Algolia-Central2 — Dev Log

2026-06-14 — Harness reliability hardening + reliability SOP

Status: Done What happened: The autocorrect optimization loop crashed twice and ran ~4h before fixes. Root-caused a single failure class (unreliable batch-LLM / agent-orchestration code) and fixed it: parallelized the judge + run-tests (mapWithConcurrency, ~5–10× faster), cached the fixed floor across rounds, guarded the propose seam + raised its token budget (was crashing on Gemini MAX_TOKENS empty), and widened the Gemini retry set to {429,500,502,503,504} + network throws. Added the previously-missing tests for the live adapters. Also fixed a UX issue (fabricated mock judge scores on idle launch → call-to-action). Key decisions: Encoded as a reusable SOP/runbook — **llm-harness-reliability** (canonical: repo docs/sop/llm-harness-reliability.md). Tests written: mapWithConcurrency (4), applyFloorCache (3), gemini retry (4) → lab/server 18 tests; judge 64, autocorrect 22, web 37 — all green; tsc clean. Next: Implement observability (timestamps/heartbeat/summary, std S8); commit the session batch; apply the staged typo-tolerance reformulation fix post-run.

2026-07-10 — Content Engagement strategic pivot named; vision + data-gathering opened

Status: Vision + data-gathering stage only. No code touched. What happened: Arijit framed a new strategic direction for Algolia-Central2: move Algolia from "search/discovery for ecommerce & product" to ContentSearch → Content Engagement — chat that doesn't just answer but discovers, understands, and actively engages a visitor like a human companion, driving them toward a journey outcome (book a call, start building, convert). First POC target: embed this content-engagement behavior across several algolia.com journeys in parallel — case study pages, pricing, and the homepage/landing page — with context/page-awareness driving personalized engagement. Backend = Algolia; delivery = Agent Studio agents. Deployment for the POC is a standalone demo environment simulating those pages, not an injection into live algolia.com. Success criteria = broad market/product-market validation (not one exec or one customer). Full synthesis: **content-engagement-vision**. Key decisions: None locked yet — this is pre-decision. Explicitly left open: whether this extends the existing live 3-agent build (General/Developer/Marketer on app 0EXRPAXB56, see SESSION.md session 7) or runs as a separate/parallel initiative. Arijit's call: "not sure yet — that's part of what we're figuring out." Data-gathering opened in parallel (tactical, not yet executed this entry): 1. Analytics API — full breadth/depth of real user queries against ALGOLIA_WWW_PROD_V2. 2. Insights API — click/conversion signal detail for the same index. 3. Index x-ray of ALGOLIA_WWW_PROD_V2 — data shape, richness gaps, enrichment targets (precedent: enrichment already done on AC2's own index, see session 5-6 enrichment A/B). 4. Screenshot sweep of representative algolia.com pages (case study / pricing / home) to ground a UI/UX brainstorm on where engagement should surface. 5. GA behavioral data — Arijit is pulling a Looker Studio export (datastudio.google.com/.../b05b44b9-.../Cr7KB); confirmed NOT fetchable via plain WebFetch (redirects to Google login) — needs either an authenticated browser session or a manual CSV/Sheets export from Arijit. Grounding correction made this session: ALGOLIA_WWW_PROD_V2 does not live in the AC2 build app — it lives in the VISIBILITY app 1QDAWL72TQ, documented in this project's own CLAUDE.md as Algolia's live keyword incumbent, read-only, benchmark target. Resolved without needing to ask Arijit, by reading CLAUDE.md rather than guessing. Next: Run the Analytics/Insights/index-x-ray pulls against 1QDAWL72TQ / ALGOLIA_WWW_PROD_V2; capture findings in a data-shape synthesis; then a joint UI/UX brainstorm (screenshots + query data + GA data together) on where engagement surfaces.

2026-07-10 (later same day) — Analytics + index x-ray pulled (raw data relocated)

Status: Data-gathering executed. Raw findings intentionally NOT kept here — Arijit called this out: analytics numbers and record dumps are transactional data, not knowledge, and pollute the wiki. Relocated to repo docs/algolia-data/analytics.md and docs/algolia-data/content.md (Algolia-Central2 repo, not the vault). What happened: Pulled Analytics API + full index content via direct REST (Arijit rejected the Algolia MCP tool — use the real keys already in .env, not MCP). Root-caused a region-host bug along the way (see memory feedback-algolia-analytics-region-host). Key decision that DOES belong here: raw analytics/content-shape numbers live in repo docs, not the vault. Vault holds the vision, the open questions, and the synthesized implications — not the transactional evidence itself. Next: Screenshot sweep of case study/pricing/homepage; GA Looker Studio export; joint UI/UX brainstorm once both data threads (Algolia Insights = what people search, GA/Looker = where traffic is) are in.

2026-07-12 — Five founding decisions resolved; full data-health audit run on both copies; merged into vault as single source of truth

Status: Done. What happened: Five open questions from the vision doc were answered directly (timeline, audience, success criteria, official naming, AC2 relationship) — see **2026-07-12-content-engagement-motion-decisions**. In parallel, ran the same full, unsampled field-health + analytics audit on both the current production index (ALGOLIA_WWW_PROD_V2, 17,138 records) and AC2's enriched copy (AC2_WWW_MULTI_NEURAL, 8,353 records) — see **data-health-current-production**, **data-health-enriched-copy**, and the composite-scored comparison in **data-health-final-verdict**. Headline: the two are statistically even (63.4 vs 62.6/100), but the enriched copy's edge comes entirely from deduplication — on content quality alone the unenriched baseline is ahead by 4.1 points. Key decisions: Content Engagement is the AC2 work (not parallel) — decision 5 above. An interactive companion site was also built (bible.chowmes.com/ac2/, source repo RAG/Algolia-Central2/algolia-content-engagement/) with the same findings presented with real record examples, a bang-for-buck action-item ranking, and a build log. This vault folder is the single source of truth — the interactive site is a presentation layer over the same facts, not a second place where the facts live. A separate, portfolio-wide live-render system was discovered running at bible.chowmes.com root (reads this exact vault folder on every request, no build step) — this project's vault content is already visible there automatically; only the ADR and data-health syntheses added today needed to be written into the vault explicitly, since they didn't exist as markdown before now. Next: Resolve the scope-vs-deadline tension (full 3-journey POC vs one journey, by Tuesday) — tracked in **open-questions**. Screenshot sweep, GA data pull, and Agent Studio build remain not started.