Connected Knowledge

Vector retrieval is necessary for most document search. It is not enough when the answer depends on source membership, cited neighbors, topic summaries, or conflicting requirements across documents.

Connected Knowledge starts with the same knowledgeBase() used by Retrieval and RAG. Add only the layer that pays for itself:

For PostgreSQL deployments, use postgresStorage({ dimensions, sparseDimensions? }) from @use-crux/postgres, run storage.setup.apply() during provisioning, and pass the bundle to knowledgeBase(). Records and pgvector rows share one pool; Runtime Engine tables remain isolated in crux_runtime.

plain knowledge base -> views -> relations -> communities/global search -> assertions

The Progressive Path

Start with a knowledge base when the model needs chunks that match a question. Add views when the corpus has typed metadata and users ask within a subset. Add relations when nearby evidence matters even when it does not match the query text. Add communities and globalSearch() when the question should scan themes across the corpus instead of starting from vector hits. Add assertions when the corpus contains propositions that need evidence, conflict handling, or resolution before prompting.

import { knowledgeBase } from "@use-crux/core/knowledge";
import { inMemoryStorage } from "@use-crux/core/storage";

const docs = knowledgeBase({
  id: "docs",
  storage: inMemoryStorage(),
  embeddings: dense,
});

await docs.index([
  {
    namespace: "docs",
    sourceId: "policy.md",
    content: "Refunds are available within 30 days.",
    metadata: { status: "published" },
  },
]);

const hits = await docs.retriever({ limit: 5 }).retrieve("refund policy");

When Each Level Pays Off

What Is Deterministic

These parts are deterministic after their inputs exist:

• view membership, revisions, and namespace scoping
• structural graph traversal over hierarchy and sequence
• deterministic relate({ run }) and assertions({ run }) stages
• relation expansion through the published graph
• assertion resolution from explicit assertion relations and deterministic policies
• community clustering and report reuse checks
• global-search preflight, batching, and reduce ordering

These parts make explicit model calls:

• model-backed relate()
• relateEntities({ model })
• model-backed assertions()
• communities({ model }) report generation, and its generic entity mapping when no authored entity mapping is present
• globalSearch({ model }) map calls over selected reports
• model-backed assertion resolution policies

Crux does not make hidden connected-knowledge generation calls for a plain knowledge base.

Next Steps