veryfront 0.1.47 → 0.1.49

package/src/src/provider/local/ai-sdk-adapter.ts

@@ -18,22 +18,30 @@ import { isLocalAIDisabled } from "./env.js";
 
 const logger = serverLogger.component("local-llm");
 
+/** Shape of a single message in the AI SDK LanguageModelV2 prompt array. */
+interface PromptMessage {
+  role: string;
+  content: string | ReadonlyArray<{ type: string; text?: string }>;
+}
+
 /**
  * Convert AI SDK LanguageModelV2 prompt format to simple ChatMessage array.
  *
  * The AI SDK prompt is an array of message objects with role and content arrays.
  * We extract text content for the local model.
  */
-// deno-lint-ignore no-explicit-any
-function convertPrompt(prompt: any[]): ChatMessage[] {
+function convertPrompt(prompt: ReadonlyArray<PromptMessage>): ChatMessage[] {
   const messages: ChatMessage[] = [];
 
   for (const msg of prompt) {
-    const role = msg.role as "system" | "user" | "assistant" | "tool";
     // Skip tool messages — local models don't support tool calling
-    if (role === "tool") continue;
+    if (msg.role === "tool") continue;
 
-    const mappedRole = role === "system" ? "system" : role === "user" ? "user" : "assistant";
+    const mappedRole = msg.role === "system"
+      ? "system"
+      : msg.role === "user"
+      ? "user"
+      : "assistant";
 
     // Extract text content from content array
     let text = "";
@@ -55,6 +63,27 @@ function convertPrompt(prompt: any[]): ChatMessage[] {
   return messages;
 }
 
+/** AI SDK LanguageModelV2 generation options that both doGenerate and doStream receive. */
+interface LocalModelOptions {
+  prompt: unknown[];
+  maxOutputTokens?: number;
+  temperature?: number;
+  topP?: number;
+  topK?: number;
+  stopSequences?: string[];
+}
+
+/** Map AI SDK generation options to local engine GenerateOptions. */
+function toGenerateOptions(options: LocalModelOptions): GenerateOptions {
+  return {
+    maxNewTokens: options.maxOutputTokens ?? 512,
+    temperature: options.temperature ?? 0.7,
+    topP: options.topP,
+    topK: options.topK,
+    stopSequences: options.stopSequences,
+  };
+}
+
 /**
  * Create a local AI SDK LanguageModel for the given model ID.
  *
@@ -75,22 +104,9 @@ export function createLocalModel(modelId?: string): LanguageModel {
 
     supportedUrls: {},
 
-    async doGenerate(options: {
-      prompt: unknown[];
-      maxOutputTokens?: number;
-      temperature?: number;
-      topP?: number;
-      topK?: number;
-      stopSequences?: string[];
-    }) {
-      const messages = convertPrompt(options.prompt as unknown[]);
-      const genOptions: GenerateOptions = {
-        maxNewTokens: options.maxOutputTokens ?? 512,
-        temperature: options.temperature ?? 0.7,
-        topP: options.topP,
-        topK: options.topK,
-        stopSequences: options.stopSequences,
-      };
+    async doGenerate(options: LocalModelOptions) {
+      const messages = convertPrompt(options.prompt as ReadonlyArray<PromptMessage>);
+      const genOptions = toGenerateOptions(options);
 
       logger.debug(`[local] doGenerate: ${messages.length} messages → ${resolvedId}`);
 
@@ -108,14 +124,7 @@ export function createLocalModel(modelId?: string): LanguageModel {
       };
     },
 
-    async doStream(options: {
-      prompt: unknown[];
-      maxOutputTokens?: number;
-      temperature?: number;
-      topP?: number;
-      topK?: number;
-      stopSequences?: string[];
-    }) {
+    async doStream(options: LocalModelOptions) {
       // Eagerly check if local AI is disabled — must throw before creating the
       // ReadableStream, otherwise the 200 response headers are already committed.
       // Note: getTransformers() in local-engine.ts also checks this, but we need
@@ -130,14 +139,8 @@ export function createLocalModel(modelId?: string): LanguageModel {
         );
       }
 
-      const messages = convertPrompt(options.prompt as unknown[]);
-      const genOptions: GenerateOptions = {
-        maxNewTokens: options.maxOutputTokens ?? 512,
-        temperature: options.temperature ?? 0.7,
-        topP: options.topP,
-        topK: options.topK,
-        stopSequences: options.stopSequences,
-      };
+      const messages = convertPrompt(options.prompt as ReadonlyArray<PromptMessage>);
+      const genOptions = toGenerateOptions(options);
 
       logger.debug(`[local] doStream: ${messages.length} messages → ${resolvedId}`);
 

package/src/src/provider/local/env.ts

@@ -18,7 +18,10 @@ export function isLocalAIDisabled(): boolean {
   const denoVal = (dntShim.dntGlobalThis as any).Deno?.env?.get?.("VERYFRONT_DISABLE_LOCAL_AI");
   if (denoVal === "1") return true;
 
-  if (typeof process !== "undefined" && process.env?.VERYFRONT_DISABLE_LOCAL_AI === "1") {
+  if (
+    typeof process !== "undefined" &&
+    process.env?.VERYFRONT_DISABLE_LOCAL_AI === "1"
+  ) {
     return true;
   }
 

package/src/src/provider/local/local-embedding-adapter.ts

@@ -0,0 +1,37 @@
+/**
+ * AI SDK Adapter for Local Embedding Models
+ *
+ * Bridges `@huggingface/transformers` local embedding inference to the
+ * AI SDK `EmbeddingModel` interface. This allows `embed()` and
+ * `embedMany()` to work with local models seamlessly.
+ *
+ * @module provider/local
+ */
+
+import type { EmbeddingModel } from "ai";
+import { embedTexts } from "./local-embedding-engine.js";
+import { DEFAULT_LOCAL_EMBEDDING_MODEL } from "./model-catalog.js";
+
+/**
+ * Create a local AI SDK EmbeddingModel for the given model ID.
+ *
+ * The returned object implements the EmbeddingModelV2 interface, making it
+ * compatible with all AI SDK embedding functions and the VeryFront
+ * embedding/RAG primitives.
+ */
+export function createLocalEmbeddingModel(modelId?: string): EmbeddingModel {
+  const resolvedId = modelId || DEFAULT_LOCAL_EMBEDDING_MODEL;
+
+  return {
+    specificationVersion: "v2",
+    provider: "local",
+    modelId: `local/${resolvedId}`,
+    maxEmbeddingsPerCall: undefined,
+    supportsParallelCalls: false,
+
+    async doEmbed({ values }: { values: string[] }) {
+      const embeddings = await embedTexts(resolvedId, values);
+      return { embeddings, usage: { tokens: 0 }, rawResponse: undefined, warnings: [] };
+    },
+  } as EmbeddingModel;
+}

package/src/src/provider/local/local-embedding-engine.ts

@@ -0,0 +1,88 @@
+/**
+ * Local Embedding Engine
+ *
+ * Singleton wrapper around `@huggingface/transformers` for server-side
+ * local embedding generation. Uses the `feature-extraction` pipeline
+ * with mean pooling and normalization.
+ *
+ * @module provider/local
+ */
+
+import { serverLogger } from "../../utils/index.js";
+import { type ModelInfo, resolveLocalEmbeddingModel } from "./model-catalog.js";
+import { getTransformers } from "./local-engine.js";
+
+const logger = serverLogger.component("local-embedding");
+
+// deno-lint-ignore no-explicit-any
+type Pipeline = any;
+
+/** Cached pipeline instances keyed by HuggingFace model ID */
+const pipelineCache = new Map<string, Pipeline>();
+
+/** Whether a model is currently being loaded (prevents concurrent loads) */
+const loadingLocks = new Map<string, Promise<Pipeline>>();
+
+/**
+ * Load a feature-extraction pipeline for the given model.
+ * Returns a cached pipeline if already loaded.
+ */
+async function loadPipeline(modelInfo: ModelInfo): Promise<Pipeline> {
+  const cacheKey = modelInfo.hfId;
+
+  const cached = pipelineCache.get(cacheKey);
+  if (cached) return cached;
+
+  const existingLock = loadingLocks.get(cacheKey);
+  if (existingLock) return existingLock;
+
+  const loadPromise = (async () => {
+    const transformers = await getTransformers();
+
+    logger.info(
+      `Loading local embedding model: ${modelInfo.hfId} (${modelInfo.dtype}, ~${modelInfo.sizeMB}MB)...`,
+    );
+
+    const pipe = await transformers.pipeline(
+      "feature-extraction",
+      modelInfo.hfId,
+      {
+        dtype: modelInfo.dtype,
+        device: "cpu",
+      },
+    );
+
+    logger.info(`Embedding model loaded: ${modelInfo.hfId}`);
+    pipelineCache.set(cacheKey, pipe);
+    loadingLocks.delete(cacheKey);
+    return pipe;
+  })();
+
+  loadingLocks.set(cacheKey, loadPromise);
+
+  try {
+    return await loadPromise;
+  } catch (error) {
+    loadingLocks.delete(cacheKey);
+    throw error;
+  }
+}
+
+/**
+ * Generate embeddings for an array of texts using a local model.
+ *
+ * Uses mean pooling and L2 normalization to produce unit vectors,
+ * matching the behavior of cloud embedding APIs.
+ */
+export async function embedTexts(
+  modelId: string,
+  texts: string[],
+): Promise<number[][]> {
+  const modelInfo = resolveLocalEmbeddingModel(modelId);
+  const pipe = await loadPipeline(modelInfo);
+
+  const pooling = modelInfo.pooling ?? "mean";
+  const output = await pipe(texts, { pooling, normalize: true });
+
+  return output.tolist() as number[][];
+}

package/src/src/provider/local/local-engine.ts

@@ -52,7 +52,7 @@ let transformersModule: TransformersModule | null = null;
  * Lazily import @huggingface/transformers.
  * Only loads when actually needed, keeping startup fast when API keys are present.
  */
-async function getTransformers(): Promise<TransformersModule> {
+export async function getTransformers(): Promise<TransformersModule> {
   if (transformersModule) return transformersModule;
 
   if (isLocalAIDisabled()) {

package/src/src/provider/local/model-catalog.ts

@@ -11,11 +11,13 @@ export interface ModelInfo {
   /** HuggingFace model repository ID */
   hfId: string;
   /** Quantization dtype for ONNX Runtime */
-  dtype: "q4" | "q8" | "fp32";
+  dtype: "q4" | "q8" | "fp16" | "fp32";
   /** Approximate download size in MB */
   sizeMB: number;
   /** Human-readable description */
   description: string;
+  /** Pooling strategy for embedding models (default: "mean") */
+  pooling?: "mean" | "last_token";
 }
 
 /**
@@ -43,11 +45,75 @@ const MODEL_CATALOG: Record<string, ModelInfo> = {
     sizeMB: 1000,
     description: "SmolLM2 1.7B — highest quality local model",
   },
+  "qwen3-1.7b": {
+    hfId: "onnx-community/Qwen3-1.7B-ONNX",
+    dtype: "q4",
+    sizeMB: 1200,
+    description: "Qwen3 1.7B — strong multilingual reasoning and tool use",
+  },
+  "gemma3-1b": {
+    hfId: "onnx-community/gemma-3-1b-it-ONNX",
+    dtype: "q4",
+    sizeMB: 700,
+    description: "Gemma 3 1B — Google's compact instruction-tuned model",
+  },
 };
 
 /** Default model used when no specific model ID is provided */
 export const DEFAULT_LOCAL_MODEL = "smollm2-135m";
 
+/**
+ * Catalog of supported local embedding models.
+ *
+ * Uses the same `q4` quantization as language models for consistency.
+ */
+const EMBEDDING_MODEL_CATALOG: Record<string, ModelInfo> = {
+  "all-MiniLM-L6-v2": {
+    hfId: "Xenova/all-MiniLM-L6-v2",
+    dtype: "q4",
+    sizeMB: 23,
+    description: "All-MiniLM-L6-v2 — fast 384-dim embeddings",
+  },
+  "nomic-embed-text-v1.5": {
+    hfId: "nomic-ai/nomic-embed-text-v1.5",
+    dtype: "q4",
+    sizeMB: 130,
+    description: "Nomic Embed Text v1.5 — 768-dim, variable-length embeddings",
+  },
+  "bge-base-en-v1.5": {
+    hfId: "Xenova/bge-base-en-v1.5",
+    dtype: "q4",
+    sizeMB: 110,
+    description: "BGE Base EN v1.5 — 768-dim, strong English embeddings",
+  },
+  "qwen3-embedding-0.6b": {
+    hfId: "onnx-community/Qwen3-Embedding-0.6B-ONNX",
+    dtype: "q8",
+    sizeMB: 620,
+    pooling: "last_token",
+    description: "Qwen3 Embedding 0.6B — SOTA multilingual embeddings",
+  },
+};
+
+/** Default embedding model used when no specific model ID is provided */
+export const DEFAULT_LOCAL_EMBEDDING_MODEL = "all-MiniLM-L6-v2";
+
+/**
+ * Resolve a friendly embedding model ID to its HuggingFace model info.
+ * Falls back to treating the ID as a raw HuggingFace repository ID.
+ */
+export function resolveLocalEmbeddingModel(modelId: string): ModelInfo {
+  const catalogEntry = EMBEDDING_MODEL_CATALOG[modelId];
+  if (catalogEntry) return catalogEntry;
+
+  return {
+    hfId: modelId,
+    dtype: "q4",
+    sizeMB: 0,
+    description: `Custom embedding model: ${modelId}`,
+  };
+}
+
 /**
  * Resolve a friendly model ID to its HuggingFace model info.
  * Falls back to treating the ID as a raw HuggingFace repository ID.

package/src/src/provider/model-registry.ts

@@ -137,6 +137,53 @@ function autoInitializeFromEnv(): void {
   }
 }
 
+/**
+ * Default cloud models to try when auto-upgrading from a local model.
+ * Ordered by preference: Anthropic → OpenAI → Google.
+ *
+ * NOTE: model IDs are hardcoded — update when default models change.
+ */
+const CLOUD_UPGRADE_CANDIDATES: Array<{
+  provider: string;
+  model: string;
+  hasKey: () => boolean;
+}> = [
+  {
+    provider: "anthropic",
+    model: "claude-sonnet-4-20250514",
+    hasKey: () => !!getAnthropicEnvConfig().apiKey,
+  },
+  {
+    provider: "openai",
+    model: "gpt-4o-mini",
+    hasKey: () => !!getOpenAIEnvConfig().apiKey,
+  },
+  {
+    provider: "google",
+    model: "gemini-2.0-flash",
+    hasKey: () => !!getGoogleGenAIEnvConfig().apiKey,
+  },
+];
+
+/**
+ * Find the first cloud provider with a valid API key.
+ *
+ * Returns a "provider/model" string that can be passed to `resolveModel`,
+ * or `null` if no cloud provider is available.
+ *
+ * This is intentionally a **query** — it does NOT resolve the model.
+ * The caller (agent runtime) decides whether to use it.
+ */
+export function findAvailableCloudModel(): string | null {
+  autoInitializeFromEnv();
+  for (const { provider, model, hasKey } of CLOUD_UPGRADE_CANDIDATES) {
+    if (!hasKey()) continue;
+    if (!manager.has(provider)) continue;
+    return `${provider}/${model}`;
+  }
+  return null;
+}
+
 /**
  * Resolve a "provider/model" string to an AI SDK LanguageModel instance.
  *

package/src/src/proxy/retry.ts

@@ -6,8 +6,6 @@
  * Check if a fetch error is a transient connection error worth retrying.
  * These occur when renderer pods are being recycled or temporarily unavailable.
  */
-import * as dntShim from "../../_dnt.shims.js";
-
 export function isRetryableConnectionError(error: unknown): boolean {
   if (!(error instanceof Error)) return false;
   const msg = error.message.toLowerCase();
@@ -19,10 +17,3 @@ export function isRetryableConnectionError(error: unknown): boolean {
     msg.includes("os error 111") // ECONNREFUSED
   );
 }
-
-/**
- * Sleep for the specified number of milliseconds.
- */
-export function sleep(ms: number): Promise<void> {
-  return new Promise((resolve) => dntShim.setTimeout(resolve, ms));
-}

package/src/src/proxy/tracing.ts

@@ -69,7 +69,7 @@ async function loadApis(): Promise<void> {
   propagationApi = await import("@opentelemetry/core");
 }
 
-export async function initializeOTLP(): Promise<void> {
+export async function initializeOTLPWithApis(): Promise<void> {
   if (initialized) return;
 
   const config = getConfig();
@@ -139,10 +139,6 @@ export async function shutdownOTLP(): Promise<void> {
   }
 }
 
-export function isOTLPEnabled(): boolean {
-  return initialized && tracerProvider !== null;
-}
-
 function getPropagator(): import("@opentelemetry/core").W3CTraceContextPropagator | null {
   if (!propagationApi) return null;
   return new propagationApi.W3CTraceContextPropagator();
@@ -239,8 +235,6 @@ export function getTraceContext(): { traceId?: string; spanId?: string } {
  * Span names for proxy tracing.
  */
 export const ProxySpanNames = {
-  PROXY_REQUEST: "proxy.request",
-  PROXY_PROCESS: "proxy.process",
   PROXY_TOKEN_FETCH: "proxy.token_fetch",
   PROXY_DOMAIN_LOOKUP: "proxy.domain_lookup",
   OAUTH_TOKEN_REQUEST: "oauth.token_request",
@@ -282,5 +276,3 @@ export async function withSpan<T>(
     span.end();
   }
 }
-
-export { initializeOTLP as initializeOTLPWithApis };

package/src/src/react/components/ai/agent-card.tsx

@@ -60,7 +60,7 @@ export const AgentCard = React.forwardRef<HTMLDivElement, AgentCardProps>(
 
           {tool.error
             ? (
-              <div className="mt-2 p-3 bg-red-50 dark:bg-red-900/20 text-red-900 dark:text-red-100 rounded-xl text-sm border border-red-200 dark:border-red-800">
+              <div className="mt-2 p-3 bg-red-50 text-red-900 rounded-xl text-sm border border-red-200">
                 Error: {tool.error}
               </div>
             )
@@ -90,7 +90,7 @@ export const AgentCard = React.forwardRef<HTMLDivElement, AgentCardProps>(
         {hasToolCalls
           ? (
             <div className="space-y-3">
-              <h3 className="text-sm font-semibold text-neutral-700 dark:text-neutral-300">
+              <h3 className="text-sm font-semibold text-[var(--card-foreground)]">
                 Tool Calls
               </h3>
               <div className="space-y-3">
@@ -107,7 +107,7 @@ export const AgentCard = React.forwardRef<HTMLDivElement, AgentCardProps>(
         {hasMessages
           ? (
             <div className="space-y-3">
-              <h3 className="text-sm font-semibold text-neutral-700 dark:text-neutral-300">
+              <h3 className="text-sm font-semibold text-[var(--card-foreground)]">
                 Messages
               </h3>
               <div className="space-y-2 max-h-96 overflow-y-auto">
@@ -119,12 +119,12 @@ export const AgentCard = React.forwardRef<HTMLDivElement, AgentCardProps>(
                   return (
                     <div
                       key={msg.id}
-                      className="text-sm p-3 rounded-xl bg-neutral-50 dark:bg-neutral-800"
+                      className="text-sm p-3 rounded-xl bg-[var(--accent)]"
                     >
-                      <span className="font-semibold capitalize text-neutral-900 dark:text-neutral-100">
+                      <span className="font-semibold capitalize text-[var(--foreground)]">
                         {msg.role}:
                       </span>
-                      <span className="text-neutral-600 dark:text-neutral-400 ml-1">
+                      <span className="text-[var(--card-foreground)] ml-1">
                         {text.substring(0, 200)}
                         {text.length > 200 ? "..." : ""}
                       </span>
@@ -145,18 +145,18 @@ AgentCard.displayName = "AgentCard";
 function getStatusColor(status: AgentStatus): string {
   switch (status) {
     case "idle":
-      return "bg-neutral-100 text-neutral-700 dark:bg-neutral-800 dark:text-neutral-300";
+      return "bg-[var(--accent)] text-[var(--card-foreground)]";
     case "thinking":
-      return "bg-blue-100 text-blue-700 dark:bg-blue-900/30 dark:text-blue-300";
+      return "bg-blue-100 text-blue-700";
     case "tool_execution":
-      return "bg-violet-100 text-violet-700 dark:bg-violet-900/30 dark:text-violet-300";
+      return "bg-violet-100 text-violet-700";
     case "streaming":
-      return "bg-green-100 text-green-700 dark:bg-green-900/30 dark:text-green-300";
+      return "bg-green-100 text-green-700";
     case "completed":
-      return "bg-emerald-100 text-emerald-700 dark:bg-emerald-900/30 dark:text-emerald-300";
+      return "bg-emerald-100 text-emerald-700";
     case "error":
-      return "bg-red-100 text-red-700 dark:bg-red-900/30 dark:text-red-300";
+      return "bg-red-100 text-red-700";
     default:
-      return "bg-neutral-100 text-neutral-700 dark:bg-neutral-800 dark:text-neutral-300";
+      return "bg-[var(--accent)] text-[var(--card-foreground)]";
   }
 }

package/src/src/react/components/ai/chat/components/animations.tsx

@@ -23,10 +23,40 @@ export function Loader({
       {["0ms", "150ms", "300ms"].map((animationDelay) => (
         <span
           key={animationDelay}
-          className="animate-bounce rounded-full bg-muted-foreground"
+          className="animate-bounce rounded-full bg-[var(--muted-foreground)]"
           style={{ width: dotSize, height: dotSize, animationDelay }}
         />
       ))}
     </div>
   );
 }
+
+export function FadeIn({
+  children,
+  duration = 300,
+  className,
+}: {
+  children: React.ReactNode;
+  duration?: number;
+  className?: string;
+}): React.ReactElement {
+  const [mounted, setMounted] = React.useState(false);
+
+  React.useEffect(() => {
+    const frame = requestAnimationFrame(() => setMounted(true));
+    return () => cancelAnimationFrame(frame);
+  }, []);
+
+  return (
+    <div
+      className={className}
+      style={{
+        opacity: mounted ? 1 : 0,
+        transform: mounted ? "translateY(0)" : "translateY(8px)",
+        transition: `opacity ${duration}ms ease-out, transform ${duration}ms ease-out`,
+      }}
+    >
+      {children}
+    </div>
+  );
+}