Centralpoint

Centralpoint — On Premise Data Aggregation & AI Governance

Tab 4 · Unified Index Architecture

Vector / NLS Index Management.

One governed index serves two audiences at once: the AI that needs vector retrieval to reason, and the people who need semantic and lexical search to find documents, policies, and decisions. We do not build two parallel indexes. We do not force a tradeoff between AI discovery and forensic precision. Each record is mapped once — and only records explicitly permissioned into the AI index ever populate the AI-governed fields.

For the AI

Vector retrieval, governed at index time.

Every record granted permission into the AI index is embedded into the vector store with pre-index sensitivity filtering already applied. PII, PHI, and proprietary content never reach the embeddings. The retrieval surface the model sees is, by construction, the surface the customer has approved.

Pre-index sensitivity filtering — PII / PHI redacted before embedding
Per-record, per-module, per-audience inclusion control
Role-scoped retrieval inherits Active Directory / Entra ID permissions
Records can be removed from the AI index without being deleted from Centralpoint

For the people

Semantic + lexical search, in the same index.

The same index serves human search: Natural Language Search expands query intent semantically, and lexical/keyword retrieval delivers forensic precision when an auditor or attorney needs the exact phrase. One index, two retrieval modes, both audience-aware.

Natural Language Search with query expansion and synonym handling
Lexical / keyword retrieval for exact-match and forensic discovery
Federated across all aggregated sources — one source of truth
Cross-walks records that share taxonomy, metadata, or classification

Live Data · Index Operations

Live index status & operations.

Below: real-time view of the Centralpoint index management environment, rendered live from the governed DataSource.

AI Vector Index Inspector

Trace each ingested record → chunks → embedding dimensions.

0Records

0Avg Chunks / Record

0Avg Dims / Chunk

0Total Embeddings

Loading records...

Chunks

Click a record to inspect its chunks

▤ No record selected yet.

◎

What you're looking at — how AI reads this

An embedding is a numerical fingerprint of meaning. When a chunk of your text is ingested, the AI model converts it into a long list of numbers (typically 384, 768, 1,536, or 3,072 dimensions). Each number is a coordinate in “meaning space” — texts about similar topics end up at similar coordinates.

At query time the AI converts your question into the same kind of vector and uses cosine similarity to find the chunks whose vectors point in the closest direction. The heatmap below shows the first dimensions of this chunk’s vector — blue cells are negative values, red are positive. An L2 Norm near 1.0 is the sanity check that the embedding is real (and not zeros).

⎁ No vector selected. Pick a chunk from the Chunks tab.

AI Vector Index — Table Schemas

The three system tables that comprise the Centralpoint AI Vector Index. Together they store the configuration, the chunked text, and the embedding vectors for every record made retrievable by the AI.

cpsys_VectorIndexConfig15 fields

ConfigId	uniqueidentifier
SourceTable	nvarchar
ModuleId	uniqueidentifier
SourceAttributes	nvarchar
EmbeddingProvider	nvarchar
EmbeddingModel	nvarchar
ChunkSize	int
ChunkOverlap	int
MaxChunksPerRecord	int
IsEnabled	bit
TotalRecords	int
IndexedRecords	int
LastFullRebuild	datetime
LastIncrementalUpdate	datetime
CreateDate	datetime

cpsys_VectorChunks6 fields

ChunkId	uniqueidentifier
ModuleId	uniqueidentifier
DataId	uniqueidentifier
ChunkText	nvarchar(max)
ChunkIndex	int
CreateDate	datetime

cpsys_VectorElements3 fields

ChunkId	uniqueidentifier
DimIndex	int
Value	float

How the three tables relate

cpsys_VectorIndexConfig registers which modules get indexed and how (chunk size, embedding model). For each record in those modules, cpsys_VectorChunks stores one row per text chunk (linked back by ModuleId + DataId). For each chunk, cpsys_VectorElements stores one row per dimension of the embedding vector. A typical record produces 2–10 chunks; each chunk produces 384–3,072 element rows depending on the embedding model.

Computing diagnostics...

<b>AI Vector Search Index (Live interactive map of Index, Elements and Chunks</b>

Data Sources

Events

Reports Library

Staff

Font Resize

Contrast