@byte_fluffy/nexra-sdk 0.1.0-alpha.46.1 → 0.1.0-alpha.48

package/README.md

@@ -439,6 +439,92 @@ const gpt6Prices = prices.filter(({ modelId }) =>
 );
 ```
 
+## Ark web search (Responses)
+
+Available since SDK `0.1.0-alpha.47` (`npm install @byte_fluffy/nexra-sdk@next`).
+
+DeepSeek V4.1 Flash (`deepseek/deepseek-v4-1-flash`), GLM 5.3 Flash
+(`z-ai/glm-5-3-flash`), Doubao Seed 2.1 Lite
+(`bytedance/doubao-seed-2-1-lite-260915`), and Kimi K3.1 Pro
+(`moonshot-ai/kimi-k3-1-pro`) support the Ark online content plugin through
+`/v1/responses`. The default SDK binding selects this endpoint. All four retain
+explicit Chat Completions bindings; DeepSeek also retains Messages.
+Search is a provider-executed tool, rather than a Chat `web_search_options` flag.
+
+```ts
+import { generateText } from "ai";
+import { openai } from "@ai-sdk/openai";
+import { createNexraLanguageModel } from "@byte_fluffy/nexra-sdk/language";
+
+const result = await generateText({
+  model: createNexraLanguageModel("deepseek/deepseek-v4-1-flash", {
+    apiKey: process.env.NEXRA_API_KEY!,
+    baseUrl: "https://vip-api.nexra-ai.com/v1",
+  }),
+  prompt: "Search for the official DeepSeek website and cite its URL.",
+  tools: { web_search: openai.tools.webSearch() },
+  providerOptions: {
+    openai: { maxToolCalls: 1 },
+    nexra: { web_search: { sources: ["search_engine"], max_keyword: 1, limit: 3 } },
+  },
+  maxOutputTokens: 512,
+  maxRetries: 0,
+});
+console.log(result.text, result.sources);
+```
+
+`providerOptions.nexra.web_search` preserves native tool fields such as `sources`,
+`max_keyword`, `limit`, and `user_location`; it requires the `web_search` tool.
+The same options work with `streamText`. Unpositioned native citations become
+URL sources; original Native bodies, raw stream events, and
+`finalStep.usage.raw.tool_usage_details.web_search` remain available.
+
+VIP configured search prices (CNY, checked 2026-11-01):
+
+| Content source | Channel cost / 1,000 uses | Nexra sale / 1,000 uses |
+| --- | ---: | ---: |
+| `search_engine` (default) | ¥4 | **¥4.4** |
+| `toutiao`, `douyin`, `moji` (each additional source) | ¥6 | ¥6.6 |
+
+A use is an actual upstream content-source search, **not an HTTP request or a
+returned result**. Multiple rounds, keywords, or content sources can incur
+multiple uses. Token charges are additional, including retrieved context.
+Declaring a tool does not itself incur a search fee if the model does not use
+it. Preauthorization reserves the configured search ceiling and settles the
+actual evidenced usage; set `maxToolCalls` and `max_keyword` to control that
+ceiling. Missing search evidence is not treated as free search.
+
+Current VIP token sale prices per million tokens:
+
+| Model | Input | Cached input | Output |
+| --- | ---: | ---: | ---: |
+| DeepSeek V4.1 Flash | ¥2.2 | ¥0.044 | ¥8.8 |
+| GLM 5.3 Flash | ¥0.88 | ¥0.253 | ¥3.08 |
+| Kimi K3.1 Pro | ¥14.3 | ¥1.43 | ¥74.8 |
+| Doubao Seed 2.1 Lite (non-audio input) | ¥0.88 | ¥0.176 | ¥2.97 |
+
+Seed audio input is ¥13.2/million tokens; cached audio input is ¥2.64/million.
+For Ark audio/video prompts in the AI SDK adapter, select
+`protocol: "openai-chat-completions"`; its native media parts retain their order. Kimi media requires bytes or Base64
+data URLs; the Chat binding derives URL support from the Native schema.
+The Responses search example above uses text. Native Responses media fields
+remain documented by each model's selected endpoint schema.
+
+GLM and Kimi K3.1 Pro thinking is always on. Use `reasoning: "low"` for minimum effort;
+`reasoning: "none"` fails locally. Native `none`/`minimal` effort aliases do not
+mean that these models can turn thinking off. Kimi does not use a `thinking`
+parameter. Its Responses endpoint permits `toolChoice: "auto"` (default) or
+`"none"`; forcing a particular tool or requiring a tool is rejected before
+fetch. With auto selection, search is billed only when actually used.
+
+These are the configured channel prices;
+the SDK does not dynamically apply upstream off-peak discounts. Provider account
+free quotas do not automatically make Nexra requests free. The selected plugin
+is separate from Doubao Search Custom; no Custom search service is used here.
+Always read the selected environment's live catalog prices before estimating
+charges. Official references: [online content plugin pricing](https://ark.volcengine.com/region:cn-beijing/docs/online-content-plugin-pricing)
+and [model pricing](https://console.volcengine.com/ark/region:cn-beijing/docs/model-pricing).
+
 ## Directory layout
 
 ```

package/dist/catalog.js

@@ -9,7 +9,7 @@ import {
   listModelSpecs,
   projectMediaGeometry,
   resolveNexraLanguageReasoning
-} from "./chunk-JK6LUGM5.js";
+} from "./chunk-2IWCEEDI.js";
 import "./chunk-NR4EWD3X.js";
 export {
   CatalogClient,