Multimodal RAG Platform · View 07 of 16
Retrieval & Generation Flow
The query path from admitted question to verified, cited answer.
Copy
PNG
PDF
⋯
Editable source
SVG
draw.io
All views
Admit
Admit
User Question
User Question
Input Guard
injection, PII
Input Guard...
Plan
Plan
Query Rewrite
history-aware
Query Rewrite...
Route
modality + tool
Route...
Search
Search
Dense Search
top 50
Dense Search...
Lexical Search
top 50, exact IDs
Lexical Search...
Text-to-SQL
aggregation over tables
Text-to-SQL...
Fuse
Fuse
RRF Fusion
rank-based merge
RRF Fusion...
Cross-Encoder Rerank
top 8
Cross-Encoder Rerank...
Parent Expansion
small to big
Parent Expansion...
Generate
Generate
Context Assembly
cached prefix
Context Assembly...
Claude Opus 5
citations enabled
Claude Opus 5...
Verify
Verify
Grounding Gate
every claim cited
Grounding Gate...
Answer + Citations
page, bbox, timecode
Answer + Citations...
unsupported claim, one retry
unsupported claim, one retry
numeric question
numeric question
Retrieval & Generation Flow
Retrieval & Generation Flow
Person or role
Person or role
Security / platform
Security / platform
Application we own
Application we own
Decision point
Decision point
External / third party
External / third party
failure / alternate
failure / alternate
synchronous
synchronous
Abstain is a valid terminal state: no supporting span means no answer.
Abstain is a valid terminal state: no supporting span means no answer.
v 1.0 · owner Solution Architecture
v 1.0 · owner Solution Architecture
Text is not SVG - cannot display
Search
Dense and lexical searches run concurrently, 50 candidates each
Numeric questions additionally route to generated SQL over tables
Every search is filtered by tenant and by document ACL
Fuse and rank
Reciprocal rank fusion merges the channels without score calibration
Cross-encoder reranks to the top 8
Parent expansion returns the section, not the fragment
Verify
Generation runs with citations enabled
The grounding gate checks every claim against a retrieved span
Unsupported claim means one retry, then abstain
