agentscope-ai/agentscope · error · ValueError

Gemini embedding API requires inline data (Base64Source). UR

Error message

Gemini embedding API requires inline data (Base64Source). URLSource is not directly supported for embedding. Got URL: {source.url}

What it means

Gemini's embed_content API only accepts inline (base64) content, so the wrapper rejects DataBlocks whose source is a URLSource. The message is deliberate: the SDK has no from_url path for embeddings, so callers must inline the bytes themselves.

Source

Thrown at src/agentscope/embedding/_gemini/_model.py:476

        """
        from google.genai import types
        from ...message import Base64Source, URLSource

        source = block.source

        if isinstance(source, Base64Source):
            import base64

            return types.Part.from_bytes(
                data=base64.b64decode(source.data),
                mime_type=source.media_type,
            )

        if isinstance(source, URLSource):
            # Gemini SDK doesn't have a direct from_url for
            # embed_content; download or use File API.
            # For now, raise — callers should use Base64Source.
            raise ValueError(
                "Gemini embedding API requires inline data "
                "(Base64Source). URLSource is not directly supported "
                f"for embedding. Got URL: {source.url}",
            )

        raise ValueError(
            f"Unsupported source type {type(source).__name__} "
            f"in DataBlock.",
        )

View on GitHub (pinned to e90f1c7592)

Solutions

Download the URL content and rebuild the DataBlock with Base64Source (inline bytes)
Migrate assets to the Gemini File API and reference them where supported, or pre-cache bytes locally
Centralize a to_base64_block(url) helper so URL blocks never reach the embedding model

Example fix

# before
block = DataBlock(source=URLSource(url=img_url, media_type="image/png"))
emb = await model(inputs=[block])

# after
import requests, base64
from agentscope.message import DataBlock, Base64Source
raw = requests.get(img_url, timeout=30).content
block = DataBlock(source=Base64Source(data=base64.b64encode(raw).decode(), media_type="image/png"))
emb = await model(inputs=[block])

Defensive patterns

Strategy: fallback

Validate before calling

from agentscope.message import URLSource
for x in inputs:
    src = getattr(x, "source", None)
    if isinstance(src, URLSource):
        raise ValueError("inline URL blocks before calling gemini embedding")  # or convert

Type guard

from agentscope.message import URLSource

def has_url_source(inputs: list) -> bool:
    return any(isinstance(getattr(x, "source", None), URLSource) for x in inputs)

Try / catch

try:
    emb = await model(inputs=inputs)
except ValueError as e:
    if "URLSource is not directly supported" in str(e):
        inputs = [inline_url_block(x) if is_url_block(x) else x for x in inputs]
        emb = await model(inputs=inputs)
    else:
        raise

Prevention

Convert remote assets to Base64Source before embedding
Keep a shared inline_url(url) helper in your codebase
Remember chat formatters fetch URLs but the embedding API cannot

When it happens

Trigger: Building DataBlock(source=URLSource(url="https://...", media_type="image/png")) and passing it to the Gemini embedding model; _call_multimodal -> _data_block_to_part hits the URLSource branch and raises.

Common situations: Reusing URL-based DataBlocks built for chat/LLM formatters (which do fetch URLs) in an embedding call; storing assets in cloud storage (S3/GCS signed URLs) and trying to embed them directly.

Related errors

AI-assisted analysis of agentscope-ai/agentscope@e90f1c7592 (2026-08-28). Data as JSON: /api/errors/f59e310ad40cec73. Report an issue: GitHub.