@agentor/dashscope weekly downloads statistics

@agentor/dashscope

English | 中文

AI SDK provider for Alibaba Cloud DashScope (Bailian) API.

Features

• Chat Completions - /chat/completions with function calling, streaming, reasoning, and vision • Completions (FIM) - /completions for code completion with Qwen Coder models • Responses - /responses endpoint with built-in tools (web search, code interpreter, MCP, etc.) • Embedding - Text vectorization via /embeddings endpoint • Reranking - Document reranking via /reranks endpoint • Image Generation - Text-to-image via multimodal generation endpoint • Video Generation - Text-to-video and image-to-video with async polling • Speech Synthesis - Text-to-speech for CosyVoice and Qwen-TTS models • Transcription - Speech-to-text for short and long audio • Built-in Tools - Web search, code interpreter, web extractor, file search, web search image, image search, MCP integration • Thinking Mode - Enable reasoning/thinking with configurable budget • Multi-region - Beijing, Singapore, US, Germany regions • AI SDK V4 - Implements the LanguageModelV4 specification • TypeScript-First - Full type safety with comprehensive TypeScript support

Installation

Quick Start

Setup

Or use the default instance (reads DASHSCOPE_API_KEY from environment):

Basic Chat

Streaming

Function Calling

Chat Completions API

Web Search

Enable web search via providerOptions:

Options: enableSearch, searchStrategy ("enable" | "enable_with_history" | "agent_max").

Code Interpreter

Enable code interpreter (requires thinking mode):

Thinking Mode

Enable reasoning with configurable budget:

Context Cache

Applies to the Chat endpoint (dashscope(modelId)). The Responses endpoint (dashscope.responses(modelId)) caches via a session header instead — see Session Cache.

DashScope supports explicit caching to reduce cost and latency for repeated prefixes. Add cacheControl via providerOptions on messages or content parts:

Cache on user message content parts:

Implicit caching is enabled automatically for supported models — no configuration needed.

Combining with structured output

By default the schema is sent via response_format (json_schema) and never enters the messages, so a cacheControl marker anywhere — system or user message — keeps hitting across calls regardless of how the schema changes.

Only models that reject json_schema fall back to json_object + prompt injection. In that fallback the schema is injected as a user message after the system block, so it stays outside the system cache prefix and a system-side cacheControl marker still hits. Qwen3.5+ merges every system message into one cache segment and only honors a cache_control marker at the segment tail, so the schema is never injected as a system message. Prefer placing cacheControl on the system message; on user content parts in fallback mode, the injected schema lands before that content and may prevent it from being cached.

JSON Output

Structured Output with Schema

Use generateText with Output.object() to generate typed JSON:

The provider sends the schema via native response_format: { type: "json_schema" } (strict). The schema is enforced server-side and travels in response_format — it never enters the messages, so a cacheControl marker anywhere keeps hitting even as the schema varies per call. Models that reject json_schema fall back automatically to json_object + prompt injection (best-effort; validate the output).

Supported models

Structured output is supported across the Qwen family and partner models, including qwen3.7-max, qwen3-max, qwen3.7-plus, qwen-plus, qwen3.7-flash, qwen3.5-flash, qwen-flash, qwen-turbo, qwen3-coder, qwen-long, the Qwen-VL models, and partner models (Kimi, DeepSeek, GLM, Stepfun). See the official guide for the full, up-to-date list.

Thinking mode

Models marked "non-thinking" may return non-strict JSON under thinking mode. For reliable structured output, avoid combining enableThinking with Output.object, or use the two-step repair: generate with the thinking model first, then normalize the result through a fast non-thinking model (e.g. qwen-flash) with Output.object.

Avoid maxOutputTokens

Do not set maxOutputTokens (max_tokens) with structured output — capping output tokens can truncate the JSON mid-stream and break parsing.

Completions (FIM)

Use completionModel() for text/code completion via the /completions endpoint (Fill-In-the-Middle):

Responses API

Use the responses namespace for the Responses API with built-in tools:

Built-in Tools

Web Search

Options: forcedSearch, searchStrategy ("enable" | "enable_with_history").

Code Interpreter

Web Extractor

Can be used together with webSearch for enhanced results.

File Search

MCP Integration

Options: serverProtocol, serverLabel, serverUrl, serverDescription, headers.

Web Search Image

Text-to-image search. Search images based on a text description.

Image Search

Image-to-image search. Search similar images based on an input image.

Multi-turn Conversation

Options via providerOptions.dashscope: previousResponseId, enableThinking, reasoning (effort levels), conversation, instructions, includeUsage.

Use previousResponseId for multi-turn conversations:

Session Cache

The Responses endpoint caches via server-side session cache, not cache_control markers. Setting cacheControl on any message automatically sends the x-dashscope-session-cache: enable header — combine it with previousResponseId (above) to carry context across turns and hit the cache (minimum 1024 tokens, 5-minute TTL).

OCR (Qwen-OCR)

Qwen-OCR models (qwen3.5-ocr, qwen-vl-ocr, qwen-vl-ocr-latest) extract text and structured data from images and PDFs. Both endpoints accept providerOptions.dashscope.ocrOptions; the Responses endpoint officially honors the built-in tasks (and is the only one that accepts PDF). On Chat, pass the image as a file part and drive extraction through the text prompt (data is a Uint8Array):

On the Responses endpoint, set a built-in task via ocrOptions.task:…