Add HAKARI-Bench results for google/gemini-embedding-2

Summary

DuckDB Nano-set Comparison

Computed from DuckDB task_results with the same Overall grouping as this PR body. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Overall nDCG@10

Reproducibility

Command

uv run hakari-bench batch dense register \
  --target gemini-embedding-2-nano-all \
  --provider gemini \
  --model gemini-embedding-2 \
  --model-alias google/gemini-embedding-2 \
  --all \
  --batch-size 32 \
  --max-input-tokens 8100 \
  --results-dir output/gemini-developer-batch-nano-all \
  --result-format json.xz

uv run hakari-bench batch dense process \
  --target gemini-embedding-2-nano-all \
  --embedding-variant truncate:128 \
  --embedding-variant truncate:256 \
  --embedding-variant truncate:512 \
  --embedding-variant truncate:768 \
  --embedding-variant truncate:1536 \
  --candidate-ranking reranking_hybrid \
  --batch-size 32 \
  --results-dir output/gemini-developer-batch-nano-all \
  --result-format json.xz

Submitter Notes

• Evaluated with the Gemini Developer API embedding batch path. Provider request text was capped to 8,100 tokens to stay below the 8,192 token model limit.
• Gemini query/document request prefixes were recorded in result metadata as task: search result | query: and title: none | text: .
• Result variants include the base 3,072-dimensional embeddings, truncation dimensions 128/256/512/768/1536, full-dimension int8/binary variants, rescore variants, and truncation x quantized/rescore variants. Truncated embeddings were produced from the full embedding prefix followed by L2 normalization.
• Batch jobs were submitted and resumed by Nano-set task, with large tasks split into smaller provider batch jobs. Some early oversized/stale attempts were cancelled and resubmitted; the final submitted set completed successfully with 4,229 succeeded jobs, 551 materialized task result files, and no remaining submit or materialization errors.
• These are standard built-in --all results covering 551 result files / 538 raw task results in the Overall grouping.
• Gemini is a provider API model, so an immutable model revision is not recorded in the result JSON.

Checklist

• Result files are committed under hakari-results/google__gemini-embedding-2/.
• Result files are compressed .json.xz; no caches, DuckDB files, HTML reports, or local scratch artifacts are included.
• The result JSON records dataset revision, runtime configuration, package versions, and provider/model metadata. Immutable model revision is not recorded because this is a hosted Gemini API model.
• Overall nDCG@10 above was generated from the submitted result files.
• Any non-default prompt, sequence length, attention implementation, candidate ranking, or reranker setting is documented above.