Aliyun Qwen Multimodal Embedding

Use when multimodal embeddings are needed from Alibaba Cloud Model Studio models such as `qwen3-vl-embedding` for image, video, and text retrieval, cross-mod...

Security Scan

The skill's code and runtime behavior are simple and offline (it only writes a JSON payload), but the README asks for an API key and pairing with external vector stores that the code never uses — an incoherence worth clarifying before install.

Purpose & Capability

Name/description claim: generate multimodal embedding requests for Alibaba Cloud Model Studio. The included Python script exactly matches that purpose (it builds/writes a request JSON and does not call any network services). However SKILL.md's 'Prerequisites' asks the user to set DASHSCOPE_API_KEY or add credentials to ~/.alibabacloud/credentials and to 'pair this skill with a vector store' — none of which are used by the script. This mismatch looks like copy-paste or over-broad documentation and should be explained by the author.

Instruction Scope

Runtime instructions contain references to environment credentials (DASHSCOPE_API_KEY and ~/.alibabacloud/credentials) and advice to stage files in object storage, but the runtime artifact (scripts/prepare_multimodal_embedding_request.py) only composes JSON and writes to disk. There are no commands that read credentials, call network endpoints, or transmit data. The documentation thus grants broader scope than the code actually performs.

Install Mechanism

This is an instruction-only skill with one small Python helper script and no install spec or remote downloads. No packages are fetched and nothing is written to system-wide locations during install — low install risk.

Credentials

No required env vars or primary credential are declared in registry metadata, but SKILL.md requests DASHSCOPE_API_KEY or an entry in ~/.alibabacloud/credentials. Because the code does not use these, the request for credentials is disproportionate and unexplained. If the skill will later be extended to call cloud APIs, requiring credentials would make sense — but as-is, asking for them is unnecessary and raises the risk of accidental credential exposure.

Persistence & Privilege

The skill does not request always: true and has no install actions that modify other skills or system config. It has normal, limited presence (a single helper script) and no special privileges.

What to consider before installing

The code only prepares and writes a JSON request for Alibaba Cloud multimodal embeddings and does not call any network services. However, the documentation asks you to set DASHSCOPE_API_KEY or add credentials to ~/.alibabacloud/credentials and to pair with a vector store — neither is used by the included script. Before installing or providing credentials: (1) Ask the publisher why an API key is mentioned and whether the skill will ever make requests on your behalf; (2) If you don't need networked calls, do not supply credentials — keep testing in a sandbox; (3) If the skill will be extended to call cloud services, provide a least-privilege key scoped only to the needed API and store it in a secure secret store; (4) Run the included validation (python -m py_compile ...) and inspect any changes the skill makes locally. The inconsistency is likely benign copy-paste, but clarify with the author before supplying secrets or chaining this into an automated pipeline.

Category: provider

Model Studio Multimodal Embedding

Validation

mkdir -p output/aliyun-qwen-multimodal-embedding
python -m py_compile skills/ai/search/aliyun-qwen-multimodal-embedding/scripts/prepare_multimodal_embedding_request.py && echo "py_compile_ok" > output/aliyun-qwen-multimodal-embedding/validate.txt

Pass criteria: command exits 0 and output/aliyun-qwen-multimodal-embedding/validate.txt is generated.

Output And Evidence

• Save normalized request payloads, selected dimensions, and sample input references under output/aliyun-qwen-multimodal-embedding/.
• Record the exact model, modality mix, and output vector dimension for reproducibility.

Use this skill when the task needs text, image, or video embeddings from Model Studio for retrieval or similarity workflows.

Critical model names

Use one of these exact model strings as needed:

• qwen3-vl-embedding
• qwen2.5-vl-embedding
• tongyi-embedding-vision-plus-2026-03-06

Selection guidance:

• Prefer qwen3-vl-embedding for the newest multimodal embedding path.
• Use qwen2.5-vl-embedding when you need compatibility with an older deployed pipeline.

Prerequisites

• Set DASHSCOPE_API_KEY in your environment, or add dashscope_api_key to ~/.alibabacloud/credentials.
• Pair this skill with a vector store such as DashVector, OpenSearch, or Milvus when building retrieval systems.

Normalized interface (embedding.multimodal)

Request

• model (string, optional): default qwen3-vl-embedding
• texts (array<string>, optional)
• images (array<string>, optional): public URLs or local paths uploaded by your client layer
• videos (array<string>, optional): public URLs where supported
• dimension (int, optional): e.g. 2560, 2048, 1536, 1024, 768, 512, 256 for qwen3-vl-embedding

Response

• embeddings (array<object>)
• dimension (int)
• usage (object, optional)

Quick start

python skills/ai/search/aliyun-qwen-multimodal-embedding/scripts/prepare_multimodal_embedding_request.py \
  --text "A cat sitting on a red chair" \
  --image "https://example.com/cat.jpg" \
  --dimension 1024

Operational guidance

• Keep input.contents as an array; malformed shapes are a common 400 cause.
• Pin the output dimension to match your index schema before writing vectors.
• Use the same model and dimension across one vector index to avoid mixed-vector incompatibility.
• For large image or video batches, stage files in object storage and reference stable URLs.

Output location

• Default output: output/aliyun-qwen-multimodal-embedding/request.json
• Override base dir with OUTPUT_DIR.

References

• references/sources.md