Multimodal RAG Platform · View 08 of 16
Data Architecture
Four zones by ownership and rebuildability. Only one of them is irreplaceable.
Copy
PNG
PDF
⋯
Editable source
SVG
draw.io
All views
Knowledge data platform
Knowledge data platform
Raw zone — immutable, system of record
Raw zone — immutable, system of record
Original Files
byte-identical
Original Files...
Ingest Manifest
checksum, ACL, source
Ingest Manifest...
Legal Hold Set
retention locked
Legal Hold Set...
Derived zone — reproducible, disposable
Derived zone — reproducible, disposable
Parsed Documents
layout JSON
Parsed Documents...
Extracted Assets
images, tables
Extracted Assets...
Transcripts
word timings
Transcripts...
Chunk Records
text + provenance
Chunk Records...
Serving zone — query-time, rebuildable
Serving zone — query-time, rebuildable
Vector Index
per-tenant namespace
Vector Index...
Lexical Index
Lexical Index
Analytical Tables
typed, SQL-queryable
Analytical Tables...
Semantic Cache
TTL bound
Semantic Cache...
Operational
Operational
Metadata DB
row-level security
Metadata DB...
Audit Log
write-once
Audit Log...
Eval & Feedback Store
golden set
Eval & Feedback Store...
Cold Archive
glacier class
Cold Archive...
Erasure Service
subject request
Erasure Service...
reparse on parser upgrade
reparse on parser upgrade
index build
index build
after 90 days
after 90 days
tombstone then purge
tombstone then purge
Data Architecture — Zones and Ownership
Data Architecture — Zones and Ownership
Data store
Data store
External / third party
External / third party
batch
batch
failure / alternate
failure / alternate
Only the raw zone is irreplaceable. Everything below it can be rebuilt from it.
Only the raw zone is irreplaceable. Everything below it can be rebuilt from it.
v 1.0 · owner Data Architecture
v 1.0 · owner Data Architecture
Text is not SVG - cannot display
Raw zone
Byte-identical originals, versioned, never mutated
Ingest manifest carries checksum, ACL and source
This is the only data that cannot be regenerated
Derived and serving
Parsed documents, extracted assets, transcripts, chunk records
Vector, lexical and analytical indexes plus semantic cache
All of it is rebuildable from the raw zone
Governance
Row-level security in the metadata database
Write-once audit log
Erasure requests tombstone the manifest, then purge downstream
Cold archive after 90 days
