xlab devcorpus — proposal, foundation, parked

Date: 2026-05-12 Initiative: Portfolio-wide conversational search over the dev-reports corpus, powered by Gemini via Vertex AI Search Status: 🅿️ Parked after Phase 2. Foundation live (zero ongoing cost), build deferred. Resume conditions documented below. Companion: dogwood_phase01_pre_flight_2026-05-13.md

TL;DR

Walked the full arc — scope, foundation, mockup, decision — in one evening. Vertex AI Search engine xlab-devcorpus-app is live in audrey-experiments, indexing 58 markdown dev-reports from gs://xlab-devcorpus-md. Zero queries fired, so zero ongoing cost. The build out to a phone-friendly corpus.dare.co.uk is real — about 3 hours, under $5/month — but parked behind two higher-leverage threads: dogwood Phase 1 close-out tomorrow morning, and the broader pivot that AI voice for real customer problems beats AI search for self. Internal corpus chat is “mildly interesting”; customer-facing voice is the frontier.

The mockup itself is the deliverable for now — a cheap deliberation before commitment, per the mockup-before-full-build pattern. NotebookLM is queued as the throwaway phone-and-voice experiment to falsify the assumption that conversational corpus access is even a habit worth tooling for.

What now exists (paid: $0, parked-cost: $0)

Live preview surface (no build required): https://console.cloud.google.com/gen-app-builder/engines/xlab-devcorpus-app/preview/search?project=audrey-experiments

The proposal in 60 seconds

Three artefacts and a sync trigger. GCS bucket holds the markdown corpus, Vertex AI Search Data Store indexes it, a Search App exposes conversational query with Gemini-grounded citations. A GitHub Action on devreports-content push keeps the corpus fresh (Phase 3, deferred). Scope:

Original scope: [proposal in chat 2026-05-12, captured here in this report].

Phases — built vs parked

Disproof criteria from the proposal still hold and now have concrete owners:

1. Quality: smoke test ≥15 of 20 queries return grounded, cited, correct answers. Test surface = the live console preview tomorrow.
2. Habit: any 7-day window post-Phase-4 with zero queries → archive.
3. Cost: unexpected spend past $20/mo at personal volume → misconfiguration.

UI sketch — what corpus.dare.co.uk would have looked like

Mockup-only at park time. The mockup is the deliverable: cheap deliberation before a 3-hour build.

┌─────────────────────────────────────────────────────────────┐
│  devcorpus                                  dan@dare.co.uk  │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Ask the corpus.                                            │
│                                                             │
│  ┌───────────────────────────────────────────────────────┐  │
│  │ what changed on dashboard.dare.co.uk last week?    🎙 │  │
│  └───────────────────────────────────────────────────────┘  │
│                                                             │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  The dashboard cron silently no-op'd for three days         │
│  in early May, root-caused to macOS TCC blocking launchd-   │
│  spawned shells from reading scripts in ~/Downloads ¹.      │
│  Smart Tiered Cache was enabled across the portfolio on     │
│  2026-05-11, with a 2026-05-14 re-check planned ².          │
│                                                             │
│  Sources                                                    │
│  ¹ feedback_downloads_is_least_favourite                    │
│  ² project_cloudflare_tiered_cache_enabled                  │
│                                                             │
│  ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─   │
│                                                             │
│  Recent asks                                                │
│  • compare audrey vs dogwood foundation reports     3m ago  │
│  • dare canonical narrative frame                  15m ago  │
│  • narrator voice tuning current state              1h ago  │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Five UI primitives: header, big input with mic, answer block, sources list (clickable, hover-preview to source report), recent-asks rail. Mobile collapses to single column. Voice via the browser’s native Web Speech API — free, no extra service account.

Architecture — the wiring that didn’t get built

┌──────────────────┐
│ Mobile / desktop │
│     browser      │
└────────┬─────────┘
         │ HTTPS
         ▼
┌──────────────────────────────────────────────────┐
│ Cloudflare edge                                  │
│  ┌────────────────────────────────────────────┐  │
│  │ Access policy: dan@dare.co.uk + dan@gf.cx  │  │
│  └────────────────────────────────────────────┘  │
│  ┌────────────────────────────────────────────┐  │
│  │ Pages: corpus.dare.co.uk                   │  │
│  │   static HTML + JS frontend                │  │
│  └────────────────────────────────────────────┘  │
│  ┌────────────────────────────────────────────┐  │
│  │ Worker: /api/ask                           │  │
│  │   reads GCP_SA_KEY secret (op-injected)    │  │
│  │   mints token → POST :answer               │  │
│  │   streams response back to browser         │  │
│  └────────────────┬───────────────────────────┘  │
└───────────────────┼──────────────────────────────┘
                    │ HTTPS, bearer auth
                    ▼
┌──────────────────────────────────────────────────┐
│ Vertex AI Search                                 │
│  engines/xlab-devcorpus-app                      │
│   └─ dataStores/xlab-devcorpus                   │
│       └─ gs://xlab-devcorpus-md/*.md             │
└──────────────────────────────────────────────────┘

Thin-edge pattern — Worker proxies to existing canonical source, no data duplication. Runtime auth via a scoped SA (xlab-devcorpus-runtime-sa, roles/discoveryengine.viewer, JSON key in op://Private/) so the existing audrey-vertex-sa stays single-purpose.

Decision — parked, not killed

Two reasons.

1. Dogwood Phase 1 close-out is the higher-priority thread

Tomorrow’s pre-flight already caught a real production bug (worker.js Twilio Auth header still uses legacy auth_token despite API Key constants being declared). That work has actual customer-facing stakes — the 7am SMS briefing breaks if it’s mishandled. Internal corpus chat doesn’t compete on importance.

2. AI voice for real customer problems > AI search for self

Stated explicitly tonight: “I would sooner develop AI VOICE for real customer problems.” Direct corollary: the same 3 hours that would build corpus.dare.co.uk would meaningfully advance the dogwood voice agent (Phase 5 in the dogwood roadmap), an audrey gift-advisor voice flow, or client-facing voice work. Internal-tool builds are an indulgence; customer-facing voice is the leverage. Park rather than kill — internal tools can resume if/when habit-data justifies, but the default investment lane is outward, not inward.

New memory landed: feedback_customer_voice_over_internal_tools.

Resume conditions — when to un-park

Any one of these is sufficient signal to resume:

1. NotebookLM-as-experiment shows habit. Upload the same 58 markdown files to NotebookLM this week. Ask the same disproof query set on the phone. If you reach for it 3+ times unprompted over the next 7 days and the manual re-upload friction is starting to bite — that’s the green light.
2. A workflow demands programmatic queries. If a dev-report-publish script wants to ask the corpus “have I already written about X?” before generating a new artefact, Phase 4 (~/bin/devcorpus CLI) is the unlock.
3. A customer engagement needs a similar surface. A client wants Q&A over their own documentation corpus — devcorpus becomes the reference implementation, and corpus.dare.co.uk becomes worth building as a portable pattern.

If none of these fire within 90 days, the cleanup move is gcloud storage rm -r gs://xlab-devcorpus-md/ + delete the Data Store + delete the Engine. Cost to leave it dormant in the meantime is rounded to zero, so the default is benign neglect.

NotebookLM — the cheaper experiment to run first

Concrete this-week test, not a build:

1. Open NotebookLM on iPhone (or web).
2. Upload ~/Code/devreports-content/*.md (drag-and-drop, ~58 files).
3. Ask the proposal’s disproof query set: - “what changed on dashboard.dare.co.uk last week?” - “summarise the dogwood Twilio rotation” - “what memories were written from the audrey foundation work?” - “compare the dare vs dogwood gap audits”
4. Try the voice mode on iPhone.
5. After 7 days, ask: did I reach for this unprompted? Did manual re-upload annoy me?

Cost: 10 minutes upload + whatever ambient ask-time. Output: an honest signal on whether conversational corpus access on a phone is a habit worth investing in.

What this teaches the portfolio

• Mockup-before-full-build pattern at work. Three hours of UI/architecture sketching + a parked decision is the same cheap-deliberation move that produced the dare A/B preview report. Pattern keeps paying.
• Thin-edge architecture pattern at the model layer. Don’t rebuild the data — point a model at the existing source-of-truth. Worker → Vertex AI Search is the same shape as agent-edge → Shopify, one level up the stack.
• Free-tier-first GCP experimentation works. Phase 0-2 cost zero, taught the API surface area, and produced a real artefact (the engine is queryable in the console tomorrow morning). Pattern transfers to any future GCP exploration.
• “Mildly interesting” is a useful disposition. Not every idea needs to be built. Surfacing the interest level honestly + parking with resume conditions beats either building-and-regretting or killing-and-forgetting.
• Cross-portfolio infra (xlab_* namespace) graduates well. Reused audrey-experiments as incubator project per the xlab-co lifecycle model. If devcorpus ever earns its own project, it migrates cleanly because the resource names are already portfolio-scoped, not audrey-scoped.

Memories written tonight

Deferred per proposal until Phase 2 quality smoke-test:

• project_devcorpus_operating_notes.md — bucket/dataStore/engine IDs + working query examples (writes after smoke test passes)
• reference_vertex_ai_search_setup.md — gcloud + REST steps for spinning new Data Stores; portable to client work
• feedback_markdown_corpus_as_rag_substrate.md — dev-reports unintentionally well-shaped for RAG; editorial discipline pays a second dividend

Linked artefacts

• Live preview surface: console.cloud.google.com/gen-app-builder/engines/xlab-devcorpus-app/preview/search?project=audrey-experiments
• Tomorrow’s dogwood pre-flight: ~/Code/dogwood-house/dev-reports/dogwood_phase01_pre_flight_2026-05-13.md
• GCP foundation reference: project_gcp_audrey_foundation
• Lifecycle pattern: project_xlab_co_lifecycle_model

Park with resume conditions, not abandonment. The foundation is free to keep; the build only commits when the habit-signal lands.