Getting 500 Errors when using cache

Gemini API
, , ,
1

I’m trying to call Gemini 2.5 Pro to infer over a video in chunks. I’m uploading the video using the files API, and I explicitly cache the system prompts and tool definition.

I receive Server Errors (500) that look like this:
ServerError("500 INTERNAL. {'error': {'code': 500, 'message': 'Internal error encountered.', 'status': 'INTERNAL'}}")

Here’s how my code looks (sort of):


async def run_model(klass: GeminiRequester, file: File, start_time: int, end_time: int, cache_name: str):
	video_msg = Part(
		file_data=FileData(file_uri=file.uri, mime_type=file.mime_type),
		video_metadata=VideoMetadata(fps=1, start_offset=f'{start_time}s', end_offset=f'{end_time}s')
	)
	klass.messages = [video_msg, Part.from_text(text=klass.get_prompt())]
	# parses the raw response to the tool model mirror behind the scenes
	response: ToolModelMirror = await klass.run(cache_name=cache_name)
	return [
		TaggedScene.model_validate({**tag.model_dump(), 'name': character_tagging.name})
		for character_tagging in response.tagged_characters for tag in character_tagging.tagged_utterances
	]


async def main(client, project, file: File, **kwargs):
	system: list[str] = [get_system_prompt(**kwargs), get_metadata_prompt(**kwargs)]
	cache: CachedContent = await client.aio.caches.create(
		model="gemini-2.5-pro",
		config=CreateCachedContentConfig(
			display_name=f'{str(project.id)}-tagging',
			system_instruction=system,
			tools=[tool],
			tool_config=ToolConfig(
				function_calling_config=FunctionCallingConfig(
					mode=FunctionCallingConfigMode.ANY,
					allowed_function_names=[tool.function_declarations[0].name]
				)
			),
			ttl="3000s",
		)
	)
	chunks = get_chunks(project)
	# GeminiRequester is our own abstraction
	requester = GeminiRequester(project)
	async with (BoundedTaskGroup(max_parallelism=5) as tg):
		tasks = [
			tg.create_task(
				run_model(klass=requester, file=file, start_time=chunk[0], end_time=chunk[1], cache_name=cache.name)
			)
			for chunk in chunks
		]

	results = dict()
	for t in tasks:
		results.update(t.result())

	return results

I used to have an implementation of similar code, except I was caching the whole video, and instead of specifying start/end intervals, each message used the entire cached video as input. In that implementation, the code worked with exactly the same system prompts and tool definition as this one. So, I think it might be related to slicing the video with intervals?

2

Hi @Yovel_Cohen,

Please wait and try your request again as a 500 error usually indicates an unexpected error on Google’s side.

Thank you!