Role aggregate
A stable public role hypothesis groups listings that describe the same opportunity.
JobCache CoreBlueprint · careervector.blueprint-graph/v5
Graph explorer
Twenty composable top-level diagrams plus recursively nested 21×21 semantic layers. Nodes descend; boundary arrows cross to neighbours.
Nested layer 1 · data · evidence · runtime
Role, ad, fact, and evidence corpus
CrateDB keeps canonical public knowledge compact while preserving source disagreement, provenance, temporal validity, and retrieval structures.
- Source ads to Role aggregate: cluster
- Source ads to Facts + evidence timeline: supports
- Facts + evidence timeline to Role projection: reduce without erasing
- Role aggregate to Role projection: aggregate
- Role projection to Vector + lexical retrieval: index
- Vector + lexical retrieval to Product-safe result: retrieve
- Role projection to Product-safe result: serve
Source ads
Canonical listing records retain source identity, fingerprints, and role membership.
CrateDBFacts + evidence timeline
Field assertions remain attributable, time-bounded, and able to disagree.
Open layer ↘Role projection
Reducers choose a useful current answer without erasing alternate observations.
JobCache CoreVector + lexical retrieval
One content-addressed embedding per recipe complements indexed structured search.
CrateDBProduct-safe result
Search and evidence APIs publish bounded public truth without App or workspace identity.
JobCache Core01 · process · mixed Role aggregate A stable public role hypothesis groups listings that describe the same opportunity.
JobCache Core02 · store · as-built Source ads Canonical listing records retain source identity, fingerprints, and role membership.
CrateDB- 03 · store · mixed Facts + evidence timeline
Field assertions remain attributable, time-bounded, and able to disagree.
Open deeper layer ↘ 04 · process · mixed Role projection Reducers choose a useful current answer without erasing alternate observations.
JobCache Core05 · store · target Vector + lexical retrieval One content-addressed embedding per recipe complements indexed structured search.
CrateDB06 · boundary · mixed Product-safe result Search and evidence APIs publish bounded public truth without App or workspace identity.
JobCache Core
Flows
- Source ads → Role aggregate cluster
- Source ads → Facts + evidence timeline supports
- Facts + evidence timeline → Role projection reduce without erasing
- Role aggregate → Role projection aggregate
- Role projection → Vector + lexical retrieval index
- Vector + lexical retrieval → Product-safe result retrieve
- Role projection → Product-safe result serve
Canonical architecture note
Open note ↗Internals
Purpose
Hold and serve the current public model of the job market. Internals is a blueprint label
for JobCache Core and its immediate processing boundary; it is not a third product.
Canonical model
role → ad → workspace job
Core owns the role and ad rungs. CareerVector owns private workspace job state. Role
facts aggregate evidence across ads without erasing disagreement or analyzing identical text
twice.
CrateDB authority
- Canonical current roles and ads.
- Structured facts with evidence and temporal validity.
- Embeddings and search projections.
- Role/ad reconciliation and revision lineage.
- Shared reusable enrichment results and public aggregate outputs.
Raw observations remain in object storage. Unfinished pipeline work remains in rebuildable operational state. Product-local session and workspace state remains in product D1/Yjs.
Invariants
- CrateDB is the only canonical public-corpus database.
- Publication is idempotent, revisioned, and traceable to immutable observations.
- A role aggregate never hides conflicting ad evidence.
- Product-local identifiers do not enter Core.
- One-node CrateDB is an explicit resource constraint, not an excuse for a second truth store.
- Customer reads take precedence over asynchronous publication pressure.
Completion evidence
Schema and query budgets for 100,000 ads; temporal role/ad reconstruction; vector and hybrid search benchmarks; publication atomicity; lineage queries; and privacy-boundary tests.