Gemini Embedding 2 Preview appears to ignore task_type for text and image embeddings

Gemini API
,
1

I believe there is a bug in gemini-embedding-2-preview where task_type has no effect on the returned embedding.

I tested these task types:

  • RETRIEVAL_DOCUMENT
  • RETRIEVAL_QUERY
  • CLASSIFICATION
  • CLUSTERING

For the same input, the returned vectors are exactly identical across all of them.

I verified this in two ways:

  • In a local app storing separate vectors per task type in a vector database
  • By directly calling the Gemini embeddings API with the same input and changing only task_type

Observed behavior

For both text input and image input, the embeddings are numerically identical across all tested task types.

Pairwise comparison results were:

  • max_abs_diff = 0.0
  • cosine similarity = 1.0

Because of that, retrieval rankings and scores are also identical across all task types.

Expected behavior

I expected task_type to affect the embedding output, especially for comparisons like:

  • RETRIEVAL_DOCUMENT vs RETRIEVAL_QUERY
  • RETRIEVAL_QUERY vs CLASSIFICATION
  • RETRIEVAL_QUERY vs CLUSTERING

If task_type is intentionally ignored for gemini-embedding-2-preview, that should be documented clearly. Otherwise this appears to be a backend bug.

Environment

  • Model: gemini-embedding-2-preview
  • API: Gemini API
  • SDK: python-genai
  • Output dimensionality: 3072
  • Date observed: 2026-03-18 UTC

Example text input

silver lever handle on a round rose

Example image input

A generic JPEG product image loaded from a public URL and embedded with the same bytes for each task_type test.

Minimal reproduction

import asyncio
import os
import httpx
import numpy as np
from google import genai
from google.genai import types

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

TASKS = [
    "RETRIEVAL_DOCUMENT",
    "RETRIEVAL_QUERY",
    "CLASSIFICATION",
    "CLUSTERING",
]

def compare_vectors(label, vectors):
    print(label)
    for i, a in enumerate(TASKS):
        for b in TASKS[i + 1:]:
            va = vectors[a]
            vb = vectors[b]
            max_abs_diff = float(np.max(np.abs(va - vb)))
            cosine = float(np.dot(va, vb) / (np.linalg.norm(va) * np.linalg.norm(vb)))
            print(a, "vs", b, "max_abs_diff=", max_abs_diff, "cosine=", cosine)

async def main():
    text = "silver lever handle on a round rose"

    text_vectors = {}
    for task in TASKS:
        result = await client.aio.models.embed_content(
            model="gemini-embedding-2-preview",
            contents=text,
            config=types.EmbedContentConfig(
                task_type=task,
                output_dimensionality=3072,
            ),
        )
        text_vectors[task] = np.array(result.embeddings[0].values, dtype=np.float64)

    compare_vectors("TEXT", text_vectors)

    image_url = "https://example.com/example.jpg"

    async with httpx.AsyncClient(timeout=30.0) as http_client:
        response = await http_client.get(image_url)
        response.raise_for_status()
        image_bytes = response.content

    image_vectors = {}
    for task in TASKS:
        result = await client.aio.models.embed_content(
            model="gemini-embedding-2-preview",
            contents=[
                types.Part.from_bytes(
                    data=image_bytes,
                    mime_type="image/jpeg",
                )
            ],
            config=types.EmbedContentConfig(
                task_type=task,
                output_dimensionality=3072,
            ),
        )
        image_vectors[task] = np.array(result.embeddings[0].values, dtype=np.float64)

    compare_vectors("IMAGE", image_vectors)

asyncio.run(main())

Actual result

For both text and image, every pair of task types produced identical vectors:

  • RETRIEVAL_DOCUMENT vs RETRIEVAL_QUERY
  • RETRIEVAL_DOCUMENT vs CLASSIFICATION
  • RETRIEVAL_DOCUMENT vs CLUSTERING
  • RETRIEVAL_QUERY vs CLASSIFICATION
  • RETRIEVAL_QUERY vs CLUSTERING
  • CLASSIFICATION vs CLUSTERING

Each pair returned:

  • max_abs_diff = 0.0
  • cosine similarity = 1.0

Question

Can you confirm whether task_type is currently expected to change the output for gemini-embedding-2-preview? If yes, this appears to be a bug. If no, the documentation should clarify that task_type is currently ignored for this model.