Datasets:

p-research
/

oreolook-research-evals

metric_id stringclasses 3 values	metric stringclasses 3 values	value float64 0.1 89.3	unit stringclasses 3 values	evaluation_scope stringclasses 1 value	environment stringclasses 2 values	source stringclasses 1 value	reported_at stringdate 2026-01-01 00:00:00 2026-01-01 00:00:00	notes stringclasses 3 values
paper-redis-keyspace-hit-rate	aggregate_redis_keyspace_hit_rate	89.3	percent	historical_production_snapshot	Single 8-vCPU Intel Cascade Lake server at 2 GHz with 32 GB RAM; 30 Hypercorn workers across three containerized replicas	https://arxiv.org/abs/2609.05463	2026	Aggregate Redis-level keyspace rate across internal operations; not a query-level semantic-cache hit rate and not an estimate of avoided model inference.
paper-redis-read-latency	redis_read_latency	0.1	milliseconds	historical_production_snapshot	Single 8-vCPU Intel Cascade Lake server at 2 GHz with 32 GB RAM; 30 Hypercorn workers across three containerized replicas	https://arxiv.org/abs/2609.05463	2026	Paper-reported Redis read latency from the same historical production snapshot; results can vary with hardware and workload.
paper-redis-memory-footprint	redis_memory_footprint	1.38	megabytes	historical_production_snapshot	Redis 7.4 in a dedicated container capped at 2 GB on the evaluated host	https://arxiv.org/abs/2609.05463	2026	Total Redis memory footprint reported after six days of uptime in the evaluated snapshot.

OreoLook Research Evaluations

A small, inspectable evaluation suite for current-information search agents. It covers routing, clarification, conversational continuity, citation discipline, freshness, PDF artifacts, protocol safety, and semantic-cache equivalence.

This repository accompanies:

Configuration	Rows	Purpose
`evaluations`	24	Synthetic user turns and expected agent behavior
`cache_pairs`	12	Paraphrase-equivalence and cache-isolation judgments
`benchmark_results`	3	Paper-reported historical production measurements

All prompts are synthetic. This release contains no production conversations, fetched page bodies, user identifiers, credentials, access tokens, or private session data.

Evaluation schema

Each evaluations row provides a conversation, expected route and context mode, freshness/citation/artifact requirements, and observable acceptance criteria. The criteria intentionally evaluate behavior rather than exact wording so that model and provider changes do not invalidate the suite.

Routes use four labels:

direct: answer without live retrieval.
tools: perform a bounded live lookup or artifact operation.
deep_research: investigate several evidence-bearing aspects.
clarify: request information that is necessary to execute safely.

Measurement provenance

benchmark_results records the historical production snapshot reported in the paper. It is not a rerun on Hugging Face infrastructure. In particular, the 89.3% value is an aggregate Redis keyspace hit rate across internal operations, not a query-level semantic-cache hit rate and not an estimate of avoided model inference.

Suggested use

Run an agent on each conversation in evaluations.
Record its route, tool trace, final answer, citations, and artifacts.
Score the observable criteria with deterministic checks plus human review.
Use cache_pairs to test whether semantic caching reuses only equivalent requests inside the permitted scope.
Keep generated outputs separate from this immutable input release.

The suite is a release gate, not a claim of universal search quality. It is English-only, intentionally compact, and does not replace adversarial, multilingual, domain-expert, or large-scale human evaluation.

Reproducibility

Validate the package using only Python's standard library:

python scripts/validate.py

The source repository also contains executable CPU-only memory and latency gates under lixsearch/memoryEval.

Versioning

This is version 1.0.0. Static examples and reported measurements are kept stable within a major version. Future generated results should state the agent revision, model/provider, execution time, region, and evaluator revision.

Citation

@article{bhattacharya2026threelayer,
  title={A Three-Layer Caching Architecture for Low-Latency LLM Web Search on Commodity CPU Hardware},
  author={Bhattacharya, Ayushman and Gazi, Nihal},
  journal={arXiv preprint arXiv:2609.05463},
  year={2026},
  url={https://arxiv.org/abs/2609.05463}
}