react-native-nobodywho 3.0.0 → 5.0.0

package/package.json

@@ -1,21 +1,25 @@
 {
   "name": "react-native-nobodywho",
-  "version": "3.0.0",
-  "description": "Run LLMs locally with offline inference — React Native bindings",
+  "version": "5.0.0",
+  "description": "Run LLMs locally and offline with React Native & Expo - Tool calling, RAG, Speech-to-Text, Text-to-Speech, and VAD",
   "main": "src/wrapper.ts",
   "types": "src/wrapper.ts",
   "exports": {
     ".": "./src/wrapper.ts"
   },
   "license": "EUPL-1.2",
-  "homepage": "https://nobodywho.ooo",
+  "homepage": "https://www.nobodywho.ai",
   "bugs": "https://github.com/nobodywho-ooo/nobodywho/issues",
   "keywords": [
     "react-native",
+    "expo",
     "ai",
     "llm",
     "local-ai",
     "inference",
+    "stt",
+    "tts",
+    "vad",
     "gguf",
     "on-device",
     "offline"

package/src/chat.ts

@@ -2,6 +2,7 @@ import {
   RustChat,
   SamplerConfig,
   MtpConfig,
+  ContextShiftOptions,
   type ChatStats,
 } from "../generated/ts/nobodywho";
 import { Model } from "./model";
@@ -10,6 +11,18 @@ import { TokenStream } from "./streaming";
 import type { Prompt } from "./prompt";
 import type { Tool } from "./tool";
 
+/**
+ * Settings to apply before a `complete` turn. An omitted field keeps what the
+ * chat has; a set one stays set, like a leading system message.
+ */
+export type Options = {
+  sampler?: SamplerConfig;
+  /** Replaces the chat's template variables wholesale. */
+  templateVariables?: Record<string, boolean>;
+  /** Re-selects the chat template, so the turn re-prefills from near token zero. */
+  tools?: Tool[];
+};
+
 /**
  * A chat session for local LLM inference.
  *
@@ -37,6 +50,9 @@ export class Chat {
    * device's physical core count (performance cores only, on Apple silicon) — hyperthreads
    * and efficiency cores make inference slower, not faster. Lower it to leave CPU headroom
    * for the rest of the app.
+   *
+   * `contextShift` sets how old turns are forgotten when the context is full; omit it for
+   * the defaults, or pass `{ enabled: false }` to disable shifting.
    */
   constructor(opts: {
     model: Model;
@@ -47,6 +63,7 @@ export class Chat {
     sampler?: SamplerConfig;
     mtp?: Partial<MtpConfig>;
     threadCount?: number;
+    contextShift?: Partial<ContextShiftOptions>;
   }) {
     this._inner = new RustChat(
       opts.model._inner,
@@ -57,6 +74,7 @@ export class Chat {
       opts.sampler ?? undefined,
       opts.mtp !== undefined ? MtpConfig.create(opts.mtp) : undefined,
       opts.threadCount ?? undefined,
+      opts.contextShift !== undefined ? ContextShiftOptions.create(opts.contextShift) : undefined,
     );
   }
 
@@ -84,6 +102,7 @@ export class Chat {
     sampler?: SamplerConfig;
     mtp?: Partial<MtpConfig>;
     threadCount?: number;
+    contextShift?: Partial<ContextShiftOptions>;
     onDownloadProgress?: (downloaded: number, total: number) => void;
   }): Promise<Chat> {
     const model = await Model.load({
@@ -107,6 +126,29 @@ export class Chat {
     return new TokenStream(this._inner.askWithPrompt(message._parts!));
   }
 
+  /**
+   * Answer a full list of messages, replacing the chat history.
+   *
+   * The list is the whole conversation, used as given: it must be non-empty and
+   * end in a user or tool message. A leading system message sets the chat's system
+   * prompt; leave it out and the prompt already on the chat is kept. A later one
+   * stays in the history, for the chat template to render in place. The response
+   * is appended, and the next `ask` continues from there.
+   *
+   * `options` follows the same rule for the chat's other settings.
+   */
+  complete(messages: Message[], options: Options = {}): TokenStream {
+    return new TokenStream(
+      this._inner.complete(messages.map(toInternal), {
+        sampler: options.sampler,
+        templateVariables: options.templateVariables
+          ? new Map(Object.entries(options.templateVariables))
+          : undefined,
+        tools: options.tools?.map((t) => t._inner),
+      }),
+    );
+  }
+
   /** Stop the current generation. */
   stopGeneration(): void {
     this._inner.stopGeneration();
@@ -145,12 +187,12 @@ export class Chat {
   }
 
   /** Tokenize a text message or multimodal prompt and return the token IDs.
-   * Each element is a token ID (number) for text, or null for image/audio embedding slots. */
-  async tokenize(message: string | Prompt): Promise<(number | null)[]> {
+   * Each element is a token ID (number) for text, or undefined for image/audio embedding slots. */
+  async tokenize(message: string | Prompt): Promise<(number | undefined)[]> {
     if (typeof message === "string") {
       return this._inner.tokenize(message);
     }
-    return this._inner.tokenizeWithPrompt(message._parts);
+    return this._inner.tokenizeWithPrompt(message._parts!);
   }
 
   /** Set the system prompt. */
@@ -158,6 +200,11 @@ export class Chat {
     return this._inner.setSystemPrompt(systemPrompt);
   }
 
+  /** Set how old turns are forgotten when the context is full. */
+  async setContextShift(options: Partial<ContextShiftOptions>): Promise<void> {
+    return this._inner.setContextShift(ContextShiftOptions.create(options));
+  }
+
   /** Set the tools available to the model. */
   async setTools(tools: Tool[]): Promise<void> {
     return this._inner.setTools(tools.map((t) => t._inner));

package/src/message.ts

@@ -1,19 +1,45 @@
 import {
+  ContentPart as InternalContentPart,
+  ContentPart_Tags,
   Message as InternalMessage,
   Message_Tags,
-  type Asset,
+  MessageContent as InternalMessageContent,
+  MessageContent_Tags,
   type ToolCall,
 } from "../generated/ts/nobodywho";
 
+/** One piece of a message: a run of text, or a media file at this position. */
+export type ContentPart =
+  | { type: "text"; text: string }
+  | { type: "image"; path: string }
+  | { type: "audio"; path: string };
+
+/**
+ * Message content: plain text, interleaved text and media, or a raw JSON value
+ * handed straight to the chat template.
+ *
+ * @example
+ * ```typescript
+ * const content: Content = [
+ *   { type: "text", text: "What is in this image?" },
+ *   { type: "image", path: "./dog.png" },
+ * ];
+ * ```
+ */
+export type Content = string | ContentPart[] | { type: "json"; json: string };
+
 /**
  * A chat message. The variant determines the message type:
  *
- * - **User message:** `{ role: "user", content, assets? }`
+ * - **User message:** `{ role: "user", content }`
  * - **Assistant message:** `{ role: "assistant", content }`
  * - **Assistant tool call:** `{ role: "assistant", content, toolCalls }`
  * - **System message:** `{ role: "system", content }`
  * - **Tool response:** `{ role: "tool", name, content }`
  *
+ * `content` is a string, a list of parts for multimodal input, or a
+ * `{ type: "json", json }` wrapper for raw template passthrough.
+ *
  * @example
  * ```typescript
  * const history = await chat.getChatHistory();
@@ -29,33 +55,75 @@ import {
  * ```
  */
 export type Message =
-  | { role: "user"; content: string; assets?: Asset[] }
-  | { role: "assistant"; content: string }
-  | { role: "assistant"; content: string; toolCalls: ToolCall[] }
-  | { role: "system"; content: string }
-  | { role: "tool"; name: string; content: string };
+  | { role: "user"; content: Content }
+  | { role: "assistant"; content: Content }
+  | { role: "assistant"; content: Content; toolCalls: ToolCall[] }
+  | { role: "system"; content: Content }
+  | { role: "tool"; name: string; content: Content };
+
+function partFromInternal(part: InternalContentPart): ContentPart {
+  if (part.tag === ContentPart_Tags.Text) {
+    return { type: "text", text: part.inner.text };
+  } else if (part.tag === ContentPart_Tags.Image) {
+    return { type: "image", path: part.inner.path };
+  } else {
+    return { type: "audio", path: part.inner.path };
+  }
+}
+
+function partToInternal(part: ContentPart): InternalContentPart {
+  if (part.type === "text") {
+    return new InternalContentPart.Text({ text: part.text });
+  } else if (part.type === "image") {
+    return new InternalContentPart.Image({ path: part.path });
+  } else {
+    return new InternalContentPart.Audio({ path: part.path });
+  }
+}
+
+/** @internal */
+export function contentFromInternal(content: InternalMessageContent): Content {
+  if (content.tag === MessageContent_Tags.Text) {
+    return content.inner.text;
+  } else if (content.tag === MessageContent_Tags.Parts) {
+    return content.inner.parts.map(partFromInternal);
+  } else {
+    return { type: "json", json: content.inner.json };
+  }
+}
+
+/** @internal */
+export function contentToInternal(content: Content): InternalMessageContent {
+  if (typeof content === "string") {
+    return new InternalMessageContent.Text({ text: content });
+  }
+  if (Array.isArray(content)) {
+    return new InternalMessageContent.Parts({
+      parts: content.map(partToInternal),
+    });
+  }
+  return new InternalMessageContent.Json({ json: content.json });
+}
 
 /** @internal Convert internal Message to Message */
 export function fromInternal(msg: InternalMessage): Message {
   if (msg.tag === Message_Tags.User) {
-    const { content, assets } = msg.inner;
-    return {
-      role: "user",
-      content,
-      ...(assets.length > 0 ? { assets } : {}),
-    };
+    return { role: "user", content: contentFromInternal(msg.inner.content) };
   } else if (msg.tag === Message_Tags.Assistant) {
     const { content, toolCalls } = msg.inner;
     if (toolCalls != null && toolCalls.length > 0) {
-      return { role: "assistant", content, toolCalls };
+      return {
+        role: "assistant",
+        content: contentFromInternal(content),
+        toolCalls,
+      };
     }
-    return { role: "assistant", content };
+    return { role: "assistant", content: contentFromInternal(content) };
   } else if (msg.tag === Message_Tags.System) {
-    const { content } = msg.inner;
-    return { role: "system", content };
+    return { role: "system", content: contentFromInternal(msg.inner.content) };
   } else {
     const { name, content } = msg.inner;
-    return { role: "tool", name, content };
+    return { role: "tool", name, content: contentFromInternal(content) };
   }
 }
 
@@ -63,27 +131,26 @@ export function fromInternal(msg: InternalMessage): Message {
 export function toInternal(msg: Message): InternalMessage {
   if (msg.role === "user") {
     return new InternalMessage.User({
-      content: msg.content,
-      assets: msg.assets ?? [],
+      content: contentToInternal(msg.content),
     });
   } else if (msg.role === "assistant" && "toolCalls" in msg) {
     return new InternalMessage.Assistant({
-      content: msg.content,
+      content: contentToInternal(msg.content),
       toolCalls: msg.toolCalls,
     });
   } else if (msg.role === "assistant") {
     return new InternalMessage.Assistant({
-      content: msg.content,
+      content: contentToInternal(msg.content),
       toolCalls: undefined,
     });
   } else if (msg.role === "system") {
     return new InternalMessage.System({
-      content: msg.content,
+      content: contentToInternal(msg.content),
     });
   } else {
     return new InternalMessage.Tool({
       name: msg.name,
-      content: msg.content,
+      content: contentToInternal(msg.content),
     });
   }
 }

package/src/model.ts

@@ -65,6 +65,11 @@ export class Model {
     return new Model(inner);
   }
 
+  /** The identifier used to load this model. */
+  get source(): string {
+    return this._inner.source();
+  }
+
   /**
    * The maximum context size this model was trained with.
    * Use this to avoid setting `contextSize` above what the model supports.

package/src/prompt.ts

@@ -1,32 +1,32 @@
-import { PromptPart } from "../generated/ts/nobodywho";
+import { ContentPart } from "../generated/ts/nobodywho";
 
 /** A text part of a multimodal prompt. */
 class TextPart {
   /** @internal */
-  readonly _inner: PromptPart;
+  readonly _inner: ContentPart;
 
   constructor(content: string) {
-    this._inner = new PromptPart.Text({ content });
+    this._inner = new ContentPart.Text({ text: content });
   }
 }
 
 /** An image part of a multimodal prompt. */
 class ImagePart {
   /** @internal */
-  readonly _inner: PromptPart;
+  readonly _inner: ContentPart;
 
   constructor(path: string) {
-    this._inner = new PromptPart.Image({ path });
+    this._inner = new ContentPart.Image({ path });
   }
 }
 
 /** An audio part of a multimodal prompt. */
 class AudioPart {
   /** @internal */
-  readonly _inner: PromptPart;
+  readonly _inner: ContentPart;
 
   constructor(path: string) {
-    this._inner = new PromptPart.Audio({ path });
+    this._inner = new ContentPart.Audio({ path });
   }
 }
 
@@ -53,7 +53,7 @@ type Part = TextPart | ImagePart | AudioPart;
  */
 export class Prompt {
   /** @internal */
-  readonly _parts: PromptPart[] | null;
+  readonly _parts: ContentPart[] | null;
   /** @internal */
   readonly _jsonString: string | null;
 

package/src/sampler_presets.ts

@@ -5,7 +5,6 @@ import {
   samplerPresetConstrainWithRegex,
   samplerPresetDefault,
   samplerPresetDry,
-  samplerPresetGrammar,
   samplerPresetGreedy,
   samplerPresetJson,
   samplerPresetTemperature,
@@ -16,6 +15,10 @@ import {
 /**
  * Static factory methods for common sampler configurations.
  *
+ * Every preset builds on the default configuration and adds its own step on top,
+ * replacing the default step of the same kind if there is one. `greedy()` is the
+ * exception: it always picks the most probable token, so it needs no steps.
+ *
  * @example
  * ```typescript
  * const sampler = SamplerPresets.temperature(0.7);
@@ -30,12 +33,12 @@ export class SamplerPresets {
     return samplerPresetDefault() as SamplerConfig;
   }
 
-  /** Create a sampler with top-k filtering only. */
+  /** Create a sampler with the default steps, but top-k overridden. */
   static topK(topK: number): SamplerConfig {
     return samplerPresetTopK(topK) as SamplerConfig;
   }
 
-  /** Create a sampler with nucleus (top-p) sampling. */
+  /** Create a sampler with the default steps, but nucleus (top-p) overridden. */
   static topP(topP: number): SamplerConfig {
     return samplerPresetTopP(topP) as SamplerConfig;
   }
@@ -55,7 +58,7 @@ export class SamplerPresets {
     return samplerPresetDry() as SamplerConfig;
   }
 
-  /** Create a sampler configured for JSON output generation (any valid JSON, GBNF-based). */
+  /** Constrain output to a JSON object of any shape. */
   static json(): SamplerConfig {
     return samplerPresetJson() as SamplerConfig;
   }
@@ -80,11 +83,4 @@ export class SamplerPresets {
   static constrainWithGrammar(grammar: string): SamplerConfig {
     return samplerPresetConstrainWithGrammar(grammar) as SamplerConfig;
   }
-
-  /**
-   * @deprecated Use {@link constrainWithGrammar} instead. It accepts both Lark and GBNF strings.
-   */
-  static grammar(grammar: string): SamplerConfig {
-    return samplerPresetGrammar(grammar) as SamplerConfig;
-  }
 }

package/src/wrapper.ts

@@ -4,7 +4,7 @@
  * This is the public entry point. It re-exports the public API:
  * - Wrapper classes: Model, Chat, Encoder, CrossEncoder, TextToSpeech, Tool, TokenStream, Prompt
  * - Direct re-exports: SamplerBuilder, SamplerConfig
- * - Types: Message, Asset, ToolCall
+ * - Types: Message, Content, ContentPart, ToolCall
  * - Utilities: SamplerPresets, cosineSimilarity, downloadModel
  *
  * This file is NOT generated — it is safe to edit.
@@ -57,6 +57,7 @@ export async function downloadModel(opts: {
 // Wrapper classes
 export { Model } from "./model";
 export { Chat } from "./chat";
+export type { Options } from "./chat";
 export { SpeechToText } from "./speech_to_text";
 export { VoiceActivityDetection } from "./vad";
 export { Encoder } from "./encoder";
@@ -67,7 +68,7 @@ export { Tool } from "./tool";
 export { TokenStream } from "./streaming";
 
 // Message types
-export type { Message } from "./message";
+export type { Message, Content, ContentPart } from "./message";
 
 // Re-export types that don't need wrapping.
 //
@@ -76,15 +77,16 @@ export type { Message } from "./message";
 export {
   SamplerBuilder,
   SamplerConfig,
+  ShiftTarget,
   cosineSimilarity,
   getCachedModels,
 } from "../generated/ts/nobodywho";
 
 export type {
-  Asset,
   ToolCall,
   CachedModel,
   ChatStats,
+  ContextShiftOptions,
 } from "../generated/ts/nobodywho";
 export type { TextToSpeechArchitecture, TextToSpeechDevice, TextToSpeechOptions } from "./text_to_speech";
 export type { SpeechToTextOptions } from "./speech_to_text";