tarekmasryo/rag-qa-logs-corpus

🧠 RAG QA Logs & Corpus

Multi-Table Synthetic RAG Telemetry for Quality, Hallucinations, Latency, and Cost

A production-style, privacy-safe synthetic dataset that mimics telemetry exported from a real Retrieval-Augmented Generation (RAG) system — from corpus → chunk index → ranked retrieval lists → evaluation outcomes.

It is designed as an analysis-ready multi-table benchmark for:

• RAG quality analysis (correctness, faithfulness, hallucination)
• retrieval strategy comparisons (dense / bm25 / hybrid / rerank variants)
• risk & meta-modeling (predict failures/hallucinations from telemetry)
• latency & cost trade-offs (ms, tokens, USD)
• dashboards and teaching materials

✅ All records are fully synthetic — no real users, customers, patients, or company data.

---

🔐 Privacy & Synthetic Data

This dataset is fully synthetic:

• No real users, customers, patients, or organizations are represented.
• No personally identifiable information (PII) is included.
• All IDs, queries, documents, and logs were programmatically generated to resemble realistic RAG telemetry while preserving privacy.

---

📌 Quick Overview

• Total size: 103,255 rows across 6 linked CSV tables
• Type: multi-table tabular logs + short text fields (queries, answers, chunk text)
• Main targets (labels): is_correct, hallucination_flag, faithfulness_label
• Stable join keys: doc_id, chunk_id, example_id, run_id, scenario_id, query_id
• Splits: train, val, test

Domains covered (documents): support_faq, hr_policies, product_docs, developer_docs, policies, financial_reports, medical_guides, research_papers, customer_success, data_platform_docs, mlops_docs, marketing_analytics

Task types (evaluation runs): factoid, explanation, summarization, multi_hop, table_qa, temporal_reasoning, comparison, instruction_following

Retrieval strategies: dense, bm25, hybrid, dense_then_rerank, bm25_then_rerank

Note: run_id is a per-example request/trace identifier (not a single run spanning many examples).

---

📂 Files

This dataset contains the following CSV files:

• rag_corpus_documents.csv — document-level corpus metadata
• rag_corpus_chunks.csv — chunk-level index + chunk text
• rag_retrieval_events.csv — per-chunk retrieval telemetry (rank, score, relevance)
• eval_runs.csv — QA evaluation runs (labels, metrics, latency, cost, configs)
• scenarios.csv — scenario templates and use cases
• data_dictionary.csv — column-level schema documentation across all tables

All files use snake_case, are tabular and ML-ready, and join cleanly via stable IDs.

---

📊 Exact Table Summary

File	Rows	Columns	Granularity
rag_corpus_documents.csv	658	19	One row per document
rag_corpus_chunks.csv	5,237	6	One row per chunk
rag_retrieval_events.csv	93,375	12	One row per retrieved chunk per example
eval_runs.csv	3,824	49	One row per QA evaluation example
scenarios.csv	62	13	One row per scenario template
data_dictionary.csv	99	5	One row per column definition

---

🔗 How to Join (Schema Map)

• Documents → Chunks rag_corpus_documents.doc_id = rag_corpus_chunks.doc_id

• Eval Runs → Retrieval Events eval_runs.example_id = rag_retrieval_events.example_id (also run_id, scenario_id, query_id, split are present for consistency)

• Eval Runs → Scenarios eval_runs.scenario_id = scenarios.scenario_id (and query_id)

• Retrieval Events → Chunks rag_retrieval_events.chunk_id = rag_corpus_chunks.chunk_id

---

🚀 Loading Examples

Option A — Pandas (local files)

import pandas as pd

docs = pd.read_csv("rag_corpus_documents.csv")
chunks = pd.read_csv("rag_corpus_chunks.csv")
events = pd.read_csv("rag_retrieval_events.csv")
runs = pd.read_csv("eval_runs.csv")
scenarios = pd.read_csv("scenarios.csv")

# Example join: runs → events → chunks → docs
df = (runs.merge(events, on="example_id", how="left", suffixes=("", "_evt"))
          .merge(chunks[["chunk_id","doc_id","domain","chunk_index","estimated_tokens"]], on="chunk_id", how="left")
          .merge(docs[["doc_id","domain","title","source_type","pii_risk_level","security_tier"]], on="doc_id", how="left",
                 suffixes=("", "_doc")))

df.head()

Option B — Hugging Face Datasets (explicit data_files)

from datasets import load_dataset

# Replace with your dataset repo id on the Hub, e.g. "tarekmasryo/<repo_name>"
repo_id = "<your-namespace>/<your-dataset-repo>"

ds = load_dataset(
    repo_id,
    data_files={
        "documents": "rag_corpus_documents.csv",
        "chunks": "rag_corpus_chunks.csv",
        "retrieval_events": "rag_retrieval_events.csv",
        "eval_runs": "eval_runs.csv",
        "scenarios": "scenarios.csv",
        "data_dictionary": "data_dictionary.csv",
    },
)
ds["eval_runs"].to_pandas().head()

---

🎯 Targets & Typical Tasks

Primary learning targets:

• is_correct — classification (correct vs incorrect)
• hallucination_flag — classification (hallucinated vs not)
• faithfulness_label — multi-class / categorical faithfulness

Paired with telemetry (retrieval ranks/scores, recall/MRR, latency, tokens, configs), this enables:

• meta-models that predict answer failure/hallucination risk
• risk scoring for guardrails (block / escalate / rerun / abstain)
• policy design for fast vs careful modes based on cost/latency/quality trade-offs

---

📘 Data Dictionary Notes

data_dictionary.csv provides column-level documentation across all tables: table_name, column_name, dtype, description, allowed_values.

Important naming note: if table_name uses logical names (e.g., rag_qa_eval_runs, rag_qa_scenarios) while your filenames are eval_runs.csv and scenarios.csv, treat table_name as a logical group name. For strict 1:1 naming, either rename the CSVs or update table_name values.

---

⚠️ Limitations

• This is a synthetic benchmark, not production data.
• rag_corpus_chunks.chunk_text may be more templated / less diverse than real-world corpora.
• Intended for research, teaching, benchmarking, and prototyping.
• Not for high-stakes decisions (clinical, legal, financial).

---

📜 License

CC BY 4.0 (Attribution Required)

---

📚 Citation

If you use this dataset in notebooks, papers, demos, or teaching material, please cite the dataset page on the Hub and reference:

“RAG QA Logs & Corpus — Multi-Table Synthetic RAG Benchmark” by Tarek Masryo