Skip to content

EntiHub Documentation

RAG / AI Grounding Taxonomy

RAG / AI Grounding Taxonomy

Beyond glossary text, a grounding taxonomy is the match layer for RAG: canonical nodes, aliases, and parent categories that stabilize retrieval and agent grounding.

Why this is needed

Chat slang and abbreviations (BP, acct) break naive retrieval.
Embedding indexes need a governed synonym surface.
Complements Corporate Vocabulary with search/match behavior.

Proposed entities in EntiHub

Governance settings

Setting	Recommended value	Why
`approvalRequired`	`true`	Taxonomy changes affect every RAG consumer.
`historyEnabled`	`true`	Track drift of synonyms over time.
`allowSameAuthorApproval`	`false`	Maker/checker separation for sensitive changes.
`deletedRetentionDays`	`90` (or higher)	Soft-delete retention for auditability.

Entity definition (YAML sketch)

name: grounding_term
displayName: Grounding term
domain: Knowledge
description: Grounding term
version: "1.0.0"
approvalRequired: true
historyEnabled: true
deletedRetentionDays: 90
displayRefColumn: code
uniqueConstraints:
  - name: uq_grounding_term_code
    columns: [code]
aliases:
  enabled: true
similarity:
  enabled: true
  topN: 10
  minScore: 0.3
  columns:
    - name: label
      weight: 0.5
      matchMode: fuzzy
    - name: aliases
      weight: 0.5
      matchMode: contains
columns:
  - name: code
    dataType: string
    nullable: false
    maxLength: 64
  - name: label
    dataType: string
    nullable: false
    maxLength: 128
  - name: parent_code
    dataType: string
    nullable: true
    maxLength: 64
  - name: aliases
    dataType: string
    nullable: true
    maxLength: 512
  - name: category
    dataType: string
    nullable: false
    maxLength: 64
  - name: embedding_hint
    dataType: string
    nullable: true
    maxLength: 256
  - name: is_active
    dataType: bool
    nullable: false

YAML sketches

Sketches highlight the important fields and governance flags. Trim or extend columns to match your domain. See Setup Entity for the full YAML reference.

Sample data

code	label	parent_code	aliases	category	is_active
BP	Business partner	PARTY	BP, biz partner	Party	true
CUST	Customer	PARTY	client, account	Party	true
PO	Purchase order	DOC	PO, P.O.	Document	true

Benefits for the company

More stable RAG mapping for slang and abbreviations.
Similarity helpers speed steward curation.
Consumption views feed vector pipelines.
Works together with Corporate Vocabulary.