localm-web 0.1.0

package/CHANGELOG.md

@@ -0,0 +1,51 @@
+# Changelog
+
+All notable changes to **localm-web** will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.1.0] - 2026-05-10
+
+### Added
+
+- Initial project scaffolding: TypeScript strict configuration, ESM-only Vite library build,
+  Vitest test runner, ESLint + Prettier.
+- Public type primitives: `Message`, `Role`, `FinishReason`, `GenerationOptions`,
+  `ModelLoadProgress`, `ProgressCallback`, `TokenChunk`, `ModelPreset`.
+- Error hierarchy rooted at `LocalmWebError` (`WebGPUUnavailableError`,
+  `ModelLoadError`, `ModelNotLoadedError`, `UnknownModelError`,
+  `GenerationAbortedError`, `QuotaExceededError`, `BackendNotAvailableError`).
+- Runtime-agnostic `Engine` interface and concrete `WebLLMEngine` backed by `@mlc-ai/web-llm`.
+- `Chat` task with `send()`, `stream()`, `setSystemPrompt()`, `resetHistory()`, `getHistory()`,
+  `unload()`. Streaming honors `AbortSignal`.
+- Curated model registry: `phi-3.5-mini-int4`, `llama-3.2-1b-int4`, `qwen2.5-1.5b-int4`.
+- Streaming helpers: `collectStream`, `tap`.
+- Vite example app under `examples/vite-chat/` demonstrating model loading,
+  streaming output and abort.
+- Unit tests for presets, exceptions, results, streaming and Chat (with a fake engine).
+- Release pipeline: `Makefile` + `scripts/release.sh` (release branch + tag + PR flow,
+  PT-BR PR template), `RELEASES.md` autogenerated from git tags.
+- GitHub Actions workflows: `ci.yml` (Node 18/20/22 matrix) and `release-npm.yml`
+  (publish on `v*.*.*` tag with `npm publish --provenance`).
+- Security policy documented in `CLAUDE.md` and `README.md`: runtime vs dev-deps split,
+  zero-CVE bar for `peerDependencies`, `npm audit` on every release, provenance signing.
+
+### Changed
+
+- Bumped `vite` from `^5.4.0` to `^7.3.3` and `vitest` from `^2.1.0` to `^3.2.4`
+  to clear advisories GHSA-67mh-4wv8-2f99 (esbuild dev-server CORS) and
+  GHSA-4w7w-66w2-5vf9 (vite path traversal in optimized deps `.map`).
+- Added `overrides.esbuild ^0.25.0` in `package.json` as defense-in-depth against
+  transitive esbuild downgrades.
+- ESLint flat config: expanded `ignores` to cover config files and `examples/**`.
+- `tsconfig.test.json`: explicit `exclude` so test files are picked up by the
+  TypeScript-aware ESLint parser.
+- `package.json` `lint` script: dropped `--ext .ts,.tsx` (no-op in flat config).
+
+### Notes
+
+- First public release. `Chat` task is the only fully implemented task — `Completion`,
+  `Embeddings`, `Reranker`, structured output and ORT-Web fallback are scheduled for v0.2–v0.5.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Mauricio Benjamin
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,225 @@
+# localm-web
+
+> ⚠️ **Status: pre-alpha.** Public API is being designed and is expected to change. Code in this repo is intentionally minimal until v0.1.
+
+Browser-only TypeScript SDK for running Language Models (LLMs and SLMs) **locally in the user's browser**, with a developer experience modeled directly on [`ort-vision-sdk-web`](https://github.com/mauriciobenjamin700/ort-vision-sdk).
+
+```typescript
+import { Chat } from "localm-web";
+
+const chat = await Chat.create("phi-3.5-mini-int4");
+
+for await (const token of chat.stream("Explain ONNX in one sentence.")) {
+  process.stdout.write(token.text);
+}
+```
+
+That's it. No server, no API key, no roundtrip — the model runs **on the user's GPU** via WebGPU.
+
+---
+
+## Why does this exist?
+
+The Python ecosystem for local Language Models is saturated: `llama-cpp-python`, Ollama, vLLM, `transformers`, text-generation-inference, and dozens more. Picking up another Python wrapper adds nothing.
+
+The **browser side is different**. The closest equivalents are:
+
+| Project                                                           | What it is                     | Why it's not enough                                               |
+| ----------------------------------------------------------------- | ------------------------------ | ----------------------------------------------------------------- |
+| [WebLLM (MLC)](https://github.com/mlc-ai/web-llm)                 | Best-in-class WebGPU runtime   | Engine-centric, low-level API, no opinionated tasks               |
+| [transformers.js](https://github.com/huggingface/transformers.js) | HF pipeline API in the browser | Slower (no WebGPU-first compilation in many paths), broad surface |
+| `onnxruntime-genai-web`                                           | Microsoft's web LM build       | Preview, unstable, no high-level tasks                            |
+
+There is no opinionated, task-oriented, strict-typed, **Ultralytics-style SDK that just works in a Vite app**. `localm-web` fills that gap.
+
+The mental model is straightforward: if [`ort-vision-sdk-web`](https://github.com/mauriciobenjamin700/ort-vision-sdk) is what `Detector` / `Classifier` / `Segmenter` look like for vision, then `localm-web` is what `Chat` / `Completion` / `Embeddings` / `Reranker` look like for language.
+
+## Design principles
+
+1. **Browser-only.** No Node target, no server runtime. If your code runs on a backend, this SDK is the wrong tool — use `transformers`, vLLM, Ollama, or any of the dozens of mature Python options.
+2. **Maximum performance.** WebGPU-first via WebLLM (MLC). Web Worker execution by default so the UI thread stays free. WASM-SIMD fallback for non-WebGPU browsers from v0.5.
+3. **Ultralytics-style DX.** `await Class.create(model)` then `predict()` / `send()` / `embed()` / `score()`. Mirrors `ort-vision-sdk-web` so a developer using both feels continuity.
+4. **ESM only.** No CJS, no UMD, no IIFE. The browser is ESM-native, modern bundlers expect ESM, and shipping multiple formats just bloats the package.
+5. **Vite-first.** The build is optimized for Vite 5+ consumers. Other bundlers will still work, but Vite is the supported smooth path.
+6. **Not tied to Vercel.** No `vercel.json`, no Next-specific helpers, no Edge runtime exports. Examples deploy to any static host (Cloudflare Pages, Netlify, GitHub Pages, S3, self-hosted).
+7. **Wrap, don't fork.** WebLLM stays a peer dependency. We add the API layer, the task abstractions, and the missing pieces (embeddings, reranker, structured output, fallback runtime).
+
+## Scope
+
+### In scope
+
+- Browser-only execution (WebGPU primary, WASM-SIMD fallback from v0.5).
+- High-level tasks: `Chat`, `Completion`, `Embeddings`, `Reranker`.
+- Streaming token output via async generators with `AbortSignal` support.
+- Tokenization, chat templates, sampling, KV cache (delegated to the underlying runtime).
+- Model caching (Cache API + OPFS) with resume on interrupted downloads.
+- Curated registry of supported SLMs: Phi-3.5-mini, Llama-3.2-1B/3B, Qwen2.5-0.5B/1.5B/3B, Gemma-2-2B, SmolLM2.
+- Structured output: JSON Schema → constrained decoding.
+- Web Worker execution out of the box.
+
+### Out of scope
+
+- Server-side execution (Node, Bun, Deno).
+- Training, fine-tuning, LoRA loading.
+- Multi-modal models at v1.0 (a future composite SDK may combine `ort-vision-sdk-web` + `localm-web`).
+- A llama.cpp / GGUF backend — community-maintained options exist; that's not our differentiation.
+- A pre-built chat UI. This is an SDK, not a chatbot kit.
+- Bundling model weights into the package — models are downloaded at runtime.
+- Non-ESM module formats.
+
+## Architecture
+
+```
+localm-web/
+├── src/
+│   ├── core/         # backend abstraction + WebLLM / ORT-Web engines
+│   ├── tasks/        # Chat, Completion, Embeddings, Reranker
+│   ├── io/           # tokenizer + chat-template loaders
+│   ├── sampling/     # greedy, top-k, top-p, temperature
+│   ├── cache/        # KV cache + model file cache (Cache API / OPFS)
+│   ├── streaming/    # async iterator + AbortSignal plumbing
+│   ├── structured/   # JSON Schema → grammar / logit-mask
+│   ├── presets/      # curated model registry
+│   ├── worker/       # Web Worker entrypoint for inference
+│   ├── results.ts    # typed result classes
+│   ├── types.ts      # primitive types (Message, ChatRequest, etc.)
+│   └── index.ts      # public API
+├── test/
+├── examples/
+├── docs/
+└── ...
+```
+
+A full layer-by-layer breakdown lives in [CLAUDE.md](./CLAUDE.md).
+
+## Tech stack
+
+- **Language:** TypeScript 5.4+, strict mode, ES2022 target.
+- **Module format:** ESM only.
+- **Build:** Vite 5+ in library mode, `tsc` for declarations.
+- **Primary runtime:** [WebLLM (MLC)](https://github.com/mlc-ai/web-llm), Apache 2.0, WebGPU-first.
+- **Fallback runtime (v0.5+):** [`onnxruntime-web`](https://github.com/microsoft/onnxruntime) + [`@huggingface/transformers`](https://github.com/huggingface/transformers.js).
+- **Tokenizer:** `@huggingface/transformers` tokenizer module.
+- **Chat templates:** `@huggingface/jinja`.
+- **Storage:** Cache API + OPFS (Origin Private File System).
+- **Concurrency:** Web Worker via `Comlink` (or native `MessagePort`).
+- **Tests:** Vitest + Playwright (real browser for WebGPU).
+- **Lint/format:** ESLint + Prettier.
+
+## Public API (target shape)
+
+```typescript
+import { Chat, Completion, Embeddings, Reranker } from "localm-web";
+
+// Chat — multi-turn conversation with chat template applied
+const chat = await Chat.create("phi-3.5-mini-int4");
+const reply = await chat.send("Explain ONNX in one sentence.");
+console.log(reply.text);
+
+// Streaming
+const controller = new AbortController();
+for await (const token of chat.stream("Explain ONNX.", { signal: controller.signal })) {
+  process.stdout.write(token.text);
+}
+
+// Completion — raw text-in text-out (no chat template)
+const comp = await Completion.create("qwen2.5-0.5b-int4");
+const out = await comp.predict("Once upon a time", { maxTokens: 100 });
+
+// Embeddings
+const emb = await Embeddings.create("bge-small-en-v1.5");
+const vectors = await emb.embed(["hello world", "another sentence"]);
+
+// Reranker
+const rerank = await Reranker.create("bge-reranker-base");
+const scores = await rerank.score("query", ["doc1", "doc2", "doc3"]);
+
+// Structured output (JSON Schema → constrained decoding)
+const json = await chat.send("Extract user info from: ...", {
+  jsonSchema: { type: "object", properties: { name: { type: "string" } } },
+});
+```
+
+The shape mirrors `ort-vision-sdk-web`: `await Class.create(model)` then `predict()` / `send()` / `embed()` / `score()`.
+
+## Versioning roadmap
+
+| Version  | Scope                                                                                        |
+| -------- | -------------------------------------------------------------------------------------------- |
+| **v0.1** | `Chat` via WebLLM. Phi-3.5-mini, Llama-3.2-1B, Qwen2.5-1.5B. Streaming with `AbortSignal`.   |
+| **v0.2** | `Completion` task. Model caching (Cache API + OPFS). Web Worker by default. Progress events. |
+| **v0.3** | `Embeddings` and `Reranker` tasks. BGE family via transformers.js.                           |
+| **v0.4** | Structured output (JSON Schema → grammar / logit masking).                                   |
+| **v0.5** | ORT-Web fallback for browsers without WebGPU. Auto-detection and graceful degradation.       |
+| **v0.6** | Function calling helper (tool use with schema-validated arguments).                          |
+| **v1.0** | Documentation site, runnable demos, stable API contract.                                     |
+
+## Browser support
+
+- **WebGPU:** Chrome 113+, Edge 113+, recent Firefox Nightly with `dom.webgpu.enabled`, Safari 18+ on macOS Sonoma+ / iOS 18+.
+- **Without WebGPU:** from v0.5, a WASM-SIMD fallback path will run smaller models acceptably. Below v0.5, a clear runtime error is raised when WebGPU is missing.
+
+## Installation
+
+> Not yet published. Once v0.1 ships:
+
+```bash
+npm install localm-web @mlc-ai/web-llm
+```
+
+`@mlc-ai/web-llm` is a peer dependency — the consumer pins the version, which keeps the SDK lightweight and avoids version conflicts.
+
+## Vite usage
+
+The package is designed to drop into a Vite app with no extra config. The Web Worker is bundled via Vite's native worker support; just import the SDK and use it.
+
+A complete example will live under `examples/vite-chat/` once v0.1 lands.
+
+## Why not server-side?
+
+Three reasons:
+
+1. **Mature alternatives exist.** Python and TS already have excellent server-side LM tooling (Ollama, vLLM, llama-cpp-python, transformers, llama.cpp Node bindings). Adding another wrapper is noise.
+2. **The browser is the underserved surface.** Running models on the user's device removes the server cost, keeps data local, and unlocks offline use cases — but the DX is currently rough.
+3. **Different concerns.** Server inference cares about throughput, batching, multi-tenant scheduling. Browser inference cares about cold-start time, model caching, UI thread isolation, WebGPU compatibility. Conflating them produces a bad SDK on both sides.
+
+## Security
+
+`localm-web` is a browser SDK — its dependencies execute in your users' browsers. Two layers, treated differently:
+
+| Layer               | What it is                              | Vuln policy                                                                             |
+| ------------------- | --------------------------------------- | --------------------------------------------------------------------------------------- |
+| **Runtime** (peers) | `@mlc-ai/web-llm`, future runtime peers | **Zero known CVEs.** Releases are blocked if `npm audit --omit=dev` reports anything.   |
+| **Dev tooling**     | Vite, Vitest, ESLint, esbuild, etc.     | Fixed promptly via dependency bumps or `overrides`. Never reaches the published bundle. |
+
+### Reporting vulnerabilities
+
+If you find a vulnerability in `localm-web` itself (not in a transitive dep), open a private security advisory at <https://github.com/mauriciobenjamin700/localm-web/security/advisories/new>. Please don't open public issues for unpatched runtime vulns.
+
+### What we do on every release
+
+- `npm ci` (locked install — no drift between dev machine and CI).
+- `npm audit` reviewed manually; nothing handwaved.
+- ESM-only build, no eval / `Function()` / dynamic remote code.
+- Signed publish via `npm publish --provenance` (provenance attestation visible on the npm package page).
+
+### What you should do as a consumer
+
+- Pin the SDK version (`localm-web@x.y.z`, not `^x.y.z`) until you've validated a release.
+- Self-host model weights or use Subresource Integrity (SRI) when the runtime fetches them — model URLs are external.
+- Models are cached locally (Cache API + OPFS) — surface this in your app's privacy policy.
+- Run inference inside a Web Worker (the SDK does this by default from v0.2). Don't bypass it to "save a thread" — it isolates faulty model code from your UI.
+
+The full maintainer policy lives in [CLAUDE.md → Security & vulnerabilities](./CLAUDE.md#security--vulnerabilities).
+
+## Contributing
+
+Pre-alpha. Issues and design discussion welcome. PRs deferred until the v0.1 surface stabilizes.
+
+## License
+
+MIT — see [LICENSE](./LICENSE).
+
+## Related projects
+
+- [`ort-vision-sdk`](https://github.com/mauriciobenjamin700/ort-vision-sdk) — sibling SDK for computer vision (classification, detection, segmentation). Same DX patterns, same author.

package/dist/index.d.ts

@@ -0,0 +1,373 @@
+/**
+ * localm-web — browser-only TypeScript SDK for running LLMs and SLMs locally.
+ *
+ * Public API surface for v0.1.
+ *
+ * @packageDocumentation
+ */
+
+/** Thrown when no usable backend is available on the current platform. */
+export declare class BackendNotAvailableError extends LocalmWebError {
+}
+
+/**
+ * Multi-turn chat task.
+ *
+ * Maintains an in-memory conversation history and applies the chat template
+ * configured for the loaded model. Use {@link Chat.create} to construct an
+ * instance — the constructor is private.
+ *
+ * @example
+ * ```ts
+ * const chat = await Chat.create("phi-3.5-mini-int4");
+ * const reply = await chat.send("Explain ONNX in one sentence.");
+ * console.log(reply.text);
+ * ```
+ *
+ * @example Streaming
+ * ```ts
+ * const controller = new AbortController();
+ * for await (const token of chat.stream("Explain ONNX.", { signal: controller.signal })) {
+ *   process.stdout.write(token.text);
+ * }
+ * ```
+ */
+export declare class Chat extends LMTask {
+    private readonly history;
+    private systemPrompt;
+    private constructor();
+    /**
+     * Create and load a `Chat` task for the given model.
+     *
+     * @param modelId - Friendly model id from the registry (e.g. `"phi-3.5-mini-int4"`).
+     * @param options - Optional creation options (progress callback, engine override).
+     */
+    static create(modelId: string, options?: LMTaskCreateOptions): Promise<Chat>;
+    /** Set or replace the system prompt prepended to every conversation. */
+    setSystemPrompt(prompt: string): void;
+    /** Clear the system prompt. */
+    clearSystemPrompt(): void;
+    /** Reset the conversation history. The system prompt is preserved. */
+    resetHistory(): void;
+    /** A read-only snapshot of the conversation history. */
+    getHistory(): readonly Message[];
+    /**
+     * Send a user message and await the full assistant reply.
+     *
+     * The user message and the assistant reply are appended to the history.
+     *
+     * @param message - The user-facing message text.
+     * @param options - Generation options.
+     * @returns A {@link ChatReply} with the assistant's reply.
+     */
+    send(message: string, options?: GenerationOptions): Promise<ChatReply>;
+    /**
+     * Stream the assistant reply token-by-token as an async iterable.
+     *
+     * The full reply is appended to the history when the stream completes
+     * normally. If the stream is aborted, neither message is appended.
+     *
+     * @param message - The user-facing message text.
+     * @param options - Generation options including an optional `signal`.
+     */
+    stream(message: string, options?: GenerationOptions): AsyncIterable<TokenChunk>;
+    private buildMessages;
+}
+
+/**
+ * Result returned by `Chat.send()`.
+ *
+ * Holds the assistant's textual reply, the structured assistant message
+ * (already appended to the chat history), and metadata about the generation.
+ */
+export declare class ChatReply {
+    /** The assistant's reply text. */
+    readonly text: string;
+    /** The structured assistant message (already appended to chat history). */
+    readonly message: Message;
+    /** Number of tokens generated. 0 when the engine does not report it. */
+    readonly tokensGenerated: number;
+    /** Why the generation loop stopped. */
+    readonly finishReason: FinishReason;
+    constructor(
+    /** The assistant's reply text. */
+    text: string, 
+    /** The structured assistant message (already appended to chat history). */
+    message: Message, 
+    /** Number of tokens generated. 0 when the engine does not report it. */
+    tokensGenerated: number, 
+    /** Why the generation loop stopped. */
+    finishReason: FinishReason);
+}
+
+/**
+ * Drain an async iterable of token chunks into a single string.
+ *
+ * Useful in tests, for non-streaming consumers, and as a one-line way to
+ * reconstruct the final text from a `Chat.stream(...)` call.
+ *
+ * @param stream - The token-chunk async iterable to consume.
+ * @returns The concatenation of every chunk's `text` field.
+ */
+export declare function collectStream(stream: AsyncIterable<TokenChunk>): Promise<string>;
+
+/**
+ * Runtime-agnostic inference contract.
+ *
+ * Tasks (`Chat`, future `Completion`, etc.) depend on this interface, not on
+ * a concrete backend. This lets the SDK swap WebLLM for the ORT-Web fallback
+ * (planned for v0.5) without touching task code.
+ */
+export declare interface Engine {
+    /**
+     * Load a model into the engine.
+     *
+     * @param modelId - Backend-specific model identifier.
+     * @param onProgress - Optional callback for load progress updates.
+     * @throws ModelLoadError on failure to fetch or initialize the model.
+     * @throws WebGPUUnavailableError when the engine requires WebGPU but it is missing.
+     */
+    load(modelId: string, onProgress?: ProgressCallback): Promise<void>;
+    /**
+     * Generate a single non-streaming response.
+     *
+     * @param messages - Conversation history including the latest user turn.
+     * @param options - Generation options.
+     * @returns The full generated text.
+     * @throws ModelNotLoadedError if called before {@link Engine.load}.
+     * @throws GenerationAbortedError if `options.signal` is triggered.
+     */
+    generate(messages: Message[], options?: GenerationOptions): Promise<string>;
+    /**
+     * Generate a streaming response as an async iterable of token chunks.
+     *
+     * @param messages - Conversation history including the latest user turn.
+     * @param options - Generation options.
+     * @returns Async iterable yielding token chunks. The final chunk has `done: true`.
+     * @throws ModelNotLoadedError if called before {@link Engine.load}.
+     * @throws GenerationAbortedError if `options.signal` is triggered.
+     */
+    stream(messages: Message[], options?: GenerationOptions): AsyncIterable<TokenChunk>;
+    /** Release any resources held by the engine. Safe to call when not loaded. */
+    unload(): Promise<void>;
+    /** Whether a model is currently loaded and ready for inference. */
+    isLoaded(): boolean;
+}
+
+/** Reason a generation loop stopped. */
+export declare type FinishReason = "stop" | "length" | "abort";
+
+/** Thrown when generation is aborted via an `AbortSignal`. */
+export declare class GenerationAbortedError extends LocalmWebError {
+}
+
+/** Options that control a single generation call. */
+export declare interface GenerationOptions {
+    /** Maximum number of tokens to generate. Engine-default if omitted. */
+    maxTokens?: number;
+    /** Sampling temperature. 0 = deterministic, higher = more random. */
+    temperature?: number;
+    /** Top-K sampling cutoff. */
+    topK?: number;
+    /** Top-P (nucleus) sampling cutoff. */
+    topP?: number;
+    /** Cancellation signal. When triggered, the engine stops generation. */
+    signal?: AbortSignal;
+    /**
+     * JSON Schema for structured output. The engine constrains decoding to
+     * produce a string parseable as JSON matching the schema. Planned for v0.4.
+     */
+    jsonSchema?: object;
+}
+
+/** Return the list of supported friendly model ids. */
+export declare function listSupportedModels(): string[];
+
+/**
+ * Base class shared by all language-model tasks (`Chat` for v0.1; `Completion`,
+ * `Embeddings` and `Reranker` planned for later versions).
+ *
+ * The base owns:
+ *   - resolving a friendly model id to a {@link ModelPreset};
+ *   - selecting and loading an {@link Engine} (defaulting to WebLLM);
+ *   - exposing `unload()` for cleanup.
+ *
+ * Subclasses add task-specific public methods (`send`, `stream`, etc.).
+ */
+export declare abstract class LMTask {
+    /** Engine used for inference. */
+    protected readonly engine: Engine;
+    /** Resolved metadata for the loaded model. */
+    readonly preset: ModelPreset;
+    protected constructor(
+    /** Engine used for inference. */
+    engine: Engine, 
+    /** Resolved metadata for the loaded model. */
+    preset: ModelPreset);
+    /**
+     * Load a model into a backend and return the wired-up engine + preset.
+     *
+     * Subclasses call this from their static `create()` factories.
+     *
+     * @param modelId - Friendly model id from the registry.
+     * @param options - Task creation options.
+     */
+    protected static createEngine(modelId: string, options?: LMTaskCreateOptions): Promise<ResolvedEngine>;
+    /** Release engine resources. Safe to call multiple times. */
+    unload(): Promise<void>;
+    /** Whether the underlying engine has a loaded model. */
+    isLoaded(): boolean;
+}
+
+/** Common options accepted by every task's `create()` factory. */
+export declare interface LMTaskCreateOptions {
+    /** Optional callback for model load progress updates. */
+    onProgress?: ProgressCallback;
+    /**
+     * Override the engine used for inference. Intended for testing.
+     * Production callers should let the SDK pick a backend automatically.
+     */
+    engine?: Engine;
+}
+
+/**
+ * Error hierarchy for localm-web.
+ *
+ * All errors thrown by the SDK extend `LocalmWebError` so consumers can
+ * distinguish SDK errors from unrelated runtime errors with a single
+ * `instanceof` check.
+ */
+/** Base class for every error raised by localm-web. */
+export declare class LocalmWebError extends Error {
+    readonly cause?: unknown | undefined;
+    /**
+     * @param message - Human-readable description of the error.
+     * @param cause - Underlying error, if any.
+     */
+    constructor(message: string, cause?: unknown | undefined);
+}
+
+/** A single message in a chat conversation. */
+export declare interface Message {
+    /** The role of the speaker. */
+    role: Role;
+    /** Text content of the message. */
+    content: string;
+    /** Optional name (used by some chat templates and tool calls). */
+    name?: string;
+}
+
+/**
+ * Curated registry of supported models for v0.1.
+ *
+ * Each entry maps a friendly id (e.g. `"phi-3.5-mini-int4"`) to the underlying
+ * runtime identifier and metadata. Friendly ids are stable; backend ids may
+ * change as upstream MLC packages evolve.
+ *
+ * Only models that have been validated to load in browsers with WebGPU and
+ * that fit the SLM target (≤ 4B parameters at INT4) are included.
+ */
+export declare const MODEL_PRESETS: Readonly<Record<string, ModelPreset>>;
+
+/** Thrown when a model fails to load (network, parsing, runtime init). */
+export declare class ModelLoadError extends LocalmWebError {
+}
+
+/** Progress event emitted while a model is loading. */
+export declare interface ModelLoadProgress {
+    /** Fraction of total work completed, in [0, 1]. */
+    progress: number;
+    /** Human-readable status message from the underlying runtime. */
+    text: string;
+    /** Bytes loaded so far. 0 when unavailable. */
+    loaded: number;
+    /** Total bytes to load. 0 when unavailable. */
+    total: number;
+}
+
+/** Thrown when an inference call is made before a model has loaded. */
+export declare class ModelNotLoadedError extends LocalmWebError {
+}
+
+/** Curated metadata for a supported model. */
+export declare interface ModelPreset {
+    /** Friendly identifier exposed to users (e.g. "phi-3.5-mini-int4"). */
+    id: string;
+    /** Model family (e.g. "Phi-3.5", "Llama-3.2"). */
+    family: string;
+    /** Parameter count as a human string (e.g. "1B", "3.8B"). */
+    parameters: string;
+    /** Quantization scheme (e.g. "q4f16_1"). */
+    quantization: string;
+    /** Identifier expected by the WebLLM runtime. */
+    webllmId: string;
+    /** Optional ONNX URL used by the future ORT-Web fallback (v0.5+). */
+    ortUrl?: string;
+    /** Maximum context window in tokens. */
+    contextWindow: number;
+    /** Short human description. */
+    description: string;
+}
+
+/** Callback signature for model load progress. */
+export declare type ProgressCallback = (progress: ModelLoadProgress) => void;
+
+/** Thrown when the browser denies storage quota for the model cache. */
+export declare class QuotaExceededError extends LocalmWebError {
+}
+
+/** Internal payload returned by {@link LMTask.createEngine}. */
+declare interface ResolvedEngine {
+    engine: Engine;
+    preset: ModelPreset;
+}
+
+/**
+ * Resolve a friendly model id to its full preset metadata.
+ *
+ * @param modelId - Friendly id (e.g. `"phi-3.5-mini-int4"`).
+ * @returns The matching preset.
+ * @throws UnknownModelError if no preset matches.
+ */
+export declare function resolveModelPreset(modelId: string): ModelPreset;
+
+/**
+ * Public type primitives for localm-web.
+ */
+/** Conversation roles supported by chat templates. */
+export declare type Role = "system" | "user" | "assistant" | "tool";
+
+/**
+ * Wrap an async iterable so that each `TokenChunk` is also passed to a
+ * caller-supplied side-effect callback before being yielded downstream.
+ *
+ * This is intentionally a passthrough — it does not buffer.
+ *
+ * @param stream - The upstream token-chunk async iterable.
+ * @param onChunk - Side-effect invoked for every chunk.
+ * @returns A new async iterable yielding the same chunks.
+ */
+export declare function tap(stream: AsyncIterable<TokenChunk>, onChunk: (chunk: TokenChunk) => void): AsyncIterable<TokenChunk>;
+
+/** A single token (or short span) produced by the streaming generator. */
+export declare interface TokenChunk {
+    /** Text fragment produced in this step. */
+    text: string;
+    /** Sequential index of this chunk in the stream, starting at 0. */
+    index: number;
+    /** True for the final chunk; the final chunk has empty text. */
+    done: boolean;
+}
+
+/** Thrown when a model id is not present in the curated registry. */
+export declare class UnknownModelError extends LocalmWebError {
+}
+
+/** Current package version. Updated at release time. */
+export declare const VERSION: string;
+
+/** Thrown when WebGPU is required but not available in the host browser. */
+export declare class WebGPUUnavailableError extends LocalmWebError {
+}
+
+export { }

package/dist/index.js

@@ -0,0 +1,336 @@
+class LocalmWebError extends Error {
+  /**
+   * @param message - Human-readable description of the error.
+   * @param cause - Underlying error, if any.
+   */
+  constructor(message, cause) {
+    super(message);
+    this.cause = cause;
+    this.name = new.target.name;
+  }
+}
+class WebGPUUnavailableError extends LocalmWebError {
+}
+class ModelLoadError extends LocalmWebError {
+}
+class ModelNotLoadedError extends LocalmWebError {
+}
+class UnknownModelError extends LocalmWebError {
+}
+class GenerationAbortedError extends LocalmWebError {
+}
+class QuotaExceededError extends LocalmWebError {
+}
+class BackendNotAvailableError extends LocalmWebError {
+}
+let webllmModulePromise = null;
+async function loadWebLLM() {
+  if (!webllmModulePromise) {
+    webllmModulePromise = import("@mlc-ai/web-llm");
+  }
+  return webllmModulePromise;
+}
+function isWebGPUAvailable() {
+  return typeof navigator !== "undefined" && "gpu" in navigator;
+}
+function buildSamplingParams(options) {
+  const params = {};
+  if (options.maxTokens !== void 0) params.max_tokens = options.maxTokens;
+  if (options.temperature !== void 0) params.temperature = options.temperature;
+  if (options.topP !== void 0) params.top_p = options.topP;
+  return params;
+}
+function toChatMessages(messages) {
+  return messages.map((m) => {
+    switch (m.role) {
+      case "system":
+        return { role: "system", content: m.content };
+      case "user":
+        return { role: "user", content: m.content };
+      case "assistant":
+        return { role: "assistant", content: m.content };
+      case "tool":
+        return { role: "tool", content: m.content, tool_call_id: m.name ?? "" };
+    }
+  });
+}
+class WebLLMEngine {
+  engine = null;
+  isLoaded() {
+    return this.engine !== null;
+  }
+  async load(modelId, onProgress) {
+    if (!isWebGPUAvailable()) {
+      throw new WebGPUUnavailableError(
+        "WebGPU is not available in this browser. The ORT-Web fallback is planned for v0.5."
+      );
+    }
+    const webllm = await loadWebLLM();
+    try {
+      this.engine = await webllm.CreateMLCEngine(modelId, {
+        initProgressCallback: (report) => {
+          onProgress?.({
+            progress: report.progress,
+            text: report.text,
+            loaded: 0,
+            total: 0
+          });
+        }
+      });
+    } catch (err) {
+      throw new ModelLoadError(`Failed to load model "${modelId}".`, err);
+    }
+  }
+  async generate(messages, options = {}) {
+    const engine = this.requireEngine();
+    if (options.signal?.aborted) {
+      throw new GenerationAbortedError("Generation aborted before start.");
+    }
+    const completion = await engine.chat.completions.create({
+      ...buildSamplingParams(options),
+      messages: toChatMessages(messages),
+      stream: false
+    });
+    return completion.choices[0]?.message?.content ?? "";
+  }
+  async *stream(messages, options = {}) {
+    const engine = this.requireEngine();
+    if (options.signal?.aborted) {
+      throw new GenerationAbortedError("Generation aborted before start.");
+    }
+    const completion = await engine.chat.completions.create({
+      ...buildSamplingParams(options),
+      messages: toChatMessages(messages),
+      stream: true
+    });
+    let index = 0;
+    let finished = false;
+    try {
+      for await (const chunk of completion) {
+        if (options.signal?.aborted) {
+          throw new GenerationAbortedError("Generation aborted by signal.");
+        }
+        const choice = chunk.choices[0];
+        const delta = choice?.delta?.content ?? "";
+        if (delta) {
+          yield { text: delta, index, done: false };
+          index += 1;
+        }
+        if (choice?.finish_reason) {
+          finished = true;
+          yield { text: "", index, done: true };
+          index += 1;
+        }
+      }
+      if (!finished) {
+        yield { text: "", index, done: true };
+      }
+    } catch (err) {
+      if (err instanceof GenerationAbortedError) throw err;
+      throw new ModelLoadError("Streaming generation failed.", err);
+    }
+  }
+  async unload() {
+    if (this.engine) {
+      await this.engine.unload();
+      this.engine = null;
+    }
+  }
+  requireEngine() {
+    if (!this.engine) {
+      throw new ModelNotLoadedError("Engine not loaded. Call load() before generation.");
+    }
+    return this.engine;
+  }
+}
+const MODEL_PRESETS = Object.freeze({
+  "phi-3.5-mini-int4": {
+    id: "phi-3.5-mini-int4",
+    family: "Phi-3.5",
+    parameters: "3.8B",
+    quantization: "q4f16_1",
+    webllmId: "Phi-3.5-mini-instruct-q4f16_1-MLC",
+    contextWindow: 4096,
+    description: "Microsoft Phi-3.5 mini, INT4 quantized for browser inference."
+  },
+  "llama-3.2-1b-int4": {
+    id: "llama-3.2-1b-int4",
+    family: "Llama-3.2",
+    parameters: "1B",
+    quantization: "q4f16_1",
+    webllmId: "Llama-3.2-1B-Instruct-q4f16_1-MLC",
+    contextWindow: 4096,
+    description: "Meta Llama 3.2 1B Instruct, INT4 quantized."
+  },
+  "qwen2.5-1.5b-int4": {
+    id: "qwen2.5-1.5b-int4",
+    family: "Qwen2.5",
+    parameters: "1.5B",
+    quantization: "q4f16_1",
+    webllmId: "Qwen2.5-1.5B-Instruct-q4f16_1-MLC",
+    contextWindow: 4096,
+    description: "Alibaba Qwen 2.5 1.5B Instruct, INT4 quantized."
+  }
+});
+function resolveModelPreset(modelId) {
+  const preset = MODEL_PRESETS[modelId];
+  if (!preset) {
+    const available = Object.keys(MODEL_PRESETS).join(", ");
+    throw new UnknownModelError(`Unknown model "${modelId}". Available models: ${available}.`);
+  }
+  return preset;
+}
+function listSupportedModels() {
+  return Object.keys(MODEL_PRESETS);
+}
+class LMTask {
+  constructor(engine, preset) {
+    this.engine = engine;
+    this.preset = preset;
+  }
+  /**
+   * Load a model into a backend and return the wired-up engine + preset.
+   *
+   * Subclasses call this from their static `create()` factories.
+   *
+   * @param modelId - Friendly model id from the registry.
+   * @param options - Task creation options.
+   */
+  static async createEngine(modelId, options = {}) {
+    const preset = resolveModelPreset(modelId);
+    const engine = options.engine ?? new WebLLMEngine();
+    if (!engine.isLoaded()) {
+      await engine.load(preset.webllmId, options.onProgress);
+    }
+    return { engine, preset };
+  }
+  /** Release engine resources. Safe to call multiple times. */
+  async unload() {
+    await this.engine.unload();
+  }
+  /** Whether the underlying engine has a loaded model. */
+  isLoaded() {
+    return this.engine.isLoaded();
+  }
+}
+class ChatReply {
+  constructor(text, message, tokensGenerated, finishReason) {
+    this.text = text;
+    this.message = message;
+    this.tokensGenerated = tokensGenerated;
+    this.finishReason = finishReason;
+  }
+}
+class Chat extends LMTask {
+  history = [];
+  systemPrompt = null;
+  constructor(engine, preset) {
+    super(engine, preset);
+  }
+  /**
+   * Create and load a `Chat` task for the given model.
+   *
+   * @param modelId - Friendly model id from the registry (e.g. `"phi-3.5-mini-int4"`).
+   * @param options - Optional creation options (progress callback, engine override).
+   */
+  static async create(modelId, options = {}) {
+    const { engine, preset } = await LMTask.createEngine(modelId, options);
+    return new Chat(engine, preset);
+  }
+  /** Set or replace the system prompt prepended to every conversation. */
+  setSystemPrompt(prompt) {
+    this.systemPrompt = prompt;
+  }
+  /** Clear the system prompt. */
+  clearSystemPrompt() {
+    this.systemPrompt = null;
+  }
+  /** Reset the conversation history. The system prompt is preserved. */
+  resetHistory() {
+    this.history.length = 0;
+  }
+  /** A read-only snapshot of the conversation history. */
+  getHistory() {
+    return this.history.slice();
+  }
+  /**
+   * Send a user message and await the full assistant reply.
+   *
+   * The user message and the assistant reply are appended to the history.
+   *
+   * @param message - The user-facing message text.
+   * @param options - Generation options.
+   * @returns A {@link ChatReply} with the assistant's reply.
+   */
+  async send(message, options = {}) {
+    const messages = this.buildMessages(message);
+    const text = await this.engine.generate(messages, options);
+    const userMsg = { role: "user", content: message };
+    const assistantMsg = { role: "assistant", content: text };
+    this.history.push(userMsg, assistantMsg);
+    return new ChatReply(text, assistantMsg, 0, "stop");
+  }
+  /**
+   * Stream the assistant reply token-by-token as an async iterable.
+   *
+   * The full reply is appended to the history when the stream completes
+   * normally. If the stream is aborted, neither message is appended.
+   *
+   * @param message - The user-facing message text.
+   * @param options - Generation options including an optional `signal`.
+   */
+  async *stream(message, options = {}) {
+    const messages = this.buildMessages(message);
+    const userMsg = { role: "user", content: message };
+    let acc = "";
+    for await (const chunk of this.engine.stream(messages, options)) {
+      acc += chunk.text;
+      yield chunk;
+    }
+    const assistantMsg = { role: "assistant", content: acc };
+    this.history.push(userMsg, assistantMsg);
+  }
+  buildMessages(userMessage) {
+    const messages = [];
+    if (this.systemPrompt) {
+      messages.push({ role: "system", content: this.systemPrompt });
+    }
+    messages.push(...this.history);
+    messages.push({ role: "user", content: userMessage });
+    return messages;
+  }
+}
+async function collectStream(stream) {
+  let acc = "";
+  for await (const chunk of stream) {
+    acc += chunk.text;
+  }
+  return acc;
+}
+async function* tap(stream, onChunk) {
+  for await (const chunk of stream) {
+    onChunk(chunk);
+    yield chunk;
+  }
+}
+const VERSION = "0.1.0";
+export {
+  BackendNotAvailableError,
+  Chat,
+  ChatReply,
+  GenerationAbortedError,
+  LMTask,
+  LocalmWebError,
+  MODEL_PRESETS,
+  ModelLoadError,
+  ModelNotLoadedError,
+  QuotaExceededError,
+  UnknownModelError,
+  VERSION,
+  WebGPUUnavailableError,
+  collectStream,
+  listSupportedModels,
+  resolveModelPreset,
+  tap
+};
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sources":["../src/core/exceptions.ts","../src/core/webllm-engine.ts","../src/presets/models.ts","../src/tasks/lm-task.ts","../src/results.ts","../src/tasks/chat.ts","../src/streaming/token-stream.ts","../src/index.ts"],"sourcesContent":["/**\n * Error hierarchy for localm-web.\n *\n * All errors thrown by the SDK extend `LocalmWebError` so consumers can\n * distinguish SDK errors from unrelated runtime errors with a single\n * `instanceof` check.\n */\n\n/** Base class for every error raised by localm-web. */\nexport class LocalmWebError extends Error {\n  /**\n   * @param message - Human-readable description of the error.\n   * @param cause - Underlying error, if any.\n   */\n  constructor(\n    message: string,\n    public readonly cause?: unknown\n  ) {\n    super(message);\n    this.name = new.target.name;\n  }\n}\n\n/** Thrown when WebGPU is required but not available in the host browser. */\nexport class WebGPUUnavailableError extends LocalmWebError {}\n\n/** Thrown when a model fails to load (network, parsing, runtime init). */\nexport class ModelLoadError extends LocalmWebError {}\n\n/** Thrown when an inference call is made before a model has loaded. */\nexport class ModelNotLoadedError extends LocalmWebError {}\n\n/** Thrown when a model id is not present in the curated registry. */\nexport class UnknownModelError extends LocalmWebError {}\n\n/** Thrown when generation is aborted via an `AbortSignal`. */\nexport class GenerationAbortedError extends LocalmWebError {}\n\n/** Thrown when the browser denies storage quota for the model cache. */\nexport class QuotaExceededError extends LocalmWebError {}\n\n/** Thrown when no usable backend is available on the current platform. */\nexport class BackendNotAvailableError extends LocalmWebError {}\n","import type { Engine } from \"./engine\";\nimport type { GenerationOptions, Message, ProgressCallback, TokenChunk } from \"../types\";\nimport {\n  GenerationAbortedError,\n  ModelLoadError,\n  ModelNotLoadedError,\n  WebGPUUnavailableError,\n} from \"./exceptions\";\n\ntype WebLLMModule = typeof import(\"@mlc-ai/web-llm\");\ntype MLCEngine = import(\"@mlc-ai/web-llm\").MLCEngineInterface;\ntype ChatCompletionMessageParam = import(\"@mlc-ai/web-llm\").ChatCompletionMessageParam;\n\nlet webllmModulePromise: Promise<WebLLMModule> | null = null;\n\nasync function loadWebLLM(): Promise<WebLLMModule> {\n  if (!webllmModulePromise) {\n    webllmModulePromise = import(\"@mlc-ai/web-llm\");\n  }\n  return webllmModulePromise;\n}\n\nfunction isWebGPUAvailable(): boolean {\n  return typeof navigator !== \"undefined\" && \"gpu\" in navigator;\n}\n\ninterface SamplingParams {\n  max_tokens?: number;\n  temperature?: number;\n  top_p?: number;\n}\n\nfunction buildSamplingParams(options: GenerationOptions): SamplingParams {\n  const params: SamplingParams = {};\n  if (options.maxTokens !== undefined) params.max_tokens = options.maxTokens;\n  if (options.temperature !== undefined) params.temperature = options.temperature;\n  if (options.topP !== undefined) params.top_p = options.topP;\n  return params;\n}\n\nfunction toChatMessages(messages: Message[]): ChatCompletionMessageParam[] {\n  return messages.map((m): ChatCompletionMessageParam => {\n    switch (m.role) {\n      case \"system\":\n        return { role: \"system\", content: m.content };\n      case \"user\":\n        return { role: \"user\", content: m.content };\n      case \"assistant\":\n        return { role: \"assistant\", content: m.content };\n      case \"tool\":\n        return { role: \"tool\", content: m.content, tool_call_id: m.name ?? \"\" };\n    }\n  });\n}\n\n/**\n * Inference engine backed by [WebLLM (MLC)](https://github.com/mlc-ai/web-llm).\n *\n * Requires WebGPU. The fallback path planned for v0.5 will route to ORT-Web\n * when WebGPU is missing.\n */\nexport class WebLLMEngine implements Engine {\n  private engine: MLCEngine | null = null;\n\n  isLoaded(): boolean {\n    return this.engine !== null;\n  }\n\n  async load(modelId: string, onProgress?: ProgressCallback): Promise<void> {\n    if (!isWebGPUAvailable()) {\n      throw new WebGPUUnavailableError(\n        \"WebGPU is not available in this browser. The ORT-Web fallback is planned for v0.5.\"\n      );\n    }\n    const webllm = await loadWebLLM();\n    try {\n      this.engine = await webllm.CreateMLCEngine(modelId, {\n        initProgressCallback: (report): void => {\n          onProgress?.({\n            progress: report.progress,\n            text: report.text,\n            loaded: 0,\n            total: 0,\n          });\n        },\n      });\n    } catch (err) {\n      throw new ModelLoadError(`Failed to load model \"${modelId}\".`, err);\n    }\n  }\n\n  async generate(messages: Message[], options: GenerationOptions = {}): Promise<string> {\n    const engine = this.requireEngine();\n    if (options.signal?.aborted) {\n      throw new GenerationAbortedError(\"Generation aborted before start.\");\n    }\n    const completion = await engine.chat.completions.create({\n      ...buildSamplingParams(options),\n      messages: toChatMessages(messages),\n      stream: false,\n    });\n    return completion.choices[0]?.message?.content ?? \"\";\n  }\n\n  async *stream(messages: Message[], options: GenerationOptions = {}): AsyncIterable<TokenChunk> {\n    const engine = this.requireEngine();\n    if (options.signal?.aborted) {\n      throw new GenerationAbortedError(\"Generation aborted before start.\");\n    }\n    const completion = await engine.chat.completions.create({\n      ...buildSamplingParams(options),\n      messages: toChatMessages(messages),\n      stream: true,\n    });\n    let index: number = 0;\n    let finished: boolean = false;\n    try {\n      for await (const chunk of completion) {\n        if (options.signal?.aborted) {\n          throw new GenerationAbortedError(\"Generation aborted by signal.\");\n        }\n        const choice = chunk.choices[0];\n        const delta = choice?.delta?.content ?? \"\";\n        if (delta) {\n          yield { text: delta, index, done: false };\n          index += 1;\n        }\n        if (choice?.finish_reason) {\n          finished = true;\n          yield { text: \"\", index, done: true };\n          index += 1;\n        }\n      }\n      if (!finished) {\n        yield { text: \"\", index, done: true };\n      }\n    } catch (err) {\n      if (err instanceof GenerationAbortedError) throw err;\n      throw new ModelLoadError(\"Streaming generation failed.\", err);\n    }\n  }\n\n  async unload(): Promise<void> {\n    if (this.engine) {\n      await this.engine.unload();\n      this.engine = null;\n    }\n  }\n\n  private requireEngine(): MLCEngine {\n    if (!this.engine) {\n      throw new ModelNotLoadedError(\"Engine not loaded. Call load() before generation.\");\n    }\n    return this.engine;\n  }\n}\n","import type { ModelPreset } from \"../types\";\nimport { UnknownModelError } from \"../core/exceptions\";\n\n/**\n * Curated registry of supported models for v0.1.\n *\n * Each entry maps a friendly id (e.g. `\"phi-3.5-mini-int4\"`) to the underlying\n * runtime identifier and metadata. Friendly ids are stable; backend ids may\n * change as upstream MLC packages evolve.\n *\n * Only models that have been validated to load in browsers with WebGPU and\n * that fit the SLM target (≤ 4B parameters at INT4) are included.\n */\nexport const MODEL_PRESETS: Readonly<Record<string, ModelPreset>> = Object.freeze({\n  \"phi-3.5-mini-int4\": {\n    id: \"phi-3.5-mini-int4\",\n    family: \"Phi-3.5\",\n    parameters: \"3.8B\",\n    quantization: \"q4f16_1\",\n    webllmId: \"Phi-3.5-mini-instruct-q4f16_1-MLC\",\n    contextWindow: 4096,\n    description: \"Microsoft Phi-3.5 mini, INT4 quantized for browser inference.\",\n  },\n  \"llama-3.2-1b-int4\": {\n    id: \"llama-3.2-1b-int4\",\n    family: \"Llama-3.2\",\n    parameters: \"1B\",\n    quantization: \"q4f16_1\",\n    webllmId: \"Llama-3.2-1B-Instruct-q4f16_1-MLC\",\n    contextWindow: 4096,\n    description: \"Meta Llama 3.2 1B Instruct, INT4 quantized.\",\n  },\n  \"qwen2.5-1.5b-int4\": {\n    id: \"qwen2.5-1.5b-int4\",\n    family: \"Qwen2.5\",\n    parameters: \"1.5B\",\n    quantization: \"q4f16_1\",\n    webllmId: \"Qwen2.5-1.5B-Instruct-q4f16_1-MLC\",\n    contextWindow: 4096,\n    description: \"Alibaba Qwen 2.5 1.5B Instruct, INT4 quantized.\",\n  },\n});\n\n/**\n * Resolve a friendly model id to its full preset metadata.\n *\n * @param modelId - Friendly id (e.g. `\"phi-3.5-mini-int4\"`).\n * @returns The matching preset.\n * @throws UnknownModelError if no preset matches.\n */\nexport function resolveModelPreset(modelId: string): ModelPreset {\n  const preset = MODEL_PRESETS[modelId];\n  if (!preset) {\n    const available = Object.keys(MODEL_PRESETS).join(\", \");\n    throw new UnknownModelError(`Unknown model \"${modelId}\". Available models: ${available}.`);\n  }\n  return preset;\n}\n\n/** Return the list of supported friendly model ids. */\nexport function listSupportedModels(): string[] {\n  return Object.keys(MODEL_PRESETS);\n}\n","import type { Engine } from \"../core/engine\";\nimport { WebLLMEngine } from \"../core/webllm-engine\";\nimport { resolveModelPreset } from \"../presets/models\";\nimport type { ModelPreset, ProgressCallback } from \"../types\";\n\n/** Common options accepted by every task's `create()` factory. */\nexport interface LMTaskCreateOptions {\n  /** Optional callback for model load progress updates. */\n  onProgress?: ProgressCallback;\n  /**\n   * Override the engine used for inference. Intended for testing.\n   * Production callers should let the SDK pick a backend automatically.\n   */\n  engine?: Engine;\n}\n\n/** Internal payload returned by {@link LMTask.createEngine}. */\nexport interface ResolvedEngine {\n  engine: Engine;\n  preset: ModelPreset;\n}\n\n/**\n * Base class shared by all language-model tasks (`Chat` for v0.1; `Completion`,\n * `Embeddings` and `Reranker` planned for later versions).\n *\n * The base owns:\n *   - resolving a friendly model id to a {@link ModelPreset};\n *   - selecting and loading an {@link Engine} (defaulting to WebLLM);\n *   - exposing `unload()` for cleanup.\n *\n * Subclasses add task-specific public methods (`send`, `stream`, etc.).\n */\nexport abstract class LMTask {\n  protected constructor(\n    /** Engine used for inference. */\n    protected readonly engine: Engine,\n    /** Resolved metadata for the loaded model. */\n    public readonly preset: ModelPreset\n  ) {}\n\n  /**\n   * Load a model into a backend and return the wired-up engine + preset.\n   *\n   * Subclasses call this from their static `create()` factories.\n   *\n   * @param modelId - Friendly model id from the registry.\n   * @param options - Task creation options.\n   */\n  protected static async createEngine(\n    modelId: string,\n    options: LMTaskCreateOptions = {}\n  ): Promise<ResolvedEngine> {\n    const preset = resolveModelPreset(modelId);\n    const engine = options.engine ?? new WebLLMEngine();\n    if (!engine.isLoaded()) {\n      await engine.load(preset.webllmId, options.onProgress);\n    }\n    return { engine, preset };\n  }\n\n  /** Release engine resources. Safe to call multiple times. */\n  async unload(): Promise<void> {\n    await this.engine.unload();\n  }\n\n  /** Whether the underlying engine has a loaded model. */\n  isLoaded(): boolean {\n    return this.engine.isLoaded();\n  }\n}\n","import type { FinishReason, Message } from \"./types\";\n\n/**\n * Result returned by `Chat.send()`.\n *\n * Holds the assistant's textual reply, the structured assistant message\n * (already appended to the chat history), and metadata about the generation.\n */\nexport class ChatReply {\n  constructor(\n    /** The assistant's reply text. */\n    public readonly text: string,\n    /** The structured assistant message (already appended to chat history). */\n    public readonly message: Message,\n    /** Number of tokens generated. 0 when the engine does not report it. */\n    public readonly tokensGenerated: number,\n    /** Why the generation loop stopped. */\n    public readonly finishReason: FinishReason\n  ) {}\n}\n","import { LMTask, type LMTaskCreateOptions } from \"./lm-task\";\nimport type { Engine } from \"../core/engine\";\nimport { ChatReply } from \"../results\";\nimport type { GenerationOptions, Message, ModelPreset, TokenChunk } from \"../types\";\n\n/**\n * Multi-turn chat task.\n *\n * Maintains an in-memory conversation history and applies the chat template\n * configured for the loaded model. Use {@link Chat.create} to construct an\n * instance — the constructor is private.\n *\n * @example\n * ```ts\n * const chat = await Chat.create(\"phi-3.5-mini-int4\");\n * const reply = await chat.send(\"Explain ONNX in one sentence.\");\n * console.log(reply.text);\n * ```\n *\n * @example Streaming\n * ```ts\n * const controller = new AbortController();\n * for await (const token of chat.stream(\"Explain ONNX.\", { signal: controller.signal })) {\n *   process.stdout.write(token.text);\n * }\n * ```\n */\nexport class Chat extends LMTask {\n  private readonly history: Message[] = [];\n  private systemPrompt: string | null = null;\n\n  private constructor(engine: Engine, preset: ModelPreset) {\n    super(engine, preset);\n  }\n\n  /**\n   * Create and load a `Chat` task for the given model.\n   *\n   * @param modelId - Friendly model id from the registry (e.g. `\"phi-3.5-mini-int4\"`).\n   * @param options - Optional creation options (progress callback, engine override).\n   */\n  static async create(modelId: string, options: LMTaskCreateOptions = {}): Promise<Chat> {\n    const { engine, preset } = await LMTask.createEngine(modelId, options);\n    return new Chat(engine, preset);\n  }\n\n  /** Set or replace the system prompt prepended to every conversation. */\n  setSystemPrompt(prompt: string): void {\n    this.systemPrompt = prompt;\n  }\n\n  /** Clear the system prompt. */\n  clearSystemPrompt(): void {\n    this.systemPrompt = null;\n  }\n\n  /** Reset the conversation history. The system prompt is preserved. */\n  resetHistory(): void {\n    this.history.length = 0;\n  }\n\n  /** A read-only snapshot of the conversation history. */\n  getHistory(): readonly Message[] {\n    return this.history.slice();\n  }\n\n  /**\n   * Send a user message and await the full assistant reply.\n   *\n   * The user message and the assistant reply are appended to the history.\n   *\n   * @param message - The user-facing message text.\n   * @param options - Generation options.\n   * @returns A {@link ChatReply} with the assistant's reply.\n   */\n  async send(message: string, options: GenerationOptions = {}): Promise<ChatReply> {\n    const messages = this.buildMessages(message);\n    const text = await this.engine.generate(messages, options);\n    const userMsg: Message = { role: \"user\", content: message };\n    const assistantMsg: Message = { role: \"assistant\", content: text };\n    this.history.push(userMsg, assistantMsg);\n    return new ChatReply(text, assistantMsg, 0, \"stop\");\n  }\n\n  /**\n   * Stream the assistant reply token-by-token as an async iterable.\n   *\n   * The full reply is appended to the history when the stream completes\n   * normally. If the stream is aborted, neither message is appended.\n   *\n   * @param message - The user-facing message text.\n   * @param options - Generation options including an optional `signal`.\n   */\n  async *stream(message: string, options: GenerationOptions = {}): AsyncIterable<TokenChunk> {\n    const messages = this.buildMessages(message);\n    const userMsg: Message = { role: \"user\", content: message };\n    let acc: string = \"\";\n    for await (const chunk of this.engine.stream(messages, options)) {\n      acc += chunk.text;\n      yield chunk;\n    }\n    const assistantMsg: Message = { role: \"assistant\", content: acc };\n    this.history.push(userMsg, assistantMsg);\n  }\n\n  private buildMessages(userMessage: string): Message[] {\n    const messages: Message[] = [];\n    if (this.systemPrompt) {\n      messages.push({ role: \"system\", content: this.systemPrompt });\n    }\n    messages.push(...this.history);\n    messages.push({ role: \"user\", content: userMessage });\n    return messages;\n  }\n}\n","import type { TokenChunk } from \"../types\";\n\n/**\n * Drain an async iterable of token chunks into a single string.\n *\n * Useful in tests, for non-streaming consumers, and as a one-line way to\n * reconstruct the final text from a `Chat.stream(...)` call.\n *\n * @param stream - The token-chunk async iterable to consume.\n * @returns The concatenation of every chunk's `text` field.\n */\nexport async function collectStream(stream: AsyncIterable<TokenChunk>): Promise<string> {\n  let acc: string = \"\";\n  for await (const chunk of stream) {\n    acc += chunk.text;\n  }\n  return acc;\n}\n\n/**\n * Wrap an async iterable so that each `TokenChunk` is also passed to a\n * caller-supplied side-effect callback before being yielded downstream.\n *\n * This is intentionally a passthrough — it does not buffer.\n *\n * @param stream - The upstream token-chunk async iterable.\n * @param onChunk - Side-effect invoked for every chunk.\n * @returns A new async iterable yielding the same chunks.\n */\nexport async function* tap(\n  stream: AsyncIterable<TokenChunk>,\n  onChunk: (chunk: TokenChunk) => void\n): AsyncIterable<TokenChunk> {\n  for await (const chunk of stream) {\n    onChunk(chunk);\n    yield chunk;\n  }\n}\n","/**\n * localm-web — browser-only TypeScript SDK for running LLMs and SLMs locally.\n *\n * Public API surface for v0.1.\n *\n * @packageDocumentation\n */\n\nexport { Chat } from \"./tasks/chat\";\nexport { LMTask } from \"./tasks/lm-task\";\nexport type { LMTaskCreateOptions } from \"./tasks/lm-task\";\n\nexport { ChatReply } from \"./results\";\n\nexport { MODEL_PRESETS, resolveModelPreset, listSupportedModels } from \"./presets/models\";\n\nexport {\n  LocalmWebError,\n  WebGPUUnavailableError,\n  ModelLoadError,\n  ModelNotLoadedError,\n  UnknownModelError,\n  GenerationAbortedError,\n  QuotaExceededError,\n  BackendNotAvailableError,\n} from \"./core/exceptions\";\n\nexport type { Engine } from \"./core/engine\";\n\nexport { collectStream, tap } from \"./streaming/token-stream\";\n\nexport type {\n  Role,\n  FinishReason,\n  Message,\n  GenerationOptions,\n  ModelLoadProgress,\n  ProgressCallback,\n  TokenChunk,\n  ModelPreset,\n} from \"./types\";\n\n/** Current package version. Updated at release time. */\nexport const VERSION: string = \"0.1.0\";\n"],"names":[],"mappings":"AASO,MAAM,uBAAuB,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA,EAKxC,YACE,SACgB,OAChB;AACA,UAAM,OAAO;AAFG,SAAA,QAAA;AAGhB,SAAK,OAAO,WAAW;AAAA,EACzB;AACF;AAGO,MAAM,+BAA+B,eAAe;AAAC;AAGrD,MAAM,uBAAuB,eAAe;AAAC;AAG7C,MAAM,4BAA4B,eAAe;AAAC;AAGlD,MAAM,0BAA0B,eAAe;AAAC;AAGhD,MAAM,+BAA+B,eAAe;AAAC;AAGrD,MAAM,2BAA2B,eAAe;AAAC;AAGjD,MAAM,iCAAiC,eAAe;AAAC;AC7B9D,IAAI,sBAAoD;AAExD,eAAe,aAAoC;AACjD,MAAI,CAAC,qBAAqB;AACxB,0BAAsB,OAAO,iBAAiB;AAAA,EAChD;AACA,SAAO;AACT;AAEA,SAAS,oBAA6B;AACpC,SAAO,OAAO,cAAc,eAAe,SAAS;AACtD;AAQA,SAAS,oBAAoB,SAA4C;AACvE,QAAM,SAAyB,CAAA;AAC/B,MAAI,QAAQ,cAAc,OAAW,QAAO,aAAa,QAAQ;AACjE,MAAI,QAAQ,gBAAgB,OAAW,QAAO,cAAc,QAAQ;AACpE,MAAI,QAAQ,SAAS,OAAW,QAAO,QAAQ,QAAQ;AACvD,SAAO;AACT;AAEA,SAAS,eAAe,UAAmD;AACzE,SAAO,SAAS,IAAI,CAAC,MAAkC;AACrD,YAAQ,EAAE,MAAA;AAAA,MACR,KAAK;AACH,eAAO,EAAE,MAAM,UAAU,SAAS,EAAE,QAAA;AAAA,MACtC,KAAK;AACH,eAAO,EAAE,MAAM,QAAQ,SAAS,EAAE,QAAA;AAAA,MACpC,KAAK;AACH,eAAO,EAAE,MAAM,aAAa,SAAS,EAAE,QAAA;AAAA,MACzC,KAAK;AACH,eAAO,EAAE,MAAM,QAAQ,SAAS,EAAE,SAAS,cAAc,EAAE,QAAQ,GAAA;AAAA,IAAG;AAAA,EAE5E,CAAC;AACH;AAQO,MAAM,aAA+B;AAAA,EAClC,SAA2B;AAAA,EAEnC,WAAoB;AAClB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA,EAEA,MAAM,KAAK,SAAiB,YAA8C;AACxE,QAAI,CAAC,qBAAqB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,MAAA;AAAA,IAEJ;AACA,UAAM,SAAS,MAAM,WAAA;AACrB,QAAI;AACF,WAAK,SAAS,MAAM,OAAO,gBAAgB,SAAS;AAAA,QAClD,sBAAsB,CAAC,WAAiB;AACtC,uBAAa;AAAA,YACX,UAAU,OAAO;AAAA,YACjB,MAAM,OAAO;AAAA,YACb,QAAQ;AAAA,YACR,OAAO;AAAA,UAAA,CACR;AAAA,QACH;AAAA,MAAA,CACD;AAAA,IACH,SAAS,KAAK;AACZ,YAAM,IAAI,eAAe,yBAAyB,OAAO,MAAM,GAAG;AAAA,IACpE;AAAA,EACF;AAAA,EAEA,MAAM,SAAS,UAAqB,UAA6B,IAAqB;AACpF,UAAM,SAAS,KAAK,cAAA;AACpB,QAAI,QAAQ,QAAQ,SAAS;AAC3B,YAAM,IAAI,uBAAuB,kCAAkC;AAAA,IACrE;AACA,UAAM,aAAa,MAAM,OAAO,KAAK,YAAY,OAAO;AAAA,MACtD,GAAG,oBAAoB,OAAO;AAAA,MAC9B,UAAU,eAAe,QAAQ;AAAA,MACjC,QAAQ;AAAA,IAAA,CACT;AACD,WAAO,WAAW,QAAQ,CAAC,GAAG,SAAS,WAAW;AAAA,EACpD;AAAA,EAEA,OAAO,OAAO,UAAqB,UAA6B,IAA+B;AAC7F,UAAM,SAAS,KAAK,cAAA;AACpB,QAAI,QAAQ,QAAQ,SAAS;AAC3B,YAAM,IAAI,uBAAuB,kCAAkC;AAAA,IACrE;AACA,UAAM,aAAa,MAAM,OAAO,KAAK,YAAY,OAAO;AAAA,MACtD,GAAG,oBAAoB,OAAO;AAAA,MAC9B,UAAU,eAAe,QAAQ;AAAA,MACjC,QAAQ;AAAA,IAAA,CACT;AACD,QAAI,QAAgB;AACpB,QAAI,WAAoB;AACxB,QAAI;AACF,uBAAiB,SAAS,YAAY;AACpC,YAAI,QAAQ,QAAQ,SAAS;AAC3B,gBAAM,IAAI,uBAAuB,+BAA+B;AAAA,QAClE;AACA,cAAM,SAAS,MAAM,QAAQ,CAAC;AAC9B,cAAM,QAAQ,QAAQ,OAAO,WAAW;AACxC,YAAI,OAAO;AACT,gBAAM,EAAE,MAAM,OAAO,OAAO,MAAM,MAAA;AAClC,mBAAS;AAAA,QACX;AACA,YAAI,QAAQ,eAAe;AACzB,qBAAW;AACX,gBAAM,EAAE,MAAM,IAAI,OAAO,MAAM,KAAA;AAC/B,mBAAS;AAAA,QACX;AAAA,MACF;AACA,UAAI,CAAC,UAAU;AACb,cAAM,EAAE,MAAM,IAAI,OAAO,MAAM,KAAA;AAAA,MACjC;AAAA,IACF,SAAS,KAAK;AACZ,UAAI,eAAe,uBAAwB,OAAM;AACjD,YAAM,IAAI,eAAe,gCAAgC,GAAG;AAAA,IAC9D;AAAA,EACF;AAAA,EAEA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,YAAM,KAAK,OAAO,OAAA;AAClB,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEQ,gBAA2B;AACjC,QAAI,CAAC,KAAK,QAAQ;AAChB,YAAM,IAAI,oBAAoB,mDAAmD;AAAA,IACnF;AACA,WAAO,KAAK;AAAA,EACd;AACF;AC9IO,MAAM,gBAAuD,OAAO,OAAO;AAAA,EAChF,qBAAqB;AAAA,IACnB,IAAI;AAAA,IACJ,QAAQ;AAAA,IACR,YAAY;AAAA,IACZ,cAAc;AAAA,IACd,UAAU;AAAA,IACV,eAAe;AAAA,IACf,aAAa;AAAA,EAAA;AAAA,EAEf,qBAAqB;AAAA,IACnB,IAAI;AAAA,IACJ,QAAQ;AAAA,IACR,YAAY;AAAA,IACZ,cAAc;AAAA,IACd,UAAU;AAAA,IACV,eAAe;AAAA,IACf,aAAa;AAAA,EAAA;AAAA,EAEf,qBAAqB;AAAA,IACnB,IAAI;AAAA,IACJ,QAAQ;AAAA,IACR,YAAY;AAAA,IACZ,cAAc;AAAA,IACd,UAAU;AAAA,IACV,eAAe;AAAA,IACf,aAAa;AAAA,EAAA;AAEjB,CAAC;AASM,SAAS,mBAAmB,SAA8B;AAC/D,QAAM,SAAS,cAAc,OAAO;AACpC,MAAI,CAAC,QAAQ;AACX,UAAM,YAAY,OAAO,KAAK,aAAa,EAAE,KAAK,IAAI;AACtD,UAAM,IAAI,kBAAkB,kBAAkB,OAAO,wBAAwB,SAAS,GAAG;AAAA,EAC3F;AACA,SAAO;AACT;AAGO,SAAS,sBAAgC;AAC9C,SAAO,OAAO,KAAK,aAAa;AAClC;AC7BO,MAAe,OAAO;AAAA,EACjB,YAEW,QAEH,QAChB;AAHmB,SAAA,SAAA;AAEH,SAAA,SAAA;AAAA,EACf;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUH,aAAuB,aACrB,SACA,UAA+B,IACN;AACzB,UAAM,SAAS,mBAAmB,OAAO;AACzC,UAAM,SAAS,QAAQ,UAAU,IAAI,aAAA;AACrC,QAAI,CAAC,OAAO,YAAY;AACtB,YAAM,OAAO,KAAK,OAAO,UAAU,QAAQ,UAAU;AAAA,IACvD;AACA,WAAO,EAAE,QAAQ,OAAA;AAAA,EACnB;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,UAAM,KAAK,OAAO,OAAA;AAAA,EACpB;AAAA;AAAA,EAGA,WAAoB;AAClB,WAAO,KAAK,OAAO,SAAA;AAAA,EACrB;AACF;AC9DO,MAAM,UAAU;AAAA,EACrB,YAEkB,MAEA,SAEA,iBAEA,cAChB;AAPgB,SAAA,OAAA;AAEA,SAAA,UAAA;AAEA,SAAA,kBAAA;AAEA,SAAA,eAAA;AAAA,EACf;AACL;ACQO,MAAM,aAAa,OAAO;AAAA,EACd,UAAqB,CAAA;AAAA,EAC9B,eAA8B;AAAA,EAE9B,YAAY,QAAgB,QAAqB;AACvD,UAAM,QAAQ,MAAM;AAAA,EACtB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,aAAa,OAAO,SAAiB,UAA+B,IAAmB;AACrF,UAAM,EAAE,QAAQ,OAAA,IAAW,MAAM,OAAO,aAAa,SAAS,OAAO;AACrE,WAAO,IAAI,KAAK,QAAQ,MAAM;AAAA,EAChC;AAAA;AAAA,EAGA,gBAAgB,QAAsB;AACpC,SAAK,eAAe;AAAA,EACtB;AAAA;AAAA,EAGA,oBAA0B;AACxB,SAAK,eAAe;AAAA,EACtB;AAAA;AAAA,EAGA,eAAqB;AACnB,SAAK,QAAQ,SAAS;AAAA,EACxB;AAAA;AAAA,EAGA,aAAiC;AAC/B,WAAO,KAAK,QAAQ,MAAA;AAAA,EACtB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWA,MAAM,KAAK,SAAiB,UAA6B,IAAwB;AAC/E,UAAM,WAAW,KAAK,cAAc,OAAO;AAC3C,UAAM,OAAO,MAAM,KAAK,OAAO,SAAS,UAAU,OAAO;AACzD,UAAM,UAAmB,EAAE,MAAM,QAAQ,SAAS,QAAA;AAClD,UAAM,eAAwB,EAAE,MAAM,aAAa,SAAS,KAAA;AAC5D,SAAK,QAAQ,KAAK,SAAS,YAAY;AACvC,WAAO,IAAI,UAAU,MAAM,cAAc,GAAG,MAAM;AAAA,EACpD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWA,OAAO,OAAO,SAAiB,UAA6B,IAA+B;AACzF,UAAM,WAAW,KAAK,cAAc,OAAO;AAC3C,UAAM,UAAmB,EAAE,MAAM,QAAQ,SAAS,QAAA;AAClD,QAAI,MAAc;AAClB,qBAAiB,SAAS,KAAK,OAAO,OAAO,UAAU,OAAO,GAAG;AAC/D,aAAO,MAAM;AACb,YAAM;AAAA,IACR;AACA,UAAM,eAAwB,EAAE,MAAM,aAAa,SAAS,IAAA;AAC5D,SAAK,QAAQ,KAAK,SAAS,YAAY;AAAA,EACzC;AAAA,EAEQ,cAAc,aAAgC;AACpD,UAAM,WAAsB,CAAA;AAC5B,QAAI,KAAK,cAAc;AACrB,eAAS,KAAK,EAAE,MAAM,UAAU,SAAS,KAAK,cAAc;AAAA,IAC9D;AACA,aAAS,KAAK,GAAG,KAAK,OAAO;AAC7B,aAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,aAAa;AACpD,WAAO;AAAA,EACT;AACF;ACvGA,eAAsB,cAAc,QAAoD;AACtF,MAAI,MAAc;AAClB,mBAAiB,SAAS,QAAQ;AAChC,WAAO,MAAM;AAAA,EACf;AACA,SAAO;AACT;AAYA,gBAAuB,IACrB,QACA,SAC2B;AAC3B,mBAAiB,SAAS,QAAQ;AAChC,YAAQ,KAAK;AACb,UAAM;AAAA,EACR;AACF;ACMO,MAAM,UAAkB;"}

package/package.json

@@ -0,0 +1,88 @@
+{
+  "name": "localm-web",
+  "version": "0.1.0",
+  "description": "Browser-only TypeScript SDK for running LLMs and SLMs locally with WebGPU. Ultralytics-style DX, Vite-first.",
+  "type": "module",
+  "main": "./dist/index.js",
+  "module": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    }
+  },
+  "files": [
+    "dist",
+    "README.md",
+    "LICENSE",
+    "CHANGELOG.md"
+  ],
+  "sideEffects": false,
+  "scripts": {
+    "build": "vite build",
+    "dev": "vite build --watch",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "lint": "eslint .",
+    "format": "prettier --write \"**/*.{ts,tsx,js,json,md}\"",
+    "format:check": "prettier --check \"**/*.{ts,tsx,js,json,md}\"",
+    "typecheck": "tsc --noEmit",
+    "clean": "rm -rf dist"
+  },
+  "peerDependencies": {
+    "@mlc-ai/web-llm": "^0.2.79"
+  },
+  "devDependencies": {
+    "@mlc-ai/web-llm": "^0.2.79",
+    "@types/node": "^22.10.0",
+    "@typescript-eslint/eslint-plugin": "^8.18.0",
+    "@typescript-eslint/parser": "^8.18.0",
+    "eslint": "^9.17.0",
+    "prettier": "^3.4.0",
+    "typescript": "^5.7.0",
+    "vite": "^7.3.3",
+    "vite-plugin-dts": "^4.4.0",
+    "vitest": "^3.2.4"
+  },
+  "engines": {
+    "node": ">=18.0.0"
+  },
+  "overrides": {
+    "esbuild": "^0.25.0"
+  },
+  "keywords": [
+    "llm",
+    "slm",
+    "language-model",
+    "webgpu",
+    "browser",
+    "onnx",
+    "webllm",
+    "mlc",
+    "phi",
+    "llama",
+    "qwen",
+    "gemma",
+    "local-llm",
+    "on-device",
+    "vite",
+    "typescript",
+    "chat",
+    "embeddings",
+    "reranker"
+  ],
+  "license": "MIT",
+  "author": {
+    "name": "Mauricio Benjamin",
+    "email": "mauricio.benjamin@reloverelations.com"
+  },
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/mauriciobenjamin700/localm-web.git"
+  },
+  "homepage": "https://github.com/mauriciobenjamin700/localm-web#readme",
+  "bugs": {
+    "url": "https://github.com/mauriciobenjamin700/localm-web/issues"
+  }
+}