HKUDS/DeepTutor · error · ValueError

Cohere v2 does not support content type '{kind}'

Error message

Cohere v2 does not support content type '{kind}'

What it means

While mapping multimodal contents to Cohere v2 inputs, the adapter handles only 'text' and 'image' kinds; any other content kind (audio, video, generic 'file') raises with the offending kind. Cohere's embed-v4.0 input schema accepts text and image_url parts only.

Source

Thrown at deeptutor/services/embedding/adapters/cohere.py:114

            if request.contents:
                # Cohere v2 multimodal: `inputs: [{content: [{type, text|image_url}]}]`
                # We translate the simple [{text|image|video}] contract into v2's
                # nested form. v2 cannot mix text+image in one input, so each
                # content dict becomes its own input item.
                inputs = []
                for item in request.contents:
                    if not isinstance(item, dict):
                        continue
                    kind, value = next(iter(item.items()))
                    if kind == "text":
                        inputs.append({"content": [{"type": "text", "text": value}]})
                    elif kind == "image":
                        inputs.append(
                            {"content": [{"type": "image_url", "image_url": {"url": value}}]}
                        )
                    else:
                        raise ValueError(f"Cohere v2 does not support content type '{kind}'")
                payload["inputs"] = inputs
            else:
                payload["texts"] = request.texts

            supported_dims = model_info.get("dimensions", [])
            if isinstance(supported_dims, list) and len(supported_dims) > 1:
                payload["output_dimension"] = dimension or model_info.get("default")

            if not request.truncate:
                payload["truncate"] = "NONE"

        url = self.base_url

        logger.debug(f"Sending embedding request to {url} with {len(request.texts)} texts")

        async with httpx.AsyncClient(
            timeout=self.request_timeout, verify=not disable_ssl_verify_enabled()
        ) as client:

View on GitHub (pinned to 3e82f13042)

Solutions

1. Filter contents to text/image parts before calling the Cohere adapter
2. Route non-text/image modalities to a provider whose SDK supports them (e.g. DashScope multimodal)
3. Reject unsupported kinds at request-build time with a clear error instead of deep in the adapter

Example fix

# before
contents = all_parts  # includes {"kind": "audio", ...}
await cohere.embed(EmbeddingRequest(contents=contents))
# after
supported = [p for p in all_parts if p["kind"] in ("text", "image")]
await cohere.embed(EmbeddingRequest(contents=supported))

ALLOWED = {"text", "image"}
contents = [p for p in contents if p["kind"] in ALLOWED]

def is_cohere_content(part: dict) -> bool:
    return part.get("kind") in ("text", "image")

try:
    await adapter.embed(req)
except ValueError as e:
    if "does not support content type" in str(e):
        req = replace(req, contents=[p for p in req.contents if p["kind"] in ("text", "image")])
        return await adapter.embed(req)
    raise

When it happens

Trigger: EmbeddingRequest(contents=[{"kind": "audio", ...}]) or kind='file'/'video' passed to the Cohere v2 adapter's embed().

Common situations: Sharing a generic multimodal request builder across providers where another provider accepts audio; upstream pipelines adding new content kinds that reach the Cohere adapter.

Related errors

• Cohere v1 API does not support multimodal `contents`. Use em
• Cohere model '{model_name}' does not support multimodal `con
• OpenAI-compatible embedding model '{model}' does not support
• openai_sdk adapter does not support multimodal `contents`. P
• dashscope SDK not installed. Run `pip install dashscope` (or

AI-assisted analysis of HKUDS/DeepTutor@3e82f13042 (2026-08-27). Data as JSON: /api/errors/3234fe6ac65205a7. Report an issue: GitHub.