Multimodal RAG Platform · View 15 of 16
Critical Flow — Question to Cited Answer
The end-to-end request as ordered messages across eight participants, including the abstain path.
Copy
PNG
PDF
⋯
Editable source
SVG
draw.io
All views
User
User
Web App
Web App
Query API
Query API
Guardrails
Guardrails
Retriever
Retriever
Reranker
Reranker
Model Gateway
Model Gateway
Claude API
Claude API
1. asks a question
1. asks a question
2. POST /v1/query
2. POST /v1/query
3. screen input
3. screen input
4. clean, PII masked
4. clean, PII masked
5. hybrid search, k=50
5. hybrid search, k=50
6. dense + BM25 + SQL, tenant filtered
6. dense + BM25 + SQL, tenant filtered
7. fused candidates
7. fused candidates
8. top 8 with parents
8. top 8 with parents
9. assemble context
9. assemble context
10. documents + question
10. documents + question
11. answer with cited spans
11. answer with cited spans
12. normalised result
12. normalised result
13. grounding gate
13. grounding gate
14. no supporting span, abstain
14. no supporting span, abstain
15. streamed answer + citations
15. streamed answer + citations
16. renders with source links
16. renders with source links
Critical Flow — Question to Cited Answer
Critical Flow — Question to Cited Answer
Semantic cache hit and token streaming are omitted for clarity.
Semantic cache hit and token streaming are omitted for clarity.
v 1.0 · owner Solution Architecture
v 1.0 · owner Solution Architecture
Text is not SVG - cannot display
Ordering that matters
Guardrails screen before any retrieval spends money
Retrieval fans out concurrently and is tenant-filtered at source
Reranking happens before context assembly, not after
Grounding gate
Runs after generation, before anything reaches the user
No supporting span means abstain, drawn here as an explicit path
An abstention is logged and counted, not hidden
Latency budget
Retrieval and rerank: 400 ms target
First token: under 1.5 s
Complete answer: under 6 s at the 95th percentile
