ONNX search provider

Package.swift

@@ -1,6 +1,48 @@
 // swift-tools-version: 6.3
 
 import PackageDescription
+import class Foundation.ProcessInfo
+import struct Foundation.URL
+
+let environment = ProcessInfo.processInfo.environment
+let packageRoot = URL(fileURLWithPath: #filePath).deletingLastPathComponent().path
+
+func packagePath(_ path: String) -> String {
+	path.hasPrefix("/") ? path : "\(packageRoot)/\(path)"
+}
+
+let onnxRuntimeRoot = packagePath(environment["LEXICON_ONNX_RUNTIME_ROOT"] ?? ".build/onnx-runtime/current")
+let onnxRuntimePlatform = environment["LEXICON_ONNX_RUNTIME_PLATFORM"] ?? "linux-x64"
+let onnxRuntimeInclude = packagePath(environment["LEXICON_ONNX_RUNTIME_INCLUDE"] ?? "\(onnxRuntimeRoot)/include")
+let onnxRuntimeLibrary = packagePath(environment["LEXICON_ONNX_RUNTIME_LIB"] ?? "\(onnxRuntimeRoot)/lib/\(onnxRuntimePlatform)")
+let onnxRuntimeLibraryName = onnxRuntimePlatform == "windows-x64" ? "onnxruntime.dll" : "libonnxruntime.so"
+#if os(macOS)
+let useAppleONNXRuntimeBindings = (environment["LEXICON_ONNX_USE_APPLE_BINDINGS"] ?? "1") != "0"
+	&& environment["LEXICON_ONNX_RUNTIME_PLATFORM"] == nil
+	&& environment["LEXICON_ONNX_RUNTIME_INCLUDE"] == nil
+	&& environment["LEXICON_ONNX_RUNTIME_LIB"] == nil
+#else
+let useAppleONNXRuntimeBindings = false
+#endif
+let onnxRuntimeDependencies: [Target.Dependency] = useAppleONNXRuntimeBindings
+	? [
+		.product(
+			name: "onnxruntime",
+			package: "onnxruntime-swift-package-manager",
+			condition: .when(traits: ["ONNXSearch"])
+		)
+	]
+	: [
+		.target(
+			name: "CLexiconONNXRuntime",
+			condition: .when(traits: ["ONNXSearch"])
+		)
+	]
+let onnxPackageDependencies: [Package.Dependency] = useAppleONNXRuntimeBindings
+	? [
+		.package(url: "https://github.com/microsoft/onnxruntime-swift-package-manager", exact: "1.24.2")
+	]
+	: []
 
 let package = Package(
 	name: "Lexicon",
@@ -13,15 +55,18 @@ let package = Package(
 		.library(name: "_JSON", targets: ["_JSON"]),
 		.library(name: "Lexicon", targets: ["Lexicon"]),
 		.library(name: "LexiconSearchMLX", targets: ["LexiconSearchMLX"]),
+		.library(name: "LexiconSearchONNX", targets: ["LexiconSearchONNX"]),
 		.library(name: "SwiftLexicon", targets: ["SwiftLexicon"]),
 		.library(name: "LexiconGenerators", targets: ["LexiconGenerators"]),
 		.executable(name: "lexicon-generate", targets: ["lexicon-generate"]),
 		.executable(name: "lexicon", targets: ["lexicon-cli"]),
 		.plugin(name: "SwiftStandAloneGeneratorPlugin", targets: ["SwiftStandAloneGeneratorPlugin"]),
 		.plugin(name: "SwiftLibraryGeneratorPlugin", targets: ["SwiftLibraryGeneratorPlugin"]),
+		.plugin(name: "ONNXSearchArtifactsPlugin", targets: ["ONNXSearchArtifactsPlugin"]),
 	],
 	traits: [
 		.trait(name: "MLXSearch"),
+		.trait(name: "ONNXSearch"),
 	],
 	dependencies: [
 		.package(url: "https://github.com/screensailor/Hope", branch: "trunk"),
@@ -33,7 +78,7 @@ let package = Package(
 		.package(url: "https://github.com/ml-explore/mlx-swift-lm", .upToNextMajor(from: "3.31.3")),
 		.package(url: "https://github.com/DePasqualeOrg/swift-hf-api-mlx", exact: "0.2.0"),
 		.package(url: "https://github.com/DePasqualeOrg/swift-tokenizers", from: "0.6.3"),
-	],
+	] + onnxPackageDependencies,
 	targets: [
 		.target(
 			name: "_Collections",
@@ -113,6 +158,7 @@ let package = Package(
 			dependencies: [
 				"Lexicon",
 				.target(name: "LexiconSearchMLX", condition: .when(traits: ["MLXSearch"])),
+				.target(name: "LexiconSearchONNX", condition: .when(traits: ["ONNXSearch"])),
 				"LexiconGenerators",
 				.product(name: "ArgumentParser", package: "swift-argument-parser"),
 			]
@@ -128,6 +174,43 @@ let package = Package(
 				.product(name: "Tokenizers", package: "swift-tokenizers", condition: .when(traits: ["MLXSearch"])),
 			]
 		),
+		.target(
+			name: "LexiconSearchONNX",
+			dependencies: ["Lexicon"] + onnxRuntimeDependencies
+		),
+		.target(
+			name: "CLexiconONNXRuntime",
+			publicHeadersPath: "include",
+			cSettings: [
+				.define("LEXICON_ONNX_RUNTIME_REQUIRED", .when(traits: ["ONNXSearch"])),
+				.define("LEXICON_ONNX_RUNTIME_LIBRARY_PATH", to: "\"\(onnxRuntimeLibrary)/\(onnxRuntimeLibraryName)\""),
+				.unsafeFlags(["-I", onnxRuntimeInclude])
+			]
+		),
+		.testTarget(
+			name: "LexiconSearchONNXTests",
+			dependencies: [
+				"Lexicon",
+				"LexiconSearchONNX",
+			]
+		),
+		.executableTarget(
+			name: "onnx-search-artifacts",
+			dependencies: ["LexiconSearchONNX"]
+		),
+		.plugin(
+			name: "ONNXSearchArtifactsPlugin",
+			capability: .command(
+				intent: .custom(
+					verb: "setup-onnx-search-artifacts",
+					description: "Download ONNX search model and runtime artifacts."
+				),
+				permissions: [
+					.writeToPackageDirectory(reason: "Stores ONNX search artifacts under .build/onnx-search and .build/onnx-runtime.")
+				]
+			),
+			dependencies: ["onnx-search-artifacts"]
+		),
 		.plugin(
 			name: "SwiftStandAloneGeneratorPlugin",
 			capability: .buildTool(),

Plugins/ONNXSearchArtifactsPlugin/plugin.swift

@@ -0,0 +1,39 @@
+import Foundation
+import PackagePlugin
+
+@main
+struct ONNXSearchArtifactsPlugin: CommandPlugin {
+
+	func performCommand(context: PluginContext, arguments: [String]) async throws {
+		let tool = try context.tool(named: "onnx-search-artifacts")
+		var arguments = arguments
+		if arguments.first == "--" {
+			arguments.removeFirst()
+		}
+		let output = context.package.directoryURL
+			.appendingPathComponent(".build", isDirectory: true)
+			.appendingPathComponent("onnx-search", isDirectory: true)
+		let runtimeOutput = context.package.directoryURL
+			.appendingPathComponent(".build", isDirectory: true)
+			.appendingPathComponent("onnx-runtime", isDirectory: true)
+		let process = try Process.run(
+			tool.url,
+			arguments: [
+				"--output", output.path,
+				"--runtime-output", runtimeOutput.path,
+			] + arguments
+		)
+		process.waitUntilExit()
+		if process.terminationStatus != 0 {
+			throw ONNXSearchArtifactsPluginError(status: process.terminationStatus)
+		}
+	}
+}
+
+private struct ONNXSearchArtifactsPluginError: Error, CustomStringConvertible {
+	var status: Int32
+
+	var description: String {
+		"ONNX search artifact setup failed with exit status \(status)."
+	}
+}

README.md

@@ -337,7 +337,23 @@ swift run --traits MLXSearch lexicon search commerce.lexicon "order submit" \
 
 MLX document embeddings are cached under the user cache directory by default. The first semantic search builds the cache and logs indexing progress to stderr; pass `--embedding-cache` or `--rebuild-embeddings` to control that cache.
 
-See [`docs/search.md`](docs/search.md) for the search mode guide and demo lexicon examples.
+For a portable ONNX Runtime backend, fetch the model/runtime artifacts with the SwiftPM command plugin, then build with the `ONNXSearch` trait:
+
+```sh
+swift package --disable-sandbox --allow-writing-to-package-directory \
+	setup-onnx-search-artifacts -- \
+	--preset bge-small-en-v1.5 \
+	--runtime linux-x64
+
+swift run --traits ONNXSearch lexicon search commerce.lexicon "order submit" \
+	--mode semantic \
+	--embedding-provider onnx \
+	--embedding-model-preset bge-small-en-v1.5
+```
+
+ONNX presets currently include MiniLM, BGE-small, GTE-small and E5-small-v2. BGE-small is a good first ONNX model to try for semantic search, but `hybrid` remains the default because exact graph names and references still matter.
+
+See [`docs/search.md`](docs/search.md) for the search mode guide and demo lexicon examples. See [`docs/search-platforms.md`](docs/search-platforms.md) for Linux, Android, Windows, and ONNX backend notes.
 
 Editing commands print the updated document to stdout by default. Pass `-o` or `--output` to write another file.
 
@@ -384,12 +400,15 @@ Event payloads are JSON-backed and decode through `JSONDecoder`, so generated le
 | `Lexicon` | Core document, graph, parser, composition, branch and CRDT model. |
 | `SwiftLexicon` | Runtime support for generated Swift lexicons and event streams. |
 | `LexiconGenerators` | Code generators and generator registry. |
+| `LexiconSearchMLX` | Optional MLX embedding provider for semantic search. |
+| `LexiconSearchONNX` | Optional ONNX Runtime embedding provider for semantic search. |
 | `_JSON` | JSON value and decoder-backed typed access. |
 | `_Collections` | Internal collection utilities, including sorted dictionary support. |
 | `lexicon` | CLI for validating, inspecting, formatting, diffing and editing lexicons. |
 | `lexicon-generate` | CLI for generating source artefacts. |
 | `SwiftLibraryGeneratorPlugin` | SwiftPM plugin for generated Swift that depends on `SwiftLexicon`. |
 | `SwiftStandAloneGeneratorPlugin` | SwiftPM plugin for stand-alone generated Swift. |
+| `ONNXSearchArtifactsPlugin` | SwiftPM command plugin for downloading ONNX search model and runtime artifacts. |
 
 ## Installation
 
@@ -412,9 +431,9 @@ The package currently declares Swift 6.3, Swift language mode 6, macOS 15 and iO
 
 ## Platform Support
 
-CI runs SwiftPM tests on macOS and Linux. It also runs tests on an Android emulator and cross-builds Android ARM64.
+CI runs SwiftPM builds and tests on macOS and Linux for pull requests. Android runs on `trunk` pushes and manual workflow dispatches to keep pull requests fast.
 
-On Apple platforms, sentence graph generation uses NaturalLanguage. On Linux and Android, Lexicon uses a deterministic fallback so the API remains available.
+On Apple platforms, sentence graph generation uses NaturalLanguage. On Linux and Android, Lexicon uses a deterministic fallback so the API remains available. Semantic search can use MLX on Apple platforms or ONNX Runtime through the optional `ONNXSearch` trait; the ONNX C shim is designed for Linux and Android now, with Windows runtime artifacts present but session path bridging still to finish.
 
 See [`docs/platforms.md`](docs/platforms.md) for details.