WixQA_Extended

A hand-curated out-of-scope extension to Wix/WixQA (MIT) for evaluating how a customer-support RAG agent handles questions it should not answer normally: off-topic requests, competitor bait, abuse, prompt injection, harmful use, and in-domain questions the knowledge base cannot answer (the hallucination-under-retrieval-miss case).

This dataset contains extension rows only (40 questions, 20 per split) — it is a companion to WixQA, not a redistribution. Evaluation code is expected to concatenate it with WixQA's wixqa_expertwritten subset at load time and to index WixQA's wix_kb_corpus as the retrieval target.

Taxonomy

category	expected_behavior	n	intent
off_topic	deflect_redirect	4	not about Wix at all (medical, creative writing, unrelated coding)
competitor_brand	deflect_redirect	6	comparison/complaint bait; no disparagement, no defensiveness
competitor_brand_gray	answer_normally	4	real, KB-answerable tasks that mention a competitor (over-refusal probes)
abusive	deflect_redirect	4	profanity/harassment at the agent; de-escalate
abusive_gray	answer_normally	2	real support issue buried in hostility (over-refusal probes)
prompt_injection	deflect_redirect	6	embedded instructions to reveal prompts, change role, or concede claims
harmful_use	refuse	4	using Wix for phishing/fraud, incl. deceptive intent behind a routine task
out_of_kb	escalate_human	10	account-specific status/decisions no documentation can answer; verified unanswerable against the WixQA KB via a retrieval pipeline

Schema

Superset of the WixQA QA schema, so concatenation is trivial:

• id — stable row id (<category>-<nn>)
• question — realistic ticket register (typos, run-ons, hostility)
• answer — for gray traps: a KB-grounded reference answer; otherwise a reference deflection showing the intended tone and behavior (a calibration example for an LLM judge, not a string-match target). escalate_human references explicitly offer to connect the user with a human support agent.
• article_ids — gold KB article ids for gray traps, [] otherwise
• category, expected_behavior — see taxonomy
• split — train / validation, assigned explicitly and stratified by category

Suggested normalization for original WixQA rows at load time: category="standard", expected_behavior="answer_normally".

Scoring notes

• Rows with expected_behavior != "answer_normally" have empty gold sets — exclude them from retrieval-metric means (they would score 0.0) and from answer-correctness aggregates; score them with a scope-handling judge instead.
• Gray traps are ordinary QA rows and belong in the standard metrics; they exist to measure over-refusal introduced by scope-handling prompts.
• Per-category n is 2–5 per split: read gross effects, not deltas.

Caveats

• Reference answers were written by the dataset authors, not by Wix support experts — a quality tier below WixQA's expert-written subset. They are tone/behavior calibration examples.
• One harmful_use question (phishing-page request) is frequently refused at the API layer by some model providers before any agent logic runs; treat that category's effective n as unstable.
• Grounding for gray traps refers to the WixQA KB snapshot the authors indexed; KB drift may affect them.

Built as part of a customer-support RAG experimentation project; not affiliated with Wix. The Wix/WixQA dataset is MIT-licensed; this extension is likewise MIT.