--- title: "Doubao Integration" description: "Call ByteDance's Doubao model family through the Volcano Engine Ark data-plane API. Agents run Chinese-language generation and embeddings on a platform that already meets local data requirements." url: https://flowrunner.ai/integrations/doubao date_modified: 2026-08-07T17:20:23-07:00 --- # Doubao [AI](https://flowrunner.ai/integrations/category/ai-llms) Call ByteDance's Doubao model family through the Volcano Engine Ark data-plane API. Agents run Chinese-language generation and embeddings on a platform that already meets local data requirements. [Verified](https://flowrunner.ai/integrations/verified "What does verified mean?") · 8 actions · API key · available [Doubao website](https://ark.cn-beijing.volces.com/) · [Platform Documentation](https://www.volcengine.com/docs/82379) · Capability data verified 2026-08-12 1. A seasonal campaign is scheduled and the product list is finalised for the China storefront 2. Agent pulls each product's specification and existing photography from the catalogue 3. Agent runs Tokenize Text on the assembled prompts to confirm they fit the model's context window and to price the batch before anything is submitted 4. Agent runs Generate Image with the product photo passed as a reference image, producing a related set in one call 5. Agent posts the image set to the campaign channel with the prompt and the 24 hour expiry on each URL called out 6. The campaign owner authorises the video generation batch, which cannot be cancelled once submitted and whose output expires the same day ## What This Integration Enables Doubao is ByteDance's model family, and the surface you actually build against is Volcano Engine Ark. That distinction is the whole reason this connector exists: Ark is a cloud data plane, not a chat product, and it behaves accordingly. There is no API-key-authenticated models endpoint, so the model field is free text taking either a Model ID or an Endpoint ID, and every model has to be activated in the Ark console before a call to it will work. Nothing here is discoverable at runtime. If your audience is in China and your content is Chinese, this is the connector you want. If you are looking for a drop-in replacement for a Western provider, the operational shape will surprise you, and it is better to learn that here than from a failed run. What Ark gives an agent is genuinely broad. Chat Completion exposes full control over sampling, deep-thinking mode, reasoning effort, stop words, tool calling and JSON Schema structured output, while Send Prompt covers the one-shot case without assembling a message array. Analyze Images sends photos to a vision model for captioning, OCR-style extraction and visual question answering, with an analysis granularity setting that trades cost against detail. Create Embeddings handles Chinese and English text for vector search. Seedream generates images guided by up to fourteen reference images, and Seedance produces short video as an asynchronous task you submit and poll. Two constraints shape every flow built on this: streaming is never used, so each call returns one complete response, and generated image and video URLs expire twenty four hours after creation. Design for the expiry. It is the difference between an asset library and a folder of dead links. ### Without FlowRunner **Creative volume capped by studio time**: Every localised variant waits for a shoot or a designer **Cost discovered after the run**: Nobody knows what a batch costs until the invoice arrives **Generated assets go stale unnoticed**: A batch produced overnight is found the next afternoon, past its download window ### With FlowRunner **Variants generated from the product photo**: Reference images guide the model, so output stays recognisably the product **Cost checked before submission**: Token counts are measured and priced while the batch is still cancellable **Expiry treated as a deadline**: Assets are downloaded and stored the moment they land, and the batch is timed around a person being awake ## Use Case Scenarios ### Product photos read into structured fields Warehouse or supplier photos arrive without usable metadata. The agent passes each image to Analyze Images with a question that names the fields it needs, using the higher analysis granularity where small print matters, and gets back model, colourway, dimensions and any visible defect as text. Chat Completion with JSON Schema structured output turns that into a clean record, which is written to [Airtable](https://flowrunner.ai/integrations/airtable) or straight onto the [Magento](https://flowrunner.ai/integrations/magento) product. Photos the model reads with low confidence are separated as [automation exceptions](https://flowrunner.ai/concepts/automation-exceptions) and queued for a person rather than filed with a guess. ### Chinese-language support answers grounded in your own documentation Create Embeddings turns the Chinese product documentation into vectors, batched around four texts per call for throughput and kept inside the model's input limit, and the vectors go to [Qdrant](https://flowrunner.ai/integrations/qdrant). When a question arrives, the agent embeds it, retrieves the matching passages, and answers with Chat Completion in deep-thinking mode for anything the retrieval step returned weakly. Tokenize Text runs first on long support threads to confirm the assembled context fits the window, which turns a class of runtime failures into a pre-flight check. ### Campaign creative produced as a batch, released as a decision A campaign is approved. The agent generates the still assets with Generate Image, passing the existing product photograph as a reference so the output stays recognisably the product, and using image-set mode to get a related series in one call rather than four uncorrelated ones. The set is posted for review. Only the video, which is the expensive and slow half, waits for authorisation. When it comes, the agent submits Create Video Generation Task, polls Get Video Generation Task until the status reads succeeded, and immediately copies the MP4 to [S3](https://flowrunner.ai/integrations/s3) before the twenty four hour URL lapses. ## Human-in-Loop Highlight Ark's video pipeline is a one-way door, so the pause on this connector sits before the spend rather than after the output. Create Video Generation Task queues an asynchronous job and returns a task ID. The status vocabulary includes cancelled, but the connector exposes no cancel action, so once a batch of Seedance jobs is submitted, it runs. Add the twenty four hour expiry on the resulting MP4 URLs and an unattended overnight batch can be both fully paid for and fully worthless by the time anyone opens the channel. So the agent does everything up to the submit and then stops. It assembles the prompts, resolves the reference images, and prices the batch from the token counts it measured with Tokenize Text. Then it posts to the campaign owner: "Twelve Seedance clips ready to submit at the resolution and duration below. These cannot be cancelled once queued, and the output URLs expire twenty four hours after they complete. Submit now, submit at 08:00 local so somebody is here to collect them, or cut the list?" The owner answers, and the batch runs on their timing. That pause costs a minute and is the difference between [human-in-the-loop](https://flowrunner.ai/concepts/human-in-the-loop) and an expensive log entry. Agent processes routinely Detects exception requiring judgment Clear match Continues automatically Ambiguous Routes to human via preferred channel Human decides Agent resumes with decision ## Agent Capabilities 8 actions ### Chat 3 - **Chat Completion** Sends a full message list to a Doubao model on Ark and returns the next assistant message. Gives complete control over sampling, deep-thinking mode, reasoning effort, stop words, tool calling and JSON Schema structured output. Streaming is never used, so the whole reply comes back in one response. - **Send Prompt** Sends a single prompt and returns just the generated text, with no message array to assemble. Accepts an optional system instruction that sets the model's role and constraints, and exposes deep-thinking control. The reasoning trace, where the model produced one, is returned separately from the answer. - **Analyze Images** Sends one or more images with a question to a Doubao vision model and returns a textual answer. Accepts public image URLs or data URIs, and trades cost against detail through the analysis granularity setting. Used for captioning, OCR-style extraction, quality checks and visual question answering. ### Embeddings 1 - **Create Embeddings** Converts Chinese or English texts into semantic vectors for vector search, clustering and retrieval flows. Each text stays within the model's input limit, and batching around four texts per call gives the best throughput. ### Images 1 - **Generate Image** Generates images from a text prompt with a Seedream model, optionally guided by up to fourteen reference images for image-to-image work and multi-image fusion. Supports resolution presets or explicit pixel dimensions, an image-set mode returning a related series in one call, prompt optimisation, and an optional AI-generated watermark. Generated URLs expire twenty four hours after creation. ### Video 2 - **Create Video Generation Task** Submits an asynchronous video generation job to a Seedance model from a text prompt and, optionally, a first frame, a last frame or up to nine reference images. Returns a task ID immediately. Resolution, aspect ratio and duration ranges vary by model, and the resulting video URL expires after twenty four hours. - **Get Video Generation Task** Retrieves the current state of a video generation task and, once the status is succeeded, the MP4 URL plus the optional last-frame image. Statuses are queued, running, succeeded, failed, cancelled and expired. Only tasks from the last seven days can be queried. ### Utilities 1 - **Tokenize Text** Splits text into the tokens a Doubao model actually sees, returning the token count, the token IDs from the model vocabulary and the character offsets of each token. Used to estimate cost, to confirm a prompt fits the context window before sending it, or to build token IDs for a logit bias setting. ## Frequently Asked Questions ### What can FlowRunner do with Doubao? FlowRunner agents can run Chat Completion, Send Prompt, and Analyze Images in Doubao, plus 5 more actions. ### Does connecting Doubao to FlowRunner require OAuth? No. Doubao connects to FlowRunner with an API key, no OAuth flow required. ### Can Doubao trigger a FlowRunner workflow automatically? Doubao doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger. **Work at Doubao?** This integration exposes Doubao to AI agents on every FlowRunner plan, including through MCP, at no cost to you. [See what FlowRunner offers integration partners](https://flowrunner.ai/integrations/partners), including how to keep this page current. --- Markdown version of https://flowrunner.ai/integrations/doubao. Site index: https://flowrunner.ai/llms.txt