Saved.archi Ideas from Cloudflare + Tempo Notes

Document

Repository research note synced from research/saved-archi-cloudflare-tempo-ideas.md

Saved.archi Ideas from Cloudflare + Tempo Notes

Source notes

cloudflare-ai-search.md
cloudflare-git-artifacts.md
mpp-tempo-for-saved-works-archi-akiyama.md

Core takeaway

The cleanest architecture for saved.archi is:

Cloudflare AI Search for hot-path retrieval
Tempo/MPP tools for enrichment and long-tail research writes
Cloudflare Artifacts for versioned research state, provenance, and reversible data changes

That split matches the product well:

users need fast hybrid retrieval across works, books, bureaus, materials, briefs, jobs, and their own notes
the corpus still needs ongoing enrichment and cleanup
provenance is already one of the strongest differentiated features in the product

Best ideas for saved.archi

1. Hybrid research search across the whole corpus

Cloudflare AI Search is a strong fit for saved.archi because the corpus is graph-shaped but users still search in messy natural language.

What it could power:

one search surface across works, books, architects, bureaus, briefs, materials, jobs, and user notes
better results when metadata is wrong or thin
natural-language retrieval for prompts like "postwar civic concrete buildings in Japan" or "brick museums with strong landscape integration"
better recall for architecture-specific language where BM25 alone is too brittle

Why it matters here:

wrong typology tags stop being fatal because vector retrieval can still match the description
exact names like architect names, building names, movements, and niche theory terms still work because BM25 stays in the loop
metadata boosting can bias results toward period, country, material, typology, architect, bureau, and public accessibility

Concrete product fits:

/search: unify keyword + semantic retrieval
/ask: use AI Search as the retrieval layer beneath graph-query interpretation
/compare: search for candidate works without depending on rigid field matches
/research/[slug]: retrieve related works, books, and briefs from the same conceptual cluster

Implementation shape:

shared namespaces for works, books, briefs, materials, jobs, source-atlas
per-user namespace like user-{id} for saved items, notes, draft collections, private briefs
one cross-namespace search call for "corpus + my notes"

2. AI Search as a fix for corpus quality, not just search UX

One of the strongest ideas in the notes is using hybrid retrieval to soften data-quality issues.

For saved.archi, that means:

typology mismatches stop poisoning search
material pages can retrieve relevant works even when material tags are shallow
bureau pages can surface works tied by description and context, not just exact refs
plan and drawing research can still be found through descriptive text even when the structured field is sparse

This is especially relevant for:

underfilled material graphs
works with incomplete material tags
buildings with noisy typology labels
spatial-content pages where the best clue lives in a caption, source note, or brief rather than a canonical field

3. Retrieval boosting tailored to architecture

Cloudflare AI Search metadata boosting maps unusually well to architecture filters.

High-value boost fields:

architect
bureau
typology
material
country
city
year range
public accessibility
plan availability
carbon/profile coverage

Examples:

boost country: Japan and year: 1945-1975 for "postwar Japanese civic buildings"
boost material: brick and typology: museum for "brick museums"
boost planAvailability: true for users doing precedent-study workflows

This could make the product feel much more like a research assistant and much less like a literal field filter.

4. Per-user research memory

The notes' "shared corpus + user namespace" pattern is a very strong match for saved.archi.

This suggests:

per-user notes become searchable with the public corpus
saved works and books can influence ranking
private brief drafts and compare sets can become first-class retrieval inputs
"continue where I left off" becomes a search problem, not a bespoke feature per surface

This is likely the cleanest path toward the "saved" in saved.archi actually becoming product-defining.

5. Versioned provenance for research and corrections

Cloudflare Artifacts is not right for the main app database, but it is a strong fit for versioned research state.

Best uses in saved.archi:

version every manual correction to typology, material, architect credit, and plan-source attribution
keep reviewable histories of enrichment runs
diff a record before and after an AI enrichment pass
revert bad updates cleanly
keep a human-readable evidence trail for why a field changed

This would strengthen Source Atlas directly:

each work could have a change history for key fields
each claim could point to the evidence set that justified the latest value
field-level provenance becomes inspectable instead of implicit

Good artifact candidates:

work research dossier
source-review packet
plan-source packet
material-profile packet
brief draft history
compare session state

6. Forkable research collections and briefs

The strongest Artifacts idea from the notes is "collections as repos." Adapted to saved.archi, that becomes:

forkable precedent collections
branchable research briefs
draft vs published research states
visible lineage between one person's precedent set and another person's adaptation

That is a strong fit for architecture school, office research, and collaborative precedent work.

Potential product expressions:

fork this brief
fork this compare set
fork this precedent collection
show what changed between my version and the original

This is more differentiated than generic bookmarks and pairs naturally with books, works, plans, and notes.

7. Tempo/MPP as the enrichment engine

Tempo/MPP fits saved.archi best on the write side, not the read side.

Best uses:

Exa for source discovery
Perplexity for fast grounded fact checks
Diffbot KG for structured entity correction
Firecrawl for scraping bureau pages, archive pages, and plan sources
OpenRouter models for synthesis and normalization

Good enrichment jobs for saved.archi:

typology correction suggestions
material backfill suggestions
bureau enrichment
architect enrichment
plan-source discovery
source-atlas evidence gathering
Instagram and social link discovery
brief evidence expansion

Important principle:

MPP/Tempo should enrich the corpus and write results back
AI Search should serve retrieval on the hot path

Do not put pay-per-call Tempo services in the main user-serving search loop.

8. A plan-source discovery pipeline

The current product already moved toward plan and drawing coverage. The notes suggest a stronger pipeline:

Exa discovers candidate pages likely to contain plans, sections, axons, or drawings
Firecrawl captures the relevant bureau or publication pages
Perplexity helps summarize what kind of drawing exists and why it matters
the result writes into a versioned plan-source packet

That would make spatial content much more systematic and would build directly on work already underway.

9. A typology and material correction queue

The most actionable MPP idea in the notes is a review queue rather than full auto-mutation.

For saved.archi, build:

a batch job that runs works through Diffbot + Perplexity + Exa as needed
a review queue for suggested typology corrections
a second queue for material enrichment suggestions
approval writes into the canonical corpus
the full decision history is stored in an Artifact packet

That gives you:

faster corpus cleanup
safer AI-assisted correction
reversible edits
explicit provenance

10. A better retrieval layer for books

The book graph is already growing. AI Search makes it much more useful if books become part of the same retrieval system as works and briefs.

Useful outcomes:

"what should I read next" can pull from books + works + briefs in one result set
architect pages can retrieve books by conceptual relevance, not just exact links
work pages can surface books about the building type, architect, place, or material family
brief generation can retrieve books alongside works and source packets

This likely gives more value than adding more isolated book metadata fields.

Priority order

Recommended now

Build a typed enrichment review queue powered by Tempo/MPP
Use that queue to fix typology and material gaps
Stand up AI Search on a limited corpus slice for works + books + briefs + user notes
Use metadata boosting tuned for architecture fields

Recommended next

Add Artifact-backed provenance packets for work-level enrichment and source review
Make plan-source discovery a first-class enrichment job
Add forkable research collections or briefs

Recommended later

Move more of /ask retrieval onto AI Search
Support per-user research memory and note retrieval
Add diffable brief history and branchable draft/public states

What not to do

Do not use Artifacts as the primary transactional app database
Do not route the main search hot path through Tempo/MPP tools
Do not auto-accept AI corrections without a review layer
Do not treat AI Search as a substitute for canonical data cleanup; it should reduce pain, not replace curation

Most promising near-term builds

If choosing the top three ideas to adapt right away:

**Typology + material correction queue**

Use Tempo/MPP tools to generate reviewable corrections for weak records.

**Hybrid search pilot**

Stand up AI Search for works + books + briefs + user notes and test it behind /search or /ask.

**Artifact-backed Source Atlas packets**

Version work research, plan-source research, and enrichment decisions so bad runs are reversible and provenance becomes even stronger.

Why this matters

The notes suggest a clean division of labor that fits saved.archi unusually well:

Cloudflare AI Search makes retrieval better
Tempo/MPP makes the corpus smarter
Artifacts makes research and provenance durable

That combination would push saved.archi beyond a static corpus into a living research system with fast retrieval, reviewable enrichment, and versioned architectural evidence.