Seed taxonomy as SMB bootstrap

@@ -14,7 +14,7 @@ cites: []
paired_with: seed-taxonomy-as-smb-bootstrap.es.md
---

Every tenant deployment begins with a **seed taxonomy**: a compact, hand-tunable four-part structure that forms the initial scaffold of the per-tenant knowledge graph. The four parts are Archetypes, Chart of Accounts, Domains, and Themes. Each entity carries gravity keywords — explainable keyword anchors that drive classification of incoming content.
Every tenant deployment begins with a **seed taxonomy**: a compact, hand-tunable four-part structure that forms the initial scaffold of the per-tenant [[knowledge-graph-grounded-apprenticeship|knowledge graph]]. The four parts are Archetypes, Chart of Accounts, Domains, and Themes. Each entity carries gravity keywords — explainable keyword anchors that drive [[service-content|classification of incoming content]].

## The four parts

@@ -28,7 +28,7 @@ Every tenant deployment begins with a **seed taxonomy**: a compact, hand-tunable

## The gravity_keywords mechanism

Classification of new content into the taxonomy uses keyword matching rather than embedding similarity. When a document arrives, the extraction service counts keyword matches against each Archetype, Chart of Accounts profile, Domain, and Theme entity. The entity with the highest match count is the proposed classification.
Classification of new content into the taxonomy uses keyword matching rather than embedding similarity. When a document arrives, the [[service-extraction|extraction service]] counts keyword matches against each Archetype, Chart of Accounts profile, Domain, and Theme entity. The entity with the highest match count is the proposed classification.

This approach is deliberate. Keyword-based classification is explainable: "this document was classified as Real Estate because the terms Leasing, Office, and Industrial matched" is a statement an operator can read and verify. An embedding similarity score is not. Keyword classification is auditable — the classification path is reproducible across model versions, which matters for regulatory record-keeping. And it is hand-tunable: an operator who sees a misclassification edits the gravity keywords list. Retraining an embedding model is not a practical operation for a small business.

@@ -36,7 +36,7 @@ Embedding similarity is provided at a separate layer to augment keyword classifi

## Provisioning a new tenant

When a new tenant is provisioned, the operator selects a Vertical Seed Pack (see [[vertical-seed-packs-marketplace]]) appropriate to their industry. The service imports the pack's JSON files into the per-tenant graph. The operator then customizes the result — adding, editing, or removing entities — using the TUI. A typical small business is intended to complete the review and customization in approximately 30 minutes.
When a new tenant is provisioned, the operator selects a Vertical Seed Pack (see [[vertical-seed-packs-marketplace]]) appropriate to their industry. The [[service-content|service]] imports the pack's JSON files into the per-tenant graph. The operator then customizes the result — adding, editing, or removing entities — using the [[tui-corpus-producer|TUI]]. A typical small business is intended to complete the review and customization in approximately 30 minutes.

## Structural difference from enterprise ontology approaches

@@ -46,7 +46,7 @@ The platform's seed taxonomy optimizes for actionability for one specific custom

## Relationship to the knowledge graph

The seeded taxonomy becomes the initial structure of the per-tenant knowledge graph in `service-content`. Every Archetype, Chart of Accounts profile, Domain, and Theme entity is a graph node. As the deployment operates, new entities discovered during inference (with accepted verdicts) are added to the graph, growing the taxonomy organically from actual use.
The seeded taxonomy becomes the initial structure of the per-tenant knowledge graph in [[service-content]]. Every Archetype, Chart of Accounts profile, Domain, and Theme entity is a graph node. As the deployment operates, new entities discovered during inference (with accepted verdicts from the [[compounding-doorman|Doorman]]) are added to the graph, growing the taxonomy organically from actual use.

The [[knowledge-graph-grounded-apprenticeship]] pattern depends on this seeded graph: the graph provides the grounding context for every subsequent inference request.