AGENT SKILL

0.2.4-understand-colbert_late_interaction

[0.2.4] Core architectural principle deferring interaction between query and document token representations until scor

License 许可
MIT
最近更新
2026-10-09
Stars
★ 0
Forks
0

01 前言 Front Matter

name
0.2.4-understand-colbert_late_interaction
description
[0.2.4] Core architectural principle deferring interaction between query and document token representations until scor

03 SKILL.md 正文

understand-colbert_late_interaction

CALL NUMBER: deep_dense_retrieval.colbert_late_interaction DEFINITION: Core architectural principle deferring interaction between query and document token representations until scoring time, enabling efficient per-token similarity computation without full cross-encoding of query-document pairs.

Invoke this skill to understand colbert_late_interaction down to its primitives. The RELATIVE ROOT below is the least-fixed-point closure of everything it bundles from — the full import cone, grouped by the lib each prim comes from. Projected from a prover-typed KB (MAP/SWI-Prolog consistency gate): every reference below resolves.

THE RELATIVE ROOT (the import cone, by lib)

from deep_dense_retrieval

  • colbert_maxsim_operator (d1): Maximum similarity operator that for each query token finds the highest-scoring document token via cosine similarity, then aggregates across query tokens to produce document relevance score.
  • colbert_reductional_scoring (d1): Late scoring mechanism reducing query-document interaction to token-level maximum similarities then aggregating, contrasted with full cross-encoder joint encoding.
  • colbert_token_aggregation (d1): Aggregation function combining per-token similarity scores (typically sum or mean over query tokens) to produce final document relevance score in late interaction framework.
  • colbert_token_embedding (d2): Dense contextualized vector representation produced per token by an encoder model, preserving token identity and positional context for late interaction scoring.
  • colbert_cross_encoder_comparison (d2): Comparison metric distinguishing ColBERT from cross-encoder reranking: ColBERT scores via token-level max-sim aggregation whereas cross-encoder produces single joint representation per pair.
  • colbert_document_encoder (d3): Encoder producing per-token document embeddings from document passages, typically run offline and indexed for fast retrieval against query embeddings.
  • colbert_query_encoder (d3): Encoder producing per-token query embeddings from input query tokens, enabling efficient single-pass encoding with late interaction against document tokens.
  • colbert_representation_rank (d4): Rank dimension of per-token embeddings determining expressiveness of token-level similarity computation in late interaction framework.

Projected from the retrieval augmented generation architecture patterns KB (207 concepts / 225 relations) — consistency-typed by MAP; the facet list after the colon IS the cross-lib dependency web.

(leaf — this is an actual skill.)

04 同仓库更多 Skills