torus-ai 0.1.0 → 0.2.1

package/src/providers/anthropic.ts

@@ -85,10 +85,19 @@ export class AnthropicProvider implements ModelProvider {
 function toApiMessage(m: Message): any {
   return {
     role: m.role,
-    content: m.content.map((b) => {
-      if (b.type === "text") return { type: "text", text: b.text };
-      if (b.type === "tool_use") return { type: "tool_use", id: b.id, name: b.name, input: b.input };
-      return { type: "tool_result", tool_use_id: b.toolUseId, content: b.content, is_error: b.isError };
+    content: m.content.flatMap((b): any[] => {
+      if (b.type === "text") return [{ type: "text", text: b.text }];
+      if (b.type === "tool_use") return [{ type: "tool_use", id: b.id, name: b.name, input: b.input }];
+      if (b.type === "tool_result") {
+        return [{ type: "tool_result", tool_use_id: b.toolUseId, content: b.content, is_error: b.isError }];
+      }
+      if (b.type === "image") {
+        const source = b.data
+          ? { type: "base64", media_type: b.mimeType ?? "image/png", data: b.data }
+          : { type: "url", url: b.url };
+        return [{ type: "image", source }];
+      }
+      return []; // video unsupported on Anthropic — drop the block
     }),
   };
 }

package/src/providers/cascade.ts

@@ -0,0 +1,118 @@
+import type { ModelProvider, ModelRequest, ModelResponse } from "../types.ts";
+import { DEEPSEEK_V4_PRO, KIMI_K2_6, LLAMA_VISION, NvidiaProvider } from "./nvidia.ts";
+import { GeminiProvider } from "./gemini.ts";
+
+// Orchestration: try a prioritized list of (provider, model) steps, falling
+// through to the next on failure (rate limit, error, or capability mismatch).
+// Capability-aware: image requests only go to vision steps; video requests only
+// to video steps — text-only models (Kimi, DeepSeek) are skipped for those.
+
+export interface CascadeStep {
+  provider: ModelProvider;
+  label: string; // e.g. "nvidia:kimi-k2.6"
+  vision: boolean; // accepts image input?
+  video?: boolean; // accepts video input? (default false)
+}
+
+export interface CascadeOptions {
+  steps: CascadeStep[];
+  /** Called when a step is skipped or fails and the cascade falls through. */
+  onFallback?: (info: { from: string; reason: string; needsVision: boolean }) => void;
+}
+
+export class CascadeProvider implements ModelProvider {
+  readonly name = "cascade";
+  private steps: CascadeStep[];
+  private onFallback?: CascadeOptions["onFallback"];
+
+  constructor(opts: CascadeOptions) {
+    if (!opts.steps.length) throw new Error("CascadeProvider needs at least one step.");
+    this.steps = opts.steps;
+    this.onFallback = opts.onFallback;
+  }
+
+  async generate(req: ModelRequest): Promise<ModelResponse> {
+    const has = (t: "image" | "video") =>
+      req.messages.some((m) => m.content.some((b) => b.type === t));
+    const needsVideo = has("video");
+    const needsImage = has("image");
+    const needsVision = needsImage || needsVideo;
+
+    const eligible = needsVideo
+      ? this.steps.filter((s) => s.video)
+      : needsImage
+        ? this.steps.filter((s) => s.vision)
+        : this.steps;
+
+    if (!eligible.length) {
+      throw new Error(
+        `Cascade: request needs ${needsVideo ? "video" : "image"} input but no step supports it.`,
+      );
+    }
+
+    let lastErr: unknown;
+    for (const step of eligible) {
+      try {
+        return await step.provider.generate(req);
+      } catch (err) {
+        lastErr = err;
+        this.onFallback?.({
+          from: step.label,
+          reason: (err as Error).message?.slice(0, 200) ?? "unknown",
+          needsVision,
+        });
+      }
+    }
+    throw new Error(`Cascade exhausted all steps. Last error: ${(lastErr as Error)?.message}`);
+  }
+}
+
+export interface DefaultProviderOptions {
+  nvidiaApiKey?: string;
+  googleApiKey?: string;
+  /** Override the main NVIDIA model (default Kimi K2.6). */
+  mainModel?: string;
+  /** Override the secondary NVIDIA model (default DeepSeek V4 Pro). */
+  secondaryModel?: string;
+  /** NVIDIA vision model for image requests (default llama-3.2-90b-vision). */
+  visionModel?: string;
+  /** Gemini model used as the final fallback option (default gemini-2.5-flash). */
+  geminiModel?: string;
+  onFallback?: CascadeOptions["onFallback"];
+}
+
+/**
+ * The SDK's recommended default: free NVIDIA endpoints first, Google as one
+ * fallback option. Capability-aware — image/video requests skip the text-only
+ * steps automatically.
+ *
+ *   1. NVIDIA Kimi K2.6                  — main; agentic + tools (text)
+ *   2. NVIDIA DeepSeek V4 Pro            — 1M-ctx text; skipped for media
+ *   3. NVIDIA Llama-3.2-90B-Vision       — image requests
+ *   4. Gemini 2.5 Flash                  — final fallback; image + video
+ */
+export function createDefaultProvider(opts: DefaultProviderOptions = {}): CascadeProvider {
+  const main = opts.mainModel ?? KIMI_K2_6;
+  const secondary = opts.secondaryModel ?? DEEPSEEK_V4_PRO;
+  const vision = opts.visionModel ?? LLAMA_VISION;
+  const gemini = opts.geminiModel ?? "gemini-2.5-flash";
+  const nv = (model: string) => new NvidiaProvider({ model, apiKey: opts.nvidiaApiKey });
+
+  return new CascadeProvider({
+    onFallback:
+      opts.onFallback ??
+      ((info) =>
+        console.warn(`[cascade] ${info.from} failed (${info.reason}); trying next`)),
+    steps: [
+      { provider: nv(main), label: `nvidia:${main}`, vision: false, video: false },
+      { provider: nv(secondary), label: `nvidia:${secondary}`, vision: false, video: false },
+      { provider: nv(vision), label: `nvidia:${vision}`, vision: true, video: false },
+      {
+        provider: new GeminiProvider({ model: gemini, apiKey: opts.googleApiKey }),
+        label: `gemini:${gemini}`,
+        vision: true,
+        video: true,
+      },
+    ],
+  });
+}

package/src/providers/gemini.ts

@@ -104,17 +104,27 @@ function toolUseNames(messages: Message[]): Map<string, string> {
 /** Translate one of our Messages into a Gemini `Content` (role + parts). */
 function toGeminiContent(m: Message, idToName: Map<string, string>): any {
   const role = m.role === "assistant" ? "model" : "user";
-  const parts = m.content.map((b) => {
-    if (b.type === "text") return { text: b.text };
-    if (b.type === "tool_use") return { functionCall: { id: b.id, name: b.name, args: b.input } };
-    // tool_result -> functionResponse
-    return {
-      functionResponse: {
-        id: b.toolUseId,
-        name: idToName.get(b.toolUseId) ?? b.toolUseId,
-        response: b.isError ? { error: b.content } : { result: b.content },
-      },
-    };
+  const parts = m.content.map((b): any => {
+    switch (b.type) {
+      case "text":
+        return { text: b.text };
+      case "image":
+      case "video":
+        // base64 -> inlineData; remote URL -> fileData
+        return b.data
+          ? { inlineData: { mimeType: b.mimeType ?? "image/png", data: b.data } }
+          : { fileData: { mimeType: b.mimeType ?? "image/png", fileUri: b.url ?? "" } };
+      case "tool_use":
+        return { functionCall: { id: b.id, name: b.name, args: b.input } };
+      case "tool_result":
+        return {
+          functionResponse: {
+            id: b.toolUseId,
+            name: idToName.get(b.toolUseId) ?? b.toolUseId,
+            response: b.isError ? { error: b.content } : { result: b.content },
+          },
+        };
+    }
   });
   return { role, parts };
 }

package/src/providers/nvidia.ts

@@ -0,0 +1,164 @@
+import type {
+  ContentBlock,
+  MediaBlock,
+  ModelProvider,
+  ModelRequest,
+  ModelResponse,
+} from "../types.ts";
+
+// NVIDIA NIM exposes an OpenAI-compatible Chat Completions API, so this provider
+// talks to it with plain `fetch` — no extra SDK dependency. Free hosted endpoints
+// (e.g. Kimi K2.6, DeepSeek V4) are the SDK's default models via the cascade.
+
+export const NVIDIA_BASE_URL = "https://integrate.api.nvidia.com/v1";
+
+// Exact IDs confirmed against GET /v1/models.
+export const KIMI_K2_6 = "moonshotai/kimi-k2.6"; // 256K ctx, tools, agentic — text-only on NIM (verified)
+export const DEEPSEEK_V4_PRO = "deepseek-ai/deepseek-v4-pro"; // 1M ctx, tools, text-only
+export const DEEPSEEK_V4_FLASH = "deepseek-ai/deepseek-v4-flash"; // faster/cheaper, text-only
+export const LLAMA_VISION = "meta/llama-3.2-90b-vision-instruct"; // free NVIDIA vision model (image), verified
+
+export interface NvidiaOptions {
+  model?: string;
+  apiKey?: string;
+  baseURL?: string;
+  maxTokens?: number;
+  temperature?: number;
+}
+
+export class NvidiaProvider implements ModelProvider {
+  readonly name = "nvidia";
+  private model: string;
+  private apiKey?: string;
+  private baseURL: string;
+  private maxTokens: number;
+  private temperature: number;
+
+  constructor(opts: NvidiaOptions = {}) {
+    this.model = opts.model ?? KIMI_K2_6;
+    this.apiKey = opts.apiKey ?? process.env.NVIDIA_API_KEY;
+    this.baseURL = opts.baseURL ?? NVIDIA_BASE_URL;
+    this.maxTokens = opts.maxTokens ?? 2048;
+    this.temperature = opts.temperature ?? 0.2; // low default for deterministic agent behavior
+  }
+
+  async generate(req: ModelRequest): Promise<ModelResponse> {
+    if (!this.apiKey) throw new Error("NvidiaProvider needs NVIDIA_API_KEY (nvapi-...).");
+
+    const body: Record<string, unknown> = {
+      model: this.model,
+      messages: toOpenAIMessages(req),
+      max_tokens: this.maxTokens,
+      temperature: this.temperature,
+    };
+    if (req.tools.length) {
+      body.tools = req.tools.map((t) => ({
+        type: "function",
+        function: { name: t.name, description: t.description, parameters: t.inputSchema },
+      }));
+      body.tool_choice = "auto";
+    }
+
+    const res = await fetch(`${this.baseURL}/chat/completions`, {
+      method: "POST",
+      headers: {
+        Authorization: `Bearer ${this.apiKey}`,
+        "Content-Type": "application/json",
+        Accept: "application/json",
+      },
+      body: JSON.stringify(body),
+    });
+
+    if (!res.ok) {
+      throw new Error(`NVIDIA ${this.model} ${res.status}: ${(await res.text()).slice(0, 300)}`);
+    }
+
+    const json: any = await res.json();
+    const choice = json.choices?.[0];
+    const msg = choice?.message ?? {};
+
+    const content: ContentBlock[] = [];
+    if (typeof msg.content === "string" && msg.content.trim()) {
+      content.push({ type: "text", text: msg.content });
+    }
+    const toolCalls: any[] = msg.tool_calls ?? [];
+    for (const tc of toolCalls) {
+      content.push({
+        type: "tool_use",
+        id: tc.id ?? "",
+        name: tc.function?.name ?? "",
+        input: safeParse(tc.function?.arguments),
+      });
+    }
+    if (content.length === 0) content.push({ type: "text", text: "" });
+
+    const stopReason =
+      choice?.finish_reason === "tool_calls" || toolCalls.length ? "tool_use" : "end_turn";
+    return { content, stopReason };
+  }
+}
+
+function safeParse(args: unknown): Record<string, unknown> {
+  if (typeof args !== "string") return (args as Record<string, unknown>) ?? {};
+  try {
+    return JSON.parse(args);
+  } catch {
+    return {};
+  }
+}
+
+/** Map our Messages into OpenAI-style chat messages (tool calls + tool results + media). */
+function toOpenAIMessages(req: ModelRequest): any[] {
+  const out: any[] = [];
+  if (req.system) out.push({ role: "system", content: req.system });
+
+  for (const m of req.messages) {
+    if (m.role === "user") {
+      // tool_result blocks become individual {role:"tool"} messages
+      for (const b of m.content) {
+        if (b.type === "tool_result") {
+          out.push({ role: "tool", tool_call_id: b.toolUseId, content: b.content });
+        }
+      }
+      const parts = m.content.filter(
+        (b) => b.type === "text" || b.type === "image" || b.type === "video",
+      );
+      if (parts.length) {
+        const multimodal = parts.some((b) => b.type !== "text");
+        out.push({
+          role: "user",
+          content: multimodal
+            ? parts.map(toOpenAIPart)
+            : parts.map((b) => (b as { text: string }).text).join("\n"),
+        });
+      }
+    } else {
+      // assistant
+      const text = m.content
+        .filter((b) => b.type === "text")
+        .map((b) => (b as { text: string }).text)
+        .join("\n");
+      const toolUses = m.content.filter((b) => b.type === "tool_use");
+      const msg: any = { role: "assistant", content: text || null };
+      if (toolUses.length) {
+        msg.tool_calls = toolUses.map((b: any) => ({
+          id: b.id,
+          type: "function",
+          function: { name: b.name, arguments: JSON.stringify(b.input) },
+        }));
+      }
+      out.push(msg);
+    }
+  }
+  return out;
+}
+
+function toOpenAIPart(b: ContentBlock): any {
+  if (b.type === "text") return { type: "text", text: b.text };
+  const media = b as MediaBlock;
+  const url =
+    media.url ??
+    (media.data ? `data:${media.mimeType ?? "application/octet-stream"};base64,${media.data}` : "");
+  if (media.type === "video") return { type: "video_url", video_url: { url } }; // experimental
+  return { type: "image_url", image_url: { url } };
+}

package/src/types.ts

@@ -22,7 +22,23 @@ export interface ToolResultBlock {
   content: string;
   isError?: boolean;
 }
-export type ContentBlock = TextBlock | ToolUseBlock | ToolResultBlock;
+/**
+ * Multimodal input. Provide either a remote `url` or base64 `data` (+ `mimeType`).
+ * Image is broadly supported; video is experimental and model-dependent (routed
+ * to a video-capable model like Kimi K2.6).
+ */
+export interface MediaBlock {
+  type: "image" | "video";
+  url?: string; // remote URL or a data: URL
+  data?: string; // raw base64 (paired with mimeType)
+  mimeType?: string; // e.g. "image/png", "image/jpeg", "video/mp4"
+}
+export type ContentBlock = TextBlock | ToolUseBlock | ToolResultBlock | MediaBlock;
+
+/** True if a message list carries any image/video content (drives vision routing). */
+export function hasMedia(messages: Message[]): boolean {
+  return messages.some((m) => m.content.some((b) => b.type === "image" || b.type === "video"));
+}
 
 export interface Message {
   role: Role;