react-native-nobodywho 4.0.0 → 5.0.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "react-native-nobodywho",
-  "version": "4.0.0",
+  "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",

package/src/chat.ts

@@ -2,6 +2,7 @@ import {
   RustChat,
   SamplerConfig,
   MtpConfig,
+  ContextShiftOptions,
   type ChatStats,
 } from "../generated/ts/nobodywho";
 import { Model } from "./model";
@@ -49,6 +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;
@@ -59,6 +63,7 @@ export class Chat {
     sampler?: SamplerConfig;
     mtp?: Partial<MtpConfig>;
     threadCount?: number;
+    contextShift?: Partial<ContextShiftOptions>;
   }) {
     this._inner = new RustChat(
       opts.model._inner,
@@ -69,6 +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,
     );
   }
 
@@ -96,6 +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({
@@ -122,11 +129,11 @@ export class Chat {
   /**
    * Answer a full list of messages, replacing the chat history.
    *
-   * The list is the whole conversation, used as given: it must be non-empty, end
-   * in a user or tool message, and carry a system message only first. That system
-   * message sets the chat's system prompt; leave it out and the prompt already on
-   * the chat is kept. The response is appended, and the next `ask` continues from
-   * there.
+   * 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.
    */
@@ -180,12 +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. */
@@ -193,6 +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/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/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

@@ -77,6 +77,7 @@ export type { Message, Content, ContentPart } from "./message";
 export {
   SamplerBuilder,
   SamplerConfig,
+  ShiftTarget,
   cosineSimilarity,
   getCachedModels,
 } from "../generated/ts/nobodywho";
@@ -85,6 +86,7 @@ export type {
   ToolCall,
   CachedModel,
   ChatStats,
+  ContextShiftOptions,
 } from "../generated/ts/nobodywho";
 export type { TextToSpeechArchitecture, TextToSpeechDevice, TextToSpeechOptions } from "./text_to_speech";
 export type { SpeechToTextOptions } from "./speech_to_text";