The Evidence Path

Structured, visual, and code RAG

9 of 12 · Multimodal · 17 named systems

Non-text RAG must preserve the structure, modality, and source coordinates needed for retrieval, reasoning, and attribution.

Structured, visual, and code RAG: architecture comparisonEach lane reads from input to output.

Tables

  1. parseSchema and cells
  2. retrieveTable or row
  3. executeSQL or program
  4. citeCell evidence
  • TableRAG
  • T-RAG
  • RAG over Tables

Visual retrieval

  1. renderPage image
  2. encodePatch vectors
  3. scoreQuery-patch interaction
  4. localizePage or region
  • ColPali
  • ColQwen
  • Region-R1
  • SCAN

Multimodal generation

  1. retrieveText and visuals
  2. structureGraph or token set
  3. compressVisual evidence
  4. generateVLM answer
  5. attributeRegion evidence
  • MuRAG
  • RA-CM3
  • VisRAG
  • VDocRAG
  • MoLoRAG
  • MegaRAG
  • RobustVisRAG

Repository retrieval

  1. searchSymbols and text
  2. inspectCode context
  3. generateCandidate change
  4. executeTests and analysis
  5. repeatFailure-guided search
  • RepoCoder
  • Repoformer
  • CODEPROMPTZIP

Read the full explanation