Multimodal RAG Platform · View 04 of 16
Container Architecture
C4 Level 2. Deployable units, their technology, and the synchronous query path between them.
Copy
PNG
PDF
⋯
Editable source
SVG
draw.io
All views
Kubernetes cluster — application namespace
Kubernetes cluster — application namespace
Edge & API
Edge & API
API Gateway
Envoy, OIDC
API Gateway...
Query API
FastAPI, 3 replicas
Query API...
Ingestion API
FastAPI, presigned PUT
Ingestion API...
Admin API
tenants, policy
Admin API...
Ingestion workers — autoscaled by queue depth
Ingestion workers — autoscaled by queue depth
Parse Worker
layout, OCR
Parse Worker...
Vision Worker
VLM caption, GPU
Vision Worker...
Media Worker
ASR, keyframes
Media Worker...
Index Worker
chunk, embed, write
Index Worker...
Retrieval & generation — synchronous path
Retrieval & generation — synchronous path
Query Orchestrator
plan and assemble
Query Orchestrator...
Retriever
hybrid fan-out
Retriever...
Reranker
cross-encoder, GPU
Reranker...
Model Gateway
route, cache, budget
Model Gateway...
Managed data services
Managed data services
Vector Index
HNSW, per-tenant
Vector Index...
Lexical Index
BM25
Lexical Index...
Analytical Tables
Parquet + SQL
Analytical Tables...
Object Store
raw + derived
Object Store...
Metadata DB
PostgreSQL
Metadata DB...
Work Queue
per-stage topics
Work Queue...
Claude API
generation + vision
Claude API...
Embedding API
Embedding API
Speech-to-Text
Speech-to-Text
Identity Provider
OIDC
Identity Provider...
HTTP/2
HTTP/2
gRPC
gRPC
search
search
top-k ANN
top-k ANN
BM25
BM25
SQL
SQL
assemble
assemble
citations on
citations on
enqueue
enqueue
Container Architecture (C4 Level 2)
Container Architecture (C4 Level 2)
Interface / broker
Interface / broker
Application we own
Application we own
Data store
Data store
Queue / topic
Queue / topic
External / third party
External / third party
synchronous
synchronous
event / async
event / async
Synchronous query path shown. Workers lease from the queue and write the indexes — that path is view 05.
Synchronous query path shown. Workers lease from the queue and write the indexes — that path is view 05.
v 1.0 · owner Solution Architecture
v 1.0 · owner Solution Architecture
Text is not SVG - cannot display
Edge and API
API gateway — Envoy, OIDC verification, per-tenant quota
Query API — FastAPI with server-sent events for streaming
Ingestion API — issues presigned uploads, never proxies bytes
Workers
Parse, vision, media and index workers scale on queue depth
Vision and rerank workers are the only GPU consumers
Every worker is idempotent and resumable
Data services
Vector index with per-tenant namespaces
Lexical BM25 index for identifiers and exact terms
Analytical tables for aggregation over extracted tables
Object store, metadata database, work queue
