@agentskit/adapters 0.10.0 → 0.12.0

package/dist/index.d.cts

@@ -374,6 +374,57 @@ interface BedrockInvokeInput {
 declare function bedrock(config: BedrockConfig): AdapterFactory;
 declare const bedrockAdapter: typeof bedrock;
 
+/**
+ * Browser-only adapter backed by WebLLM (https://github.com/mlc-ai/web-llm).
+ * Models run on-device via WebGPU; no network for inference. The MLCEngine
+ * is loaded lazily on first stream so apps can ship the import without
+ * paying the wasm cost up front.
+ *
+ * `@mlc-ai/web-llm` is an **optional peer dependency** — install it
+ * alongside this package when you opt into browser-only inference.
+ */
+interface WebLlmConfig {
+    /** Model id from MLC's catalog, e.g. `Llama-3.1-8B-Instruct-q4f16_1-MLC`. */
+    model: string;
+    /**
+     * Override the engine to inject a pre-loaded one (the MLCEngine spin-up
+     * is non-trivial — apps usually warm it once, not per turn).
+     */
+    engine?: WebLlmEngineLike;
+    /** Engine progress callback (model download / compile percent). */
+    onProgress?: (info: {
+        progress: number;
+        text: string;
+    }) => void;
+}
+interface WebLlmEngineLike {
+    reload(model: string, opts?: {
+        initProgressCallback?: (i: {
+            progress: number;
+            text: string;
+        }) => void;
+    }): Promise<void>;
+    chat: {
+        completions: {
+            create(params: {
+                messages: Array<{
+                    role: string;
+                    content: string;
+                }>;
+                stream: true;
+            }): AsyncIterable<{
+                choices: Array<{
+                    delta?: {
+                        content?: string;
+                    };
+                }>;
+            }>;
+        };
+    };
+}
+declare function webllm(config: WebLlmConfig): AdapterFactory;
+declare const webllmAdapter: typeof webllm;
+
 interface RouterCandidate {
     id: string;
     adapter: AdapterFactory;
@@ -605,4 +656,4 @@ interface KimiEmbedderConfig extends OpenAICompatibleEmbedderConfig {
 }
 declare const kimiEmbedder: (config: OpenAICompatibleEmbedderConfig) => _agentskit_core.EmbedFn;
 
-export { type AnthropicConfig, type AzureOpenAIConfig, type BailConfig, type BedrockConfig, type BedrockRuntimeClientLike, type CerebrasConfig, type CohereConfig, type CreateAdapterConfig, type DeepSeekConfig, type DeepSeekEmbedderConfig, type EnsembleAggregator, type EnsembleBranchResult, type EnsembleCandidate, type EnsembleOptions, type FallbackCandidate, type FallbackOptions, type FireworksConfig, type GeminiConfig, type GeminiEmbedderConfig, type GenericAdapterConfig, type GrokConfig, type GrokEmbedderConfig, type GroqConfig, type HuggingFaceConfig, type KimiConfig, type KimiEmbedderConfig, type LMStudioConfig, type LangChainConfig, type LangGraphConfig, type LlamaCppConfig, type MistralConfig, type MockAdapterOptions, type MockResponse, type OllamaConfig, type OllamaEmbedderConfig, type OpenAICompatibleEmbedderConfig, type OpenAIConfig, type OpenAIEmbedderConfig, type OpenRouterConfig, type RecordedTurn, type RecordingFixture, type RecordingSink, type ReplicateConfig, type RetryOptions, type RouterCandidate, type RouterOptions, type RouterPolicy, type TogetherConfig, type VLLMConfig, type VercelAIConfig, type VertexConfig, anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm };
+export { type AnthropicConfig, type AzureOpenAIConfig, type BailConfig, type BedrockConfig, type BedrockRuntimeClientLike, type CerebrasConfig, type CohereConfig, type CreateAdapterConfig, type DeepSeekConfig, type DeepSeekEmbedderConfig, type EnsembleAggregator, type EnsembleBranchResult, type EnsembleCandidate, type EnsembleOptions, type FallbackCandidate, type FallbackOptions, type FireworksConfig, type GeminiConfig, type GeminiEmbedderConfig, type GenericAdapterConfig, type GrokConfig, type GrokEmbedderConfig, type GroqConfig, type HuggingFaceConfig, type KimiConfig, type KimiEmbedderConfig, type LMStudioConfig, type LangChainConfig, type LangGraphConfig, type LlamaCppConfig, type MistralConfig, type MockAdapterOptions, type MockResponse, type OllamaConfig, type OllamaEmbedderConfig, type OpenAICompatibleEmbedderConfig, type OpenAIConfig, type OpenAIEmbedderConfig, type OpenRouterConfig, type RecordedTurn, type RecordingFixture, type RecordingSink, type ReplicateConfig, type RetryOptions, type RouterCandidate, type RouterOptions, type RouterPolicy, type TogetherConfig, type VLLMConfig, type VercelAIConfig, type VertexConfig, type WebLlmConfig, type WebLlmEngineLike, anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm, webllm, webllmAdapter };

package/dist/index.d.ts

@@ -374,6 +374,57 @@ interface BedrockInvokeInput {
 declare function bedrock(config: BedrockConfig): AdapterFactory;
 declare const bedrockAdapter: typeof bedrock;
 
+/**
+ * Browser-only adapter backed by WebLLM (https://github.com/mlc-ai/web-llm).
+ * Models run on-device via WebGPU; no network for inference. The MLCEngine
+ * is loaded lazily on first stream so apps can ship the import without
+ * paying the wasm cost up front.
+ *
+ * `@mlc-ai/web-llm` is an **optional peer dependency** — install it
+ * alongside this package when you opt into browser-only inference.
+ */
+interface WebLlmConfig {
+    /** Model id from MLC's catalog, e.g. `Llama-3.1-8B-Instruct-q4f16_1-MLC`. */
+    model: string;
+    /**
+     * Override the engine to inject a pre-loaded one (the MLCEngine spin-up
+     * is non-trivial — apps usually warm it once, not per turn).
+     */
+    engine?: WebLlmEngineLike;
+    /** Engine progress callback (model download / compile percent). */
+    onProgress?: (info: {
+        progress: number;
+        text: string;
+    }) => void;
+}
+interface WebLlmEngineLike {
+    reload(model: string, opts?: {
+        initProgressCallback?: (i: {
+            progress: number;
+            text: string;
+        }) => void;
+    }): Promise<void>;
+    chat: {
+        completions: {
+            create(params: {
+                messages: Array<{
+                    role: string;
+                    content: string;
+                }>;
+                stream: true;
+            }): AsyncIterable<{
+                choices: Array<{
+                    delta?: {
+                        content?: string;
+                    };
+                }>;
+            }>;
+        };
+    };
+}
+declare function webllm(config: WebLlmConfig): AdapterFactory;
+declare const webllmAdapter: typeof webllm;
+
 interface RouterCandidate {
     id: string;
     adapter: AdapterFactory;
@@ -605,4 +656,4 @@ interface KimiEmbedderConfig extends OpenAICompatibleEmbedderConfig {
 }
 declare const kimiEmbedder: (config: OpenAICompatibleEmbedderConfig) => _agentskit_core.EmbedFn;
 
-export { type AnthropicConfig, type AzureOpenAIConfig, type BailConfig, type BedrockConfig, type BedrockRuntimeClientLike, type CerebrasConfig, type CohereConfig, type CreateAdapterConfig, type DeepSeekConfig, type DeepSeekEmbedderConfig, type EnsembleAggregator, type EnsembleBranchResult, type EnsembleCandidate, type EnsembleOptions, type FallbackCandidate, type FallbackOptions, type FireworksConfig, type GeminiConfig, type GeminiEmbedderConfig, type GenericAdapterConfig, type GrokConfig, type GrokEmbedderConfig, type GroqConfig, type HuggingFaceConfig, type KimiConfig, type KimiEmbedderConfig, type LMStudioConfig, type LangChainConfig, type LangGraphConfig, type LlamaCppConfig, type MistralConfig, type MockAdapterOptions, type MockResponse, type OllamaConfig, type OllamaEmbedderConfig, type OpenAICompatibleEmbedderConfig, type OpenAIConfig, type OpenAIEmbedderConfig, type OpenRouterConfig, type RecordedTurn, type RecordingFixture, type RecordingSink, type ReplicateConfig, type RetryOptions, type RouterCandidate, type RouterOptions, type RouterPolicy, type TogetherConfig, type VLLMConfig, type VercelAIConfig, type VertexConfig, anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm };
+export { type AnthropicConfig, type AzureOpenAIConfig, type BailConfig, type BedrockConfig, type BedrockRuntimeClientLike, type CerebrasConfig, type CohereConfig, type CreateAdapterConfig, type DeepSeekConfig, type DeepSeekEmbedderConfig, type EnsembleAggregator, type EnsembleBranchResult, type EnsembleCandidate, type EnsembleOptions, type FallbackCandidate, type FallbackOptions, type FireworksConfig, type GeminiConfig, type GeminiEmbedderConfig, type GenericAdapterConfig, type GrokConfig, type GrokEmbedderConfig, type GroqConfig, type HuggingFaceConfig, type KimiConfig, type KimiEmbedderConfig, type LMStudioConfig, type LangChainConfig, type LangGraphConfig, type LlamaCppConfig, type MistralConfig, type MockAdapterOptions, type MockResponse, type OllamaConfig, type OllamaEmbedderConfig, type OpenAICompatibleEmbedderConfig, type OpenAIConfig, type OpenAIEmbedderConfig, type OpenRouterConfig, type RecordedTurn, type RecordingFixture, type RecordingSink, type ReplicateConfig, type RetryOptions, type RouterCandidate, type RouterOptions, type RouterPolicy, type TogetherConfig, type VLLMConfig, type VercelAIConfig, type VertexConfig, type WebLlmConfig, type WebLlmEngineLike, anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm, webllm, webllmAdapter };

package/dist/index.js

@@ -1272,6 +1272,69 @@ function bedrock(config) {
 }
 var bedrockAdapter = bedrock;
 
+// src/webllm.ts
+var cachedSdk2 = null;
+async function loadSdk2() {
+  if (!cachedSdk2) {
+    cachedSdk2 = (async () => {
+      try {
+        const moduleId = "@mlc-ai/web-llm";
+        return await import(
+          /* @vite-ignore */
+          moduleId
+        );
+      } catch {
+        throw new Error("Install @mlc-ai/web-llm to use the webllm adapter: npm install @mlc-ai/web-llm");
+      }
+    })();
+  }
+  return cachedSdk2;
+}
+function webllm(config) {
+  let enginePromise = null;
+  const getEngine = () => {
+    if (config.engine) return Promise.resolve(config.engine);
+    if (!enginePromise) {
+      enginePromise = (async () => {
+        const sdk = await loadSdk2();
+        return sdk.CreateMLCEngine(config.model, { initProgressCallback: config.onProgress });
+      })();
+    }
+    return enginePromise;
+  };
+  return {
+    capabilities: {
+      streaming: true,
+      tools: false
+    },
+    createSource: (request) => {
+      let aborted = false;
+      return {
+        stream: async function* () {
+          if (aborted) return;
+          try {
+            const engine = await getEngine();
+            const messages = request.messages.map((m) => ({ role: m.role, content: m.content }));
+            const completion = await engine.chat.completions.create({ messages, stream: true });
+            for await (const chunk of completion) {
+              if (aborted) return;
+              const delta = chunk.choices[0]?.delta?.content;
+              if (delta) yield { type: "text", content: delta };
+            }
+            yield { type: "done" };
+          } catch (err) {
+            yield { type: "error", content: err instanceof Error ? err.message : String(err) };
+          }
+        },
+        abort: () => {
+          aborted = true;
+        }
+      };
+    }
+  };
+}
+var webllmAdapter = webllm;
+
 // src/router.ts
 function requireCapabilities(request) {
   const needs = {};
@@ -1796,6 +1859,6 @@ var grokEmbedder = createOpenAICompatibleEmbedder("Grok", "https://api.x.ai");
 // src/embedders/kimi.ts
 var kimiEmbedder = createOpenAICompatibleEmbedder("Kimi", "https://api.moonshot.ai");
 
-export { anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm };
+export { anthropic, azureOpenAI, azureOpenAIAdapter, bail, bailAdapter, bedrock, bedrockAdapter, cerebras, cerebrasAdapter, chunkText, cohere, cohereAdapter, createAdapter, createEnsembleAdapter, createFallbackAdapter, createOpenAICompatibleEmbedder, createRouter, deepseek, deepseekEmbedder, fetchWithRetry, fireworks, gemini, geminiEmbedder, generic, grok, grokEmbedder, groq, groqAdapter, huggingface, inMemorySink, kimi, kimiEmbedder, langchain, langgraph, llamacpp, lmstudio, mistral, mockAdapter, ollama, ollamaEmbedder, openai, openaiEmbedder, openrouter, qwen, recordingAdapter, replayAdapter, replicate, replicateAdapter, simulateStream, together, vercelAI, vertex, vertexAdapter, vllm, webllm, webllmAdapter };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map