A1 · Grounded assistance · professional services, life sciences, public sector, financial services, manufacturing
Grounded research desk
Answers that cite the exact passage, or say they cannot.
- Customer job
- Find a defensible answer in policies, research reports and filings without reading forty documents.
- Today
- Analysts search shared drives and intranet pages, read in full, and paste excerpts into notes; answers vary by analyst and sources are rarely recorded.
- Accepted result
- An answer the analyst forwards without rework, with each claim traceable to a passage they are permitted to see.
Analysts ask questions of a governed corpus and receive answers in which every claim links to a retrieved passage the asker is allowed to read. Unsupported questions are refused, not guessed.
Google Cloud / primary provider
What carries each step.
Agent Search handles hybrid retrieval and permission-aware indexing; RAG Engine covers custom chunking. Keep the corpus in Cloud Storage with object-level IAM mirrored into index metadata.
Data and model estate in one place: Gemini on Agent Platform, Agent Search and RAG Engine, AlloyDB and BigQuery, Cloud Run for everything else.
| Component | Service | Evidence |
|---|---|---|
| Person | ||
| Cloud Run | prior, ga | |
| Identity-Aware Proxy | prior, ga | |
| Cloud Run | prior, ga | |
| Apigee (AI gateway policies) | prior, ga | |
| Gemini 3.8 Flash (Agent Platform) | source, ga | |
| Agent Search | source, ga | |
| RAG Engine (Agent Platform) | source, ga | |
| Cloud Storage | prior, ga | |
| Firestore | prior, ga | |
| Cloud Logging, Monitoring, Trace | prior, ga | |
| Evaluation service (Agent Platform) | confirm |
Customer view / judgement, not benchmark
Weigh what this customer values.
Remove providers that fail a hard constraint first: residency, data rights, identity estate, latency, operating owner, budget. Then weigh the rest. Scores are architect judgement on a 1-5 scale for the pattern as specified, recorded with a rationale. They are not benchmarks. A provider's consumption opportunity never changes a customer score.
Highest weighted: Vercel, 4.33 of 5.
Performance 4Time to value 5Speed to execution 5Cost per accepted result 4Exit cost 5Governance fit 3
Model and data layers stay swappable; enterprise identity, residency and audit depend on the chosen Marketplace partners.
Performance 5Time to value 4Speed to execution 4Cost per accepted result 5Exit cost 3Governance fit 4
Edge latency and low request cost; AI Search is remote-only in development and its permission model must be validated per corpus.
Performance 4Time to value 4Speed to execution 4Cost per accepted result 3Exit cost 3Governance fit 5
Managed retrieval with permission controls and residency options; index and model tie the workload to Agent Platform, which raises exit cost.
Performance 4Time to value 4Speed to execution 3Cost per accepted result 3Exit cost 3Governance fit 5
Strong account-level controls; more services to wire than the primary platforms.
Performance 4Time to value 4Speed to execution 3Cost per accepted result 3Exit cost 3Governance fit 5
Best fit where documents already live in SharePoint and identity is Entra; search unit pricing is step-shaped.
Performance 3Time to value 3Speed to execution 4Cost per accepted result 4Exit cost 5Governance fit 2
Lowest lock-in; identity, audit and guardrails are custom work.
Provider ledger / shown after the choice
What grows on Google Cloud.
These meters are the provider team’s consumption view. They explain the cloud workload; they never raise a customer score.
- Gemini input and output tokens
- Cloud Run vCPU-seconds and requests
- Agent Search queries and index storage
- AlloyDB or Cloud SQL instance hours
- BigQuery slots or bytes scanned
Take it with you / generated from one specification
Exact views, and a running start.
Google Cloud views
- SVG diagram for decks and documents
- diagrams.net file editable, shared IDs
- React Flow JSON for your own canvas
- Mermaid flowchart
Every provider
- Structurizr workspace C4 containers, one deployment view per provider
- Journey sequence (Mermaid)
One-click template / ai-architect-answer/v1
Grounded answers on Cloud Run with Gemini 3.8 Flash on Agent Platform
Model: gemini-3.8-flash on the global endpoint. Auth: Service account Application Default Credentials.
Evidence: local-tested: conformance suite, installed @google/genai client options and a local server run in fixture mode. Not yet verified: Container build, Cloud Run Button flow and a live Gemini call. The button builds from the engine’s main branch.
Customer ledger / assumptions you can move
Cost per accepted result.
Every attempt is charged, so failed runs count. Value comes only from accepted results. Illustrative assumption for a 120-analyst team; replace with measured volume and baseline time before any business case.
Primary metric: Accepted cited answers per analyst hour. Answers marked forwarded or reused without edit, divided by analyst hours spent in the workspace.
- Monthly value
- €54,000
- Monthly cost
- €2,680
- Net
- €51,320
- Cost per accepted result
- €1.12
Experience / drawn from the specification
What the person sees, including no.
Screens are rendered from structured blocks in the catalog, so a refusal or an exception is designed before it happens. Values marked fixture are illustrations, not measurements.
What do you need to know?
Searches only documents you can open.
Three active contracts allow termination for convenience below 60 days: Northwind Logistics (30 days) [1], Contoso Packaging (45 days) [2], and Fabrikam Metals (30 days, only after year one) [3].
- MSA Northwind s.14.2
- Contoso SoW-7 s.9
- Fabrikam MSA s.11.1
No passage you can access answers this question.
- Closest: Procurement policy 2026, section 4 (termination rules)
- Ask the contracts team to add the 2024 amendments
Notice period in the Northwind MSA?
30 days for convenience, 10 days for cause [1].
Team workflow / who carries what
Every lane leaves evidence.
| Lane | Responsibility | Evidence it leaves |
|---|---|---|
| Customer | Ask, read the cited answer, forward or flag it | Forward and flag events in the ledger |
| AI | Draft from passages; return passage IDs per claim | Citation coverage per answer |
| System | Filter by permission, enforce budget, log every call | Trace with scope and token count |
| Reviewer | Knowledge owner reviews flagged answers weekly and fixes the source | Closed flags with source revision |
| Operator | Watch refusal rate, cost per accepted answer and stale-source alerts | Weekly dashboard export |
Release evaluations
What must pass before anyone ships.
Citation precision
>= 0.95 of claims map to a retrieved passage that supports them
Frozen set of 200 questions with expected passages; scorer checks each claim span.
Must-refuse accuracy
>= 0.90 refusal on 50 out-of-corpus questions
Negative set built from questions owners confirmed are unanswerable.
Permission leak
0 passages returned outside the caller's permissions
Adversarial personas with restricted groups.
Answer latency
p95 <= 6 s end to end
Gateway telemetry over the frozen set.
Risks and controls
What would go wrong first.
Answers cite superseded document versions.
Source revision stored per citation; re-index triggers review of affected answers.
Instructions hidden in a document steer the model.
Passages are quoted as data; tool use disabled; output limited to passage-backed claims.
Retrieval bypasses permissions.
Permission filter applied in the index query, tested with adversarial personas.
Image recipes / concept art only
When to generate, and when never to.
16:9 · gpt-image-2 or a current Gemini image model
Customer moment for a proposal cover or portal hero.
Cinematic editorial photograph, late afternoon light, shallow depth of field. Find a defensible answer in policies, research reports and filings without reading forty documents. The moment just after the accepted result: An answer the analyst forwards without rework, with each claim traceable to a passage they are permitted to see. Real workplace, restrained palette of deep navy, warm paper and one cobalt accent. Natural expressions, no screens facing the camera. Concept art only. Do not render product logos, cloud icons, UI text or architecture labels; compose exact labels from the specification in code.
3:2 · gpt-image-2 (consistent characters across panels)
Team workflow storyboard for a workshop or executive briefing.
Five-panel storyboard, clean ink line with flat colour, consistent characters across panels. Panels follow the lanes customer, ai, system, reviewer, operator: (1) Ask, read the cited answer, forward or flag it (2) Draft from passages; return passage IDs per claim (3) Filter by permission, enforce budget, log every call (4) Knowledge owner reviews flagged answers weekly and fixes the source (5) Watch refusal rate, cost per accepted answer and stale-source alerts Each panel shows hands and decisions, not screens. Concept art only. Do not render product logos, cloud icons, UI text or architecture labels; compose exact labels from the specification in code.
Do not generate
Any image a customer will read for component names, data classes or provider services.
Do not generate. Export the compiled view for the chosen provider. A generated architecture image can invent services, misspell names and draw icons the vendor never published.
A design, not a deployment: no realization on this page carries a provider receipt. Generated from the pattern source at revision c43825efda38. Apache-2.0 for catalog content and exports; provider names are trademarks of their owners; Google Cloud icons retain Google rights.
Next pattern: In-product copilot