The Evidence Path

Retrieval cascade

1 of 12 · Retrieval · 44 named systems

A retrieval cascade trades progressively greater interaction cost for better evidence ranking and collective utility.

Retrieval cascade: architecture comparisonEach lane reads from input to output.

Sparse

  1. inputQuery terms
  2. transformWeight or expand
  3. indexPostings lookup
  4. outputSparse candidates
  • TF-IDF
  • Query likelihood
  • BM25
  • BM25F
  • RM3
  • DeepCT
  • doc2query
  • docT5query
  • DeepImpact
  • uniCOIL
  • COIL
  • SPLADE
  • SPLADE v2
  • SPLADE++

Dense

  1. inputQuery
  2. encodeQuery vector
  3. searchVector search
  4. outputDense candidates
  • DPR
  • ORQA
  • ANCE
  • RocketQA
  • Contriever
  • Condenser
  • coCondenser
  • RetroMAE
  • SimLM
  • GPL
  • E5
  • INSTRUCTOR
  • GTR
  • BGE
  • GritLM
  • DRAGON
  • ReasonIR

Late interaction

  1. encodeQuery tokens
  2. retrieveToken candidates
  3. scoreMaxSim
  4. outputFine-grained rank
  • ColBERT
  • ColBERTv2
  • PLAID
  • XTR
  • CITADEL

Rerank and select

  1. inputCandidate union
  2. rankJoint scoring
  3. selectCoverage and risk
  4. outputEvidence set
  • monoT5
  • RankT5
  • RankGPT
  • RankRAG
  • HYRR
  • SetR
  • LDAR
  • NEST

Read the full explanation