Gemini-embedding-001: Native title param vs manual prepend for RETRIEVAL_DOCUMENT

Gemini API
,
1

I’m building a RAG pipeline with gemini-embedding-001 and RETRIEVAL_DOCUMENT task type. My documents are chunked, and each chunk has a short title (1-5 words).

I have two approaches and would appreciate clarity on which is optimal:

1. Native title parameter:

config=types.EmbedContentConfig(
    task_type="RETRIEVAL_DOCUMENT",
    title=content_title,
    output_dimensionality=768,
)
contents = [chunk_text]

2. Manual prepend:

I don’t feed in title to the config and have the following

config=types.EmbedContentConfig(
    task_type="RETRIEVAL_DOCUMENT",
    output_dimensionality=768,
)
contents = [f"Title: {title}\nContent: {chunk_text}"]

Questions:

  • Does the native title parameter handle the title differently from simply prepending it to the content string?

  • For short titles (1-5 words), is the native param meaningfully better than manual prepend?