OpenViking/benchmark/longmemeval/openviking

LongMemEval OpenViking Benchmark

This directory contains the OpenViking evaluation flow for LongMemEval:

1. import each user's haystack sessions into OpenViking;
2. run one retrieval call per question;
3. optionally rerank the retrieved memories;
4. answer from the selected memory context;
5. judge and summarize the result CSV.

The benchmark expects an OpenViking server to already be running. The commands below use the default local OpenViking client configuration. If you need a different endpoint, pass --openviking-url to both import and eval.

Data

Set the dataset path once:

DATA=/path/to/longmemeval_s_cleaned.json

The importer uses each sample's original question_id-derived user id, such as lm_user_<id>, so the eval script searches the same user space created during import.

Import

python benchmark/longmemeval/openviking/import_to_ov.py \
  --input "$DATA" \
  --parallel 16 \
  --submit-parallel 16 \
  --wait-mode deferred \
  --success-csv result/longmemeval_openviking_import_success.csv \
  --error-log result/longmemeval_openviking_import_errors.log

Use --force-ingest only when intentionally re-importing existing sessions.

Smoke Test

Run one question before a full evaluation:

python benchmark/longmemeval/openviking/run_eval.py "$DATA" \
  --output result/longmemeval_openviking_smoke.csv \
  --count 1 \
  --threads 1 \
  --single-search-context-limit 50 \
  --single-search-rerank-limit 10 \
  --single-search-max-context-chars 30000 \
  --debug-print-model-input

With --debug-print-model-input, the CSV includes the full answer prompt and retrieval trace. Check retrieved_uris_by_iteration to confirm rerank is enabled and context_uris contains the expected number of memories.

Full Eval

This example retrieves 50 memories, keeps the top 10 after rerank, and caps the memory text passed to the answer model at 30000 characters.

OUT=result/longmemeval_openviking_search50_rerank10_chars30000.csv

python benchmark/longmemeval/openviking/run_eval.py "$DATA" \
  --output "$OUT" \
  --threads 8 \
  --timeout 900 \
  --single-search-context-limit 50 \
  --single-search-rerank-limit 10 \
  --single-search-max-context-chars 30000

Set --single-search-rerank-limit 0 to disable rerank. Set --single-search-max-context-chars 0 to disable the character budget.

Judge And Stat

python benchmark/longmemeval/openviking/judge.py \
  --input "$OUT" \
  --parallel 40

python benchmark/longmemeval/openviking/stat_judge_result.py \
  --input "$OUT"

The judge writes result values back into the same CSV. Use --force to re-grade rows that already have a result. Use --strict-prompt when you want the stricter LongMemEval judge prompt instead of the default lenient prompt.

stat_judge_result.py prints overall accuracy, average memory token/character usage, and accuracy grouped by LongMemEval question type.