Overview — Answer-Quality Lab (2×2)

The pivot

The lab previously compared incumbent vs us (① algolia.com website search · ② Ask AI as a quality floor · ③ our single Agent-Studio agent on app VVKSSPDMJX). We are tearing that framing down.

The new lab is a 2×2 experiment over our own system. Two independent variables, four cells, all on identical data:

The question shifts from "do we beat Ask AI" to "does neural beat keyword, does multi-agent beat single, and do they compound?" — each answered by an LLM judge scoring answer quality, not assumed.

Where it runs

Everything moves to the demo app FLAGSHIP_Accelerator_Program_APP (NeuralSearch enabled). We abandon VVKSSPDMJX (ARIJIT-TEST). The live Visibility app ALGOLIA_WWW_PROD_V2 (1QDAWL72TQ) is the READ-ONLY upstream source we copy from.

The four indices

All seeded from one fresh copy of the Visibility ALGOLIA_WWW_PROD_V2 (English only for now). Naming: ALGOLIA_WWW_PROD_V2_{SINGLE|MULTI}_{KEYWORD|NEURAL}. Keyword vs neural is a retrieval-config difference on identical records; single vs multi pairs hold the same data so the only variable across a pair is architecture.

The multi-agent design (Panels 3 & 4) — RATIONALIZED to 3 agents (2026-07-01)

Not the rc2/rc3 sales-discovery machine. Content-source specialist agents, each a shared-index view scoped by a source facet filter (native Algolia, minimal code). Rationalized from 4 specialists → 3 agents — kept for demo/personality/knowledge-scope, NOT quality (the thesis A/B showed specialists give no lift over one generic agent). See 2026-07-01-rationalize-to-3-agents (supersedes-in-part 2026-06-18-content-source-multi-agent).

• General (concierge / front desk) — sources Website, Other. Always interfaces first; routes + opportunistically hands off.
• Developer — sources Documentation, Developers, Support, Academy.
• Marketer / Merchandiser — sources Blog, Customer Stories, Resources.

General leads every conversation and warm-baton hands off to Developer (technical intent) or Marketer (value/merchandising intent). source facet validated against the live index.

What stays

The AI Judge + Autocorrect loop (provider-agnostic, zero-flicker) carries over. The judge now scores all four panels on three dimensions (grounding hard-floor + scored, confidence, breadth/depth; mean of 3 judges) and produces the 2×2 scorecard. See 2026-06-18-judge-3dim-per-panel.

Non-negotiable

110% grounded: every factual claim traceable to the index, or it doesn't ship. Grounding enforced by hardened agent instructions + bait-query verification. Minimise custom code; prefer native Algolia.

Context

This is also a flagship accelerator demo and feeds a POC direction with Adobe + Contentstack (both already tagged in the data's facets.facet4), then production on algolia.com.