Interaction Paradigms

Neural retrieval models can be organized along a single axis: when the query and the document are allowed to interact. This timing decision is the master trade-off in neural IR — it determines whether documents can be encoded offline (and therefore whether the model can serve first-stage retrieval at scale) and how much fine-grained relevance signal the scorer can capture. Moving “later” in interaction buys quality; moving “earlier” (toward joint encoding) costs the ability to pre-compute.

This is the broader frame behind the comparison table that lives inside the Late Interaction concept. The three named points on the spectrum each have a dedicated concept note: Bi-Encoder (no interaction), Late Interaction / ColBERT (late, token-level), and Cross-Encoder (early, joint).

The Spectrum

No Interaction        Late Interaction      Early Interaction
─────────────────────────────────────────────────────────────
  Bi-Encoder              ColBERT             Cross-Encoder
  (pre-compute)         (at query time,      (at query time,
                        token-level MaxSim)  full cross-attn)
     Fast ◄─────────────────────────────────────────► Accurate
Early interaction  (Cross-Encoder):  [Q + D] → Encoder → score
No interaction     (Bi-Encoder):     Q → enc → q; D → enc → d; score(q,d)
Late interaction   (ColBERT):        Q → enc → [q_tokens]; D → enc → [d_tokens]; MaxSim

Comparison

ParadigmWhen Q·D interactDoc pre-encode?Signal granularitySpeedQualityTypical role
No interaction (Bi-Encoder)Never — only final vectors comparedYes (single vector)Coarse (whole-doc)Very fastGoodFirst-stage retrieval
Late interaction (ColBERT / Late Interaction)At query time, after independent encodingYes (per-token vectors)Fine (token-level via MaxSim)MediumNear cross-encoderRetrieval or reranking
Early interaction (Cross-Encoder)During encoding (joint cross-attention)No — must encode each pairFinest (full cross-attn)SlowHighestReranking top-k

The order of the table is the order of the Retrieval Pipeline: cheap-and-coarse interaction filters the corpus, richer interaction reranks the survivors. Because rerankers only see ~100–1000 candidates, they can afford the expensive end of the spectrum — this is the same ladder explored in Reasoning Reranking.

Why Timing Is the Master Trade-off

  • Pre-computation gate. If interaction is deferred until after the document is encoded (no / late), document representations can be built offline and indexed. Once interaction happens during encoding (early), every query–document pair must be encoded at query time — O(candidates) work that rules out first-stage use.
  • Granularity vs. storage. Later interaction generally means keeping more per-document state: a single vector (bi-encoder) → many token vectors (late interaction, hence Token Pooling and compression concerns) → nothing stored but nothing reusable either (cross-encoder).
  • Quality climbs with interaction richness. Full cross-attention (early) captures the subtlest relevance; whole-document dot products (none) capture the least; token-level MaxSim (late) lands in between.

A Second Axis: Sparse vs. Dense

The interaction axis is orthogonal to what the representation is. The same “no / late / early” question can be asked of sparse models: Learned Sparse Retrieval (SPLADE, ELSER) are effectively no-interaction models over learned sparse term weights, scored with an inverted index instead of dot products over dense vectors. Hybrid Search combines a sparse and a dense no-interaction retriever before any later-interaction reranking.