtime, "Security violation: Path traversal is not allowed.");
-    }
-    
-    return path;
-}
-
-// Class that encapsulates a USearch index instance
-class VectorIndex : public facebook::jsi::HostObject {
-public:
-    using Index = index_dense_t;
-
-    VectorIndex(int dimensions, bool quantized) {
-        metric_kind_t metric_kind = metric_kind_t::cos_k;
-        scalar_kind_t scalar_kind = quantized ? scalar_kind_t::i8_k : scalar_kind_t::f32_k;
-        metric_punned_t metric(dimensions, metric_kind, scalar_kind);
-        
-        _index = std::make_unique<Index>(Index::make(metric));
-        if (!_index) throw std::runtime_error("Failed to initialize USearch index");
-        
-        if (!_index->reserve(100)) LOGE("Failed to reserve initial capacity");
-    }
-
-    // Expose methods to JavaScript
-    jsi::Value get(jsi::Runtime& runtime, const jsi::PropNameID& name) override {
-        std::string methodName = name.utf8(runtime);
-
-        // dimensions
-        if (methodName == "dimensions") {
-            return jsi::Value((double)_index->dimensions());
-        }
-        
-        // count
-        if (methodName == "count") {
-             if (!_index) return jsi::Value(0);
-             return jsi::Value((double)_index->size());
-        }
-
-        // memoryUsage (bytes)
-        if (methodName == "memoryUsage") {
-             if (!_index) return jsi::Value(0);
-             return jsi::Value((double)_index->memory_usage());
-        }
-
-        // delete() - Explicit Memory Release
-        if (methodName == "delete") {
-            return jsi::Function::createFromHostFunction(runtime, name, 0, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    _index.reset(); // Destroy the unique_ptr and release memory
-                    return jsi::Value::undefined();
-                });
-        }
-
-        // add(key: number, vector: Float32Array)
-        if (methodName == "add") {
-            return jsi::Function::createFromHostFunction(runtime, name, 2, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    if (count < 2) throw jsi::JSError(runtime, "add expects 2 arguments: key, vector");
-                    if (!_index) throw jsi::JSError(runtime, "VectorIndex has been deleted.");
-                    
-                    double keyDouble = arguments[0].asNumber();
-                    default_key_t key = static_cast<default_key_t>(keyDouble);
-                    
-                    auto [vecData, vecSize] = getRawVector(runtime, arguments[1]);
-                    
-                    if (vecSize != _index->dimensions()) {
-                         throw jsi::JSError(runtime, "Incorrect dimension.");
-                    }
-
-                    if (_index->size() >= _index->capacity()) _index->reserve(_index->capacity() * 2);
-                    auto result = _index->add(key, vecData);
-                    if (!result) throw jsi::JSError(runtime, "Error adding: " + std::string(result.error.what()));
-                    
-                    return jsi::Value::undefined();
-                });
-        }
-
-        // addBatch(keys: Int32Array, vectors: Float32Array)
-        if (methodName == "addBatch") {
-            return jsi::Function::createFromHostFunction(runtime, name, 2, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    if (count < 2) throw jsi::JSError(runtime, "addBatch expects 2 arguments: keys, vectors");
-                    if (!_index) throw jsi::JSError(runtime, "VectorIndex has been deleted.");
-
-                    // 1. Get Keys (Int32Array)
-                    jsi::Object keysArray = arguments[0].asObject(runtime);
-                    auto keysBufferValue = keysArray.getProperty(runtime, "buffer");
-                    auto keysBuffer = keysBufferValue.asObject(runtime).getArrayBuffer(runtime);
-                    
-                    size_t keysByteOffset = keysArray.hasProperty(runtime, "byteOffset") ? 
-                        (size_t)keysArray.getProperty(runtime, "byteOffset").asNumber() : 0;
-                    size_t keysCount = (keysArray.hasProperty(runtime, "byteLength") ? 
-                        (size_t)keysArray.getProperty(runtime, "byteLength").asNumber() : keysBuffer.size(runtime)) / sizeof(int32_t);
-                    
-                    const int32_t* keysData = reinterpret_cast<const int32_t*>(keysBuffer.data(runtime) + keysByteOffset);
-
-                    // 2. Get Vectors (Float32Array)
-                    auto [vecData, vecTotalElements] = getRawVector(runtime, arguments[1]);
-                    size_t dims = _index->dimensions();
-                    size_t batchCount = vecTotalElements / dims;
-
-                    // 3. Validation
-                    if (batchCount != keysCount) {
-                        throw jsi::JSError(runtime, "Batch mismatch: keys and vectors must have compatible sizes.");
-                    }
-
-                    // 4. Optimization: Single Reserve
-                    if (_index->size() + batchCount > _index->capacity()) {
-                        _index->reserve(_index->size() + batchCount);
-                    }
-
-                    // 5. Bulk Add
-                    for (size_t i = 0; i < batchCount; ++i) {
-                        auto result = _index->add((default_key_t)keysData[i], vecData + (i * dims));
-                        if (!result) throw jsi::JSError(runtime, "Error adding in batch at index " + std::to_string(i));
-                    }
-
-                    return jsi::Value::undefined();
-                });
-        }
-
-        // search(vector: Float32Array, count: number)
-        if (methodName == "search") {
-            return jsi::Function::createFromHostFunction(runtime, name, 2, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    if (count < 2) throw jsi::JSError(runtime, "search expects 2 arguments: vector, count");
-                    if (!_index) throw jsi::JSError(runtime, "VectorIndex has been deleted.");
-
-                    auto [queryData, querySize] = getRawVector(runtime, arguments[0]);
-                    int resultsCount = static_cast<int>(arguments[1].asNumber());
-
-                    if (querySize != _index->dimensions()) {
-                         throw jsi::JSError(runtime, "Query vector dimension mismatch.");
-                    }
-
-                    auto results = _index->search(queryData, resultsCount);
-                    jsi::Array returnArray(runtime, results.size());
-                    for (size_t i = 0; i < results.size(); ++i) {
-                        auto pair = results[i];
-                        jsi::Object resultObj(runtime);
-                        resultObj.setProperty(runtime, "key", static_cast<double>(pair.member.key));
-                        resultObj.setProperty(runtime, "distance", static_cast<double>(pair.distance));
-                        returnArray.setValueAtIndex(runtime, i, resultObj);
-                    }
-                    return returnArray;
-                });
-        }
-
-        // save(path: string)
-        if (methodName == "save") {
-            return jsi::Function::createFromHostFunction(runtime, name, 1, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    if (count < 1 || !arguments[0].isString()) throw jsi::JSError(runtime, "save expects path");
-                    std::string path = normalizePath(runtime, arguments[0].asString(runtime).utf8(runtime));
-                    
-                    if (!_index->save(path.c_str())) throw jsi::JSError(runtime, "Critical error saving index to disk: " + path);
-                    return jsi::Value::undefined();
-                });
-        }
-
-        // load(path: string)
-        if (methodName == "load") {
-            return jsi::Function::createFromHostFunction(runtime, name, 1, 
-                [this](jsi::Runtime& runtime, const jsi::Value& thisValue, const jsi::Value* arguments, size_t count) -> jsi::Value {
-                    if (count < 1 || !arguments[0].isString()) throw jsi::JSError(runtime, "load expects path");
-                    std::string path = normalizePath(runtime, arguments[0].asString(runtime).utf8(runtime));
-                    
-                    if (!_index->load(path.c_str())) throw jsi::JSError(runtime, "Critical error loading index from disk: " + path);
-                    return jsi::Value::undefined();
-                });
-        }
-
-        return jsi::Value::undefined();
-    }
-
-private:
-    std::unique_ptr<Index> _index;
-};
-
-// --------------------------------------------------------------------------
-// Module Installation (Factory)
-// --------------------------------------------------------------------------
-
 extern "C" JNIEXPORT void JNICALL
-Java_expo_modules_vectorsearch_ExpoVectorSearchModule_nativeInstall(JNIEnv* env, jobject thiz, jlong jsiPtr) {
-    auto runtime = reinterpret_cast<jsi::Runtime*>(jsiPtr);
-    if (!runtime) return;
-    auto &rt = *runtime;
-
-    auto global = rt.global();
-    auto moduleObj = jsi::Object(rt);
-
-    // ExpoVectorSearch.createIndex(dimensions: number, options?: { quantization: 'f32' | 'i8' }) -> VectorIndex
-    moduleObj.setProperty(rt, "createIndex", jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, "createIndex"), 1, 
-        [](jsi::Runtime &rt, const jsi::Value &thisValue, const jsi::Value *args, size_t count) -> jsi::Value {
-            if (count < 1 || !args[0].isNumber()) {
-                throw jsi::JSError(rt, "createIndex expects at least 1 argument: dimensions");
-            }
-            int dims = static_cast<int>(args[0].asNumber());
-            
-            bool quantized = false;
-            if (count > 1 && args[1].isObject()) {
-                jsi::Object options = args[1].asObject(rt);
-                if (options.hasProperty(rt, "quantization")) {
-                    std::string q = options.getProperty(rt, "quantization").asString(rt).utf8(rt);
-                    if (q == "i8") quantized = true;
-                }
-            }
-            
-            auto indexInstance = std::make_shared<VectorIndex>(dims, quantized);
-            return jsi::Object::createFromHostObject(rt, indexInstance);
-        }));
-
-    global.setProperty(rt, "ExpoVectorSearch", moduleObj);
+Java_expo_modules_vectorsearch_ExpoVectorSearchModule_nativeInstall(
+    JNIEnv *env, jobject thiz, jlong jsiPtr) {
+  auto runtime = reinterpret_cast<facebook::jsi::Runtime *>(jsiPtr);
+  if (runtime) {
+    expo::vectorsearch::install(*runtime);
+  }
 }

package/build/index.d.ts

@@ -0,0 +1,5 @@
+export { VectorIndex } from './src/ExpoVectorSearchModule';
+export { useVectorSearch } from './src/useVectorSearch';
+import { VectorIndex } from './src/ExpoVectorSearchModule';
+export default VectorIndex;
+//# sourceMappingURL=index.d.ts.map

package/build/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAGxD,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,eAAe,WAAW,CAAC"}

package/build/index.js

@@ -0,0 +1,6 @@
+export { VectorIndex } from './src/ExpoVectorSearchModule';
+export { useVectorSearch } from './src/useVectorSearch';
+// Export default module (VectorIndex class)
+import { VectorIndex } from './src/ExpoVectorSearchModule';
+export default VectorIndex;
+//# sourceMappingURL=index.js.map

package/build/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAExD,4CAA4C;AAC5C,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,eAAe,WAAW,CAAC","sourcesContent":["export { VectorIndex } from './src/ExpoVectorSearchModule';\nexport { useVectorSearch } from './src/useVectorSearch';\n\n// Export default module (VectorIndex class)\nimport { VectorIndex } from './src/ExpoVectorSearchModule';\nexport default VectorIndex;"]}

package/build/src/ExpoVectorSearch.types.d.ts

@@ -0,0 +1,7 @@
+export type Vector = Float32Array;
+export type DistanceMetric = 'cos' | 'l2sq' | 'ip' | 'hamming' | 'jaccard';
+export type SearchResult = {
+    key: number;
+    distance: number;
+};
+//# sourceMappingURL=ExpoVectorSearch.types.d.ts.map

package/build/src/ExpoVectorSearch.types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ExpoVectorSearch.types.d.ts","sourceRoot":"","sources":["../../src/ExpoVectorSearch.types.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,MAAM,GAAG,YAAY,CAAC;AAElC,MAAM,MAAM,cAAc,GAAG,KAAK,GAAG,MAAM,GAAG,IAAI,GAAG,SAAS,GAAG,SAAS,CAAC;AAE3E,MAAM,MAAM,YAAY,GAAG;IACzB,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;CAClB,CAAC"}

package/build/src/ExpoVectorSearch.types.js

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=ExpoVectorSearch.types.js.map

package/build/src/ExpoVectorSearch.types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ExpoVectorSearch.types.js","sourceRoot":"","sources":["../../src/ExpoVectorSearch.types.ts"],"names":[],"mappings":"","sourcesContent":["export type Vector = Float32Array;\n\nexport type DistanceMetric = 'cos' | 'l2sq' | 'ip' | 'hamming' | 'jaccard';\n\nexport type SearchResult = {\n  key: number;\n  distance: number;\n};\n"]}