v2/index.md
Algolia-Central2 — Answer-Quality Lab — Wiki Index
Type: project Created: 2026-06-18 Last Updated: 2026-07-04 (migrated coordinator-native-findings + entity-boost ADR + full Algolia API reference from repo — these existed only in-repo, never in the vault, until now)
Read this file first. It maps the entire wiki.
STATUS (2026-06-18 evening): BUILD PLAN WRITTEN + HARDENED → ready for UltraCode. Plan lives in the repo:
docs/superpowers/plans/2026-06-18-answer-quality-lab-2x2-refactor.md(authoritative). Corrections that OVERRIDE the ADRs below: build home = CENTRAL app0EXRPAXB56(NOT a separate "FLAGSHIP" app); indices namedAC2_WWW_<ARCH>_<RETRIEVAL>, agentsac2-*(NOTALGOLIA_WWW_PROD_V2_*); Maverick = coded coordinator + native source-scoped specialist agents; both OpenAI + Gemini providers; FAIRNESS INVARIANT = identical model across all 4 panels; multi-turn = generated follow-up. See repoSESSION.md+ memoryproject-build-home-central-app.md.
What this is (one line)
A 2×2 four-panel "Answer-Quality Lab" on Algolia that proves, with an LLM judge, whether neural beats keyword and whether multi-agent beats single-agent — the multi-agent built as content-source specialists orchestrated by Maverick.
Files
| File | Purpose | Last Updated |
|---|---|---|
| wiki/overview.md | Pivot, the 2×2, the multi-agent design | 2026-06-18 |
| wiki/requirements.md | Spec: panels, indices, judge, sync | 2026-06-18 |
| wiki/decisions/ | Architectural decisions (ADRs) | 2026-06-18 |
| wiki/dev-log.md | Build chronicle | 2026-06-18 |
| wiki/open-questions.md | Open questions | 2026-06-18 |
| wiki/sources/ | Research source summaries | — |
| wiki/concepts/ | Domain concept pages | — |
| wiki/entities/ | People, orgs, tools | — |
| wiki/syntheses/ | Cross-cutting analysis | — |
| wiki/research/2026-06-27-coordinator-algolia-native-findings.md | Live-probed Agent Studio native capability inventory (what's native vs custom in the coordinator) — migrated from repo 2026-07-04, was never in the vault before | 2026-06-27 |
| wiki/reference/algolia-api/ | Full local Algolia REST/MCP API reference (10 files: search, indexing, settings, relevance, analytics, engagement, ingestion, ops/security, MCP tool map) — migrated from repo 2026-07-04, was never in the vault before | 2026-06-XX |
Repo source of truth for exhaustive detail: docs/refactor/demo-app-retrieval-comparison-brief.md (in the Algolia-Central2 repo).
Decision Records
| ADR | Summary | Date | Status |
|---|---|---|---|
| pivot to 2×2 four-panel lab | Drop incumbent-vs-us; compare our own retrieval × architecture | 2026-06-18 | Accepted |
| FLAGSHIP app + four indices | Move off ARIJIT-TEST; 4 indices seeded from Visibility | 2026-06-18 | Accepted |
| content-source multi-agent | Maverick + 4 source-scoped specialists | 2026-06-18 | Accepted |
| judge: 3-dim per panel | Judge scores all 4 panels; 2×2 scorecard | 2026-06-18 | Accepted |
| daily delta-sync | Data-only sync Visibility → 4 indices | 2026-06-18 | Accepted |
| UX + judge design | 2×2 grid scoreboard + slide-in judge drawer; lean judge panel; grounding gate ×1; 1-judge live / 3-judge batch | 2026-06-18 | Accepted* |
| drop expandedQuery from retrieval | Send raw user turn; rephrase gave no lift + a grounding hazard on bait queries; NeuralSearch fires on raw NL | 2026-06-28 | Accepted |
| stateless replay vs Redis | Keep messages[] replay + dossier, no Redis; native memory doesn't carry; Redis≠token-saving |
2026-06-28 | Accepted |
| content-source routing spike verdict | Source SCOPING = no lift (KILL) on a fair baseline; depth ceiling is DATA (only Support+"Other" deep); honed-prompt+warm-baton phase next | 2026-06-28 | Accepted (supersedes-in-part content-source-multi-agent) |
| rationalize to 3 agents | Keep multi-agent for demo/personality/scope (NOT quality — thesis A/B flat); 4→3: General (fronts+routes, Website+Other) · Developer (Documentation+Developers+Support+Academy) · Marketer/Merchandiser (Blog+CustomerStories+Resources); General hands off by intent | 2026-07-01 | Accepted (supersedes-in-part content-source-multi-agent) |
| entity-boost: native optionalFilters, not Atlas | RC2's Atlas pattern (separate entity index + query expansion) was a crutch for a non-neural retriever; AC2 biases retrieval via native optionalFilters on existing facets instead — zero extra hops, same grounding |
2026-07-01 | Accepted (migrated from repo 2026-07-04, was not in vault) |
Sources
| Title | Slug | Date |
|---|---|---|
| (none yet) |
Recent Activity
| Date | Event |
|---|---|
| 2026-06-18 | Wiki initialized; recorded pivot, architecture, 5 ADRs, open questions |
| 2026-06-18 | Finalized agent roster (Maverick · Technical · Marketer · Academy · Support); locked "data-driven not legacy" v2 principle |
| 2026-06-18 | Designed lab UX + judge panel (2×2 grid, slide-in judge drawer, lean content, 1/3-judge tiers); dedup profiled; build roadmap set |
| 2026-06-28 | Backlog A: dropped brain.expandedQuery from retrieval (scored A/B + Gate-1 proof) — shipped main dc03bc7 |
| 2026-06-28 | Backlog B: native memory does NOT carry (probe negative) → keep stateless replay; ADR stateless-vs-Redis — main 03e9184 |
| 2026-06-28 | Content-source routing SPIKE → KILL on scoping (fair baseline +0.25 noise); data-hardened (only Support+"Other" deep); honed-prompt+warm-baton phase next. ADR content-source-routing-spike-verdict |
| 2026-07-01 | Core-thesis A/B (33Q rounds=3): 4 specialists give NO lift over 1 generic agent (clean Δ−0.33=flat). Native crawler-army enriched the index (deep bodies); enrichment A/B (body searchable) also null. Decision: rationalize 4→3 agents (General/Developer/Marketer-Merchandiser), General fronts + hands off. ADR rationalize-to-3-agents |
| 2026-07-04 | Vault-sync gap found and closed (from a PRISM-project session, PIP repo): coordinator-algolia-native-findings.md, the entity-boost ADR, and the full docs/algolia-api/ reference (10 files) existed only in the repo, never committed here — migrated in. Reason: PRISM V2 planning needed these findings and almost re-researched them from scratch before checking the vault first. |