Karmane/enterprise-rag-internal-knowledge-search-benchmark-sample

Enterprise RAG and Internal Knowledge Search Benchmark Dataset -- Free Evaluation Sample

This dataset packages a synthetic internal company workspace and an evidence-linked benchmark table into one product for teams building enterprise RAG systems, internal search assistants, knowledge-base copilots, and deep-search evaluation pipelines.

The benchmark is designed around a realistic fictional company, AsteraOps Cloud, with multiple departments, renamed projects, stale roadmaps, support escalations, compliance controls, customer success notes, engineering tickets, email decisions, and changelog entries. The point is not to provide another generic question-answer dataset. The point is to provide a corpus-grounded evaluation set where every question is tied to explicit enterprise-style artifacts and where the benchmark can punish common production failures: retrieving the wrong version of a document, ignoring email decisions, over-trusting stale FAQs, flattening contradictions, or answering confidently when the corpus does not actually contain the requested fact.

The workspace contains four coherent internal programs:

• Atlas Assist for evidence-grounded internal search
• Nimbus SSO for enterprise SSO and SCIM rollout
• Beacon Vault for governed audit exports and legal-hold workflows
• Harbor Queue for support escalation and SLA operations

Each query row includes the benchmark question, the gold answer, answerability labels, evidence document identifiers, evidence document titles, exact citation spans, retrieval notes, benchmark task tags, and a JSON bundle of the supporting source documents. The rows also include a compact manifest of the full synthetic workspace so buyers can reconstruct retrieval pools, create subsets, or feed the benchmark into their own evaluation harnesses without reverse-engineering undocumented context.

This matters because enterprise retrieval is usually judged on private corpora that cannot be shared. Buyers often need a realistic benchmark but do not want to expose real Slack exports, customer tickets, policy wikis, or production email. Synthetic benchmark work solves that confidentiality problem, but many public datasets still feel too clean, too academic, or too detached from how internal search fails in practice. This build intentionally adds the kinds of friction teams see in production: renamed programs, owner changes, stale timelines, policy overrides, customer milestones, and unresolved conflicts where the correct system behavior is to abstain or call out ambiguity instead of inventing certainty.

The dataset is useful for several practical workflows:

• Compare retrievers on current-versus-stale document selection
• Measure whether answer generation cites the right evidence documents
• Test no-answer behavior when a likely enterprise detail is missing
• Score contradiction handling when old and new artifacts disagree
• Benchmark customer-history reconstruction across support and CS notes
• Audit policy interpretation against current governance docs
• Evaluate deep-search agents that need to gather evidence from 2-8 internal artifacts before answering

The corpus was generated locally with deterministic templates rather than scraped from a real company. Shared fictional people, projects, dates, customer accounts, and control changes are propagated across all artifacts so the workspace stays coherent. Rule-based generation was then used to produce benchmark questions tied to explicit evidence sentences. The benchmark rows were normalized into analysis-ready columns, JSON-serialized bundle fields, consistent answerability labels, and typed metadata that can be loaded directly in pandas, notebooks, BI tools, or custom evaluation services.

This is easier than building a benchmark manually from scratch. Buyers do not need to author a fictional company graph, write dozens of enterprise artifacts, create query-answer pairs, hand-link evidence docs, generate citation spans, or add no-answer and contradiction cases on their own. The free sample proves the schema, artifact style, evidence linking, and evaluation framing. The full dataset adds the complete benchmark table, the full synthetic workspace manifest, more cross-project reasoning rows, and enough coverage to test retrieval, citation grounding, and abstention behavior without starting from a blank page.

Limitations still apply. The workspace is synthetic, not a replay of a real company. The benchmark focuses on text-based internal artifacts rather than screenshots or binary documents. It is meant for evaluation and benchmarking, not for claiming guaranteed production performance. Teams should still adapt the benchmark to their own security model, data model, and retrieval stack. Even with those limits, it gives enterprise AI builders a reusable internal-knowledge benchmark that is much closer to day-to-day enterprise search than a generic public QA set.

This repository contains a 10-row free evaluation sample of the full 104-row production dataset.

The sample is fully open and loadable without authentication so you can inspect the schema, explore features, and validate quality before purchasing.

---

Full Production Dataset

This public repository is a 10-row evaluation sample. The full production-grade dataset contains 104 rows across 43 columns.

• Get Access to the Full Dataset: https://thearticulated.gumroad.com/l/enterprise-rag-internal-knowledge-search-benchmark
• Full Dataset on Hugging Face (gated): https://huggingface.co/datasets/Karmane/enterprise-rag-internal-knowledge-search-benchmark
• Sample rows: 10
• Full dataset rows: 104
• Columns: 43

After purchase, provide your Hugging Face username and request access on the full dataset page.

---

Machine Learning and Analysis Use Cases

This dataset is a rich tabular and natural-language playground suitable for:

• Text Classification: Use text columns like query_id to predict categorical targets such as query_text.
• Tabular Regression / Classification: Predict conflict_flag or stale_document_flag using the other numeric and categorical features.
• NLP Feature Extraction: Extract embeddings or features from text columns like query_id for downstream tasks.
• Exploratory Data Analysis: Filter, pivot, and visualize across all columns in pandas, Excel, or any BI tool.

---

Dataset Structure

• Rows: 10
• Columns: 43
• Split: data
• File: data/data.parquet

Column Descriptions

• query_id: Stable benchmark query identifier.
• query_text: User-style benchmark question grounded in the synthetic enterprise corpus.
• gold_answer: Auditable target answer, including no-answer or conflict language when appropriate.
• answerability: Label indicating whether the query is answerable, no-answer, or conflicting.
• reasoning_type: Primary reasoning pattern required to answer the query.
• difficulty: Heuristic difficulty label based on evidence count and conflict handling.
• evidence_doc_ids: JSON array of supporting document identifiers.
• evidence_doc_titles: JSON array of supporting document titles.
• citation_spans: JSON array of evidence quote spans with character offsets.
• evidence_summary: Plain-English summary of why the cited documents support the answer.
• judge_rubric: Compact grading rubric for exactness, grounding, and abstention behavior.
• temporal_scope: Whether the query requires current-state, historical, or mixed temporal reasoning.
• conflict_flag: Boolean flag indicating the query includes contradictory or competing evidence.
• stale_document_flag: Boolean flag indicating stale-document handling is part of the task.
• no_answer_flag: Boolean flag indicating the correct behavior is to abstain or report missing information.
• department_scope: Comma-separated departments involved in the evidence set.
• project_scope: Project or cross-project scope of the query.
• source_count: Actual number of evidence documents attached to the query row.
• expected_source_count: Expected number of source documents a strong system should consult.
• primary_artifact_type: Primary evidence artifact type such as wiki, roadmap, or ticket.
• supporting_artifact_types: JSON array of secondary artifact types involved in the answer.
• company_persona: Short description of the synthetic company context.
• synthetic_company_name: Name of the fictional company represented in the corpus.
• created_date: Earliest evidence document date tied to this query.
• updated_date: Latest evidence document date tied to this query.
• document_time_window: Date range covered by the evidence documents.
• query_category: Higher-level benchmark category used for slicing the evaluation set.
• evaluation_split: Train, validation, or test label for downstream eval pipelines.
• benchmark_task: Primary evaluation task family.
• retrieval_notes: Notes about retrieval strategy, stale docs, or cross-document hops required.
• failure_mode_targeted: Primary model failure mode this query is designed to expose.
• answer_format: Expected answer format such as short_text, date, list, or abstain.
• source_document_titles: JSON array of evidence document titles duplicated for convenience.
• source_document_types: JSON array of evidence document artifact types.
• source_document_bundle_json: JSON payload containing the full evidence documents and metadata.
• synthetic_workspace_bundle_json: JSON payload containing a compact manifest of the entire synthetic workspace.
• provenance_note: Statement describing how the synthetic benchmark was generated.
• license_status: Rights statement for the benchmark content.
• synthetic_generation_method: High-level summary of the deterministic corpus generation approach.
• quality_review_status: Status of rule-based consistency and type checks applied to the row.
• source_domain: Synthetic source domain used for provenance tracking.
• corpus_document_count: Number of documents in the full synthetic workspace manifest.
• query_intent: Concise buyer-facing description of the retrieval task under test.

---

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Karmane/enterprise-rag-internal-knowledge-search-benchmark-sample")
print(dataset)
print(dataset["data"][0])

# Get as a pandas DataFrame
# df = dataset["data"].to_pandas()

---

Intended Use

This dataset is intended for research, experimentation, analysis, model prototyping, dashboard building, and market research.

---

Karmane. (2025). Enterprise RAG and Internal Knowledge Search Benchmark Dataset (Free Sample). Hugging Face. https://huggingface.co/datasets/Karmane/enterprise-rag-internal-knowledge-search-benchmark-sample