core/model_management/model_specs library

Model specification value types — dart:io-free, shared across all platforms (mobile, desktop, web).

These types used to be part of flutter_edge_ai_mobile.dart, which pulled dart:io (and path_provider) into the public import graph and broke dart2wasm compatibility (pub.dev "Platform support" WASM check). They are extracted here as a standalone, platform-neutral library so the public API surface and the platform interface can depend on the specs without dragging in the mobile implementation's dart:io.

Classes

DirectoryBundleFile
One file of a DIRECTORY inference model (ORT-GenAI): the model is a set of files that must live together in a per-model subdirectory with their BARE leaf names (genai_config.json references its siblings by bare name and the native loader is handed the directory). filename is the EXPLICIT install identity <modelId>/<bareLeaf> (a real subpath) — NEVER derived from source, so the on-disk id and the repository key always agree (a FileSource-derived basename would drop the subdir and break the repo lookup on every restore). The PRIMARY file (genai_config.json) uses PreferencesKeys.installedModelFileName as its prefsKey so the persist-identity + modelFilePaths.values.first both find it; siblings use their bare leaf name so each maps to a distinct getModelFilePaths entry.
DownloadProgress
Progress information for model downloads
EmbeddingModelFile
Model file for embedding models (.bin files)
EmbeddingModelSpec
Specification for embedding models (model.bin + tokenizer.json)
EmbeddingTokenizerFile
Tokenizer file for embedding models (.json files)
InferenceModelFile
Model file for inference models (.bin, .task files)
InferenceModelSpec
Specification for inference models (main model + optional LoRA)
LoraModelFile
Model file for LoRA weights. A COMPANION of the inference model — its filename gets the modelId__ prefix so a LoRA adapter travels with the exact base model it was fine-tuned for (mirrors the tokenizer/aux namespacing used by embedding/STT/TTS companions).
ModelFile
Represents a single file that belongs to a model
ModelSpec
Base specification for any model (inference or embedding)
OrphanedFileInfo
Information about a potentially orphaned file
StorageStats
Storage statistics
SttModelFile
Model file for STT models (.tflite files)
SttModelSpec
Specification for STT models (model.tflite + tokenizer.json).
SttTokenizerFile
Tokenizer file for STT models (tokenizer.json)
TtsAbsoluteUrl
A full absolute URL fetched as-is — the install base URL is ignored.
TtsAssetBase
A Flutter asset directory, laid out like the model's Hugging Face repo.
TtsBundleBase
Where a whole TTS bundle is installed from — the one base every manifest file is resolved against by TtsModelTypeManifest.sourceFor. Internal: the public surface is TtsInstallationBuilder's fromNetwork / fromAsset / fromFile / fromBundled.
TtsBundledBase
Native bundled resources, one per manifest file, named <type>__<basename> (flat: a bundled resource name cannot contain /).
TtsBundleFile
One file of a TTS model bundle. filename is namespaced by the owning TtsModelType (every bundle member — including the .tflite graphs — since a TTS bundle has no single distinguished "model" file the way Inference/Embedding/STT specs do). prefsKey stays the PLAIN manifest basename — NOT the namespaced filename — because flutter_edge_ai_speech's TtsModelProfile (e.g. configFile = 'config.json') and TtsCore/MatchaTextFrontend look bundle paths up by that plain name (paths[profile.configFile]); decoupling keeps that cross-package contract intact while the on-disk/repository identity is namespaced.
TtsFetchLocation
Where one TTS bundle member is fetched from — the typed result of TtsModelTypeManifest.fetchLocationFor. Exactly two shapes exist: a TtsRelativeSuffix resolved against the model's install base URL, or a TtsAbsoluteUrl fetched as-is (a cross-repo file, e.g. Inflect's reused Matcha G2P bundle). TtsModelTypeManifest.sourceFor exhaustively switches on this, so the two cases can never be confused by string sniffing.
TtsFileBase
An absolute directory on disk, laid out like the model's Hugging Face repo.
TtsModelSpec
Specification for a TTS model — a SELECTABLE bundle. ttsModelType carries the model family so one generic backend dispatches to the right runtime profile; sources is one entry per file in that type's manifest.
TtsNetworkBase
A URL the bundle is served under, laid out like the model's Hugging Face repo.
TtsRelativeSuffix
A relative path segment appended to the install base URL (the plain basename itself, or a tables//voices/ subpath for qwen3's embedding-table/demo-voice members).

Enums

ModelManagementType
Base enumeration for different model management types
ModelReplacePolicy
Policy for what happens to a previously-installed model when a new one is set active. Lives with the spec value types (it's a per-spec install policy); re-exported from model_file_manager_interface.dart for backward compat.
SttModelType
Speech-to-text model families supported by the pluggable STT backends. Only moonshine has a shipped SttModelProfile/pipeline (flutter_edge_ai_speech); the others are follow-ons that need a log-mel frontend.
TtsModelType
Text-to-speech model families supported by the pluggable TTS backends. matcha, qwen3 and inflect have shipped TtsModelProfile/pipelines (flutter_edge_ai_speech); kokoro/supertonic are documented follow-ons (fail-loud until wired).

Extensions

TtsModelTypeManifest on TtsModelType
The filenames a given TTS model needs, installed together as a bundle from one source. Fail-loud for unwired families.

Exceptions / Errors

ModelDownloadException
Exception thrown when model download fails
ModelException
Base exception for model management operations
ModelStorageException
Exception thrown when model storage operations fail
ModelValidationException
Exception thrown when model file validation fails