# Hybrid First-Stage Retrieval + Late Interaction Reranking

## Current Claim

The evidence now supports a sharper paper claim than the initial BM25-vs-late-interaction framing:

> Late-interaction reranking is strongest when its candidate pool combines lexical and semantic first-stage retrieval. BM25 alone is still useful, but the reranker is bounded by first-stage recall; RRF over BM25 and BGE-M3 improves that candidate pool and consistently improves reranked top-10 quality.

## Setup

- Datasets with complete BM25, BGE-M3 dense, BM25+BGE-M3 RRF, BM25-rerank, and RRF-rerank results: 7.
- Additional BM25 and BM25-rerank rows: CQADupStack English and Gaming.
- Lexical retriever: BM25.
- Dense retriever: `BAAI/bge-m3`.
- First-stage hybrid: reciprocal-rank fusion over BM25 and BGE-M3, top 100 candidates.
- Reranker: `lightonai/GTE-ModernColBERT-v1`, exact MaxSim reranking over the first-stage candidates.
- Chunking: 180-word chunks, 120-word stride, max chunk score per source document.
- Device: Apple MPS where available, with CPU fallback for non-finite embeddings.

## Main Table

See `paper_tables.md` for the full table. The complete-seven-dataset average is:

| System | nDCG@10 | Recall@100 |
| --- | ---: | ---: |
| BM25 first stage | 0.2956 | 0.5262 |
| BGE-M3 dense first stage | 0.3831 | 0.6738 |
| BM25 + BGE-M3 RRF first stage | 0.3606 | 0.6743 |
| ModernColBERT rerank of BM25 top-100 | 0.3826 | 0.5262 |
| ModernColBERT rerank of RRF top-100 | 0.4211 | 0.6743 |

## Key Results

1. RRF first-stage retrieval beats BM25 first-stage retrieval on all 7 complete datasets.
2. BGE-M3 dense first-stage retrieval beats BM25 on 6 of 7 complete datasets; SciFact is the one exception.
3. RRF+ModernColBERT beats BM25-top-100+ModernColBERT on all 7 complete datasets.
4. Average nDCG@10 improves from 0.3826 to 0.4211 when ModernColBERT reranks RRF candidates instead of BM25 candidates, a +0.0385 absolute gain.
5. Average recall@100 rises from 0.5262 for BM25 to 0.6743 for RRF, explaining the reranking gain: the reranker cannot recover relevant documents absent from its candidate pool.

## Dataset-Level Pattern

| Dataset | Relevant-doc lexical Jaccard | BM25 - Dense nDCG@10 | RRF-rerank gain over BM25-rerank |
| --- | ---: | ---: | ---: |
| SciFact | 0.0574 | +0.0072 | +0.0310 |
| NFCorpus | 0.0074 | -0.0125 | +0.0245 |
| ArguAna | 0.1704 | -0.1727 | +0.0526 |
| SciDocs | 0.0367 | -0.0233 | +0.0088 |
| FiQA | 0.0566 | -0.1974 | +0.0820 |
| CQADupStack Android | 0.1146 | -0.1243 | +0.0440 |
| CQADupStack Mathematica | 0.0569 | -0.0892 | +0.0269 |

The simple lexical-overlap feature is not enough by itself to predict BM25 superiority: ArguAna and Android have relatively high overlap but dense retrieval still wins strongly. This gives the paper a useful negative result: "lexical helps" is not equivalent to "high token overlap"; the more robust operational rule is to use lexical retrieval as one side of a hybrid candidate generator.

## Interpretation

The result matches the original paper hypothesis, but with a more nuanced answer. Lexical retrieval helps most as candidate-pool insurance, not as a standalone top-10 ranker. Dense retrieval supplies semantic coverage, BM25 supplies exact-match coverage, and late interaction is best used as a precision layer over the fused candidate set.

This turns the paper into a clean systems/IR study:

**When does lexical help semantic retrieval?** It helps when used as a complementary first-stage channel before late-interaction reranking, because hybrid candidate pools raise recall without giving up lexical anchors.

## Caveats

- Dense-only reranking is complete only for SciFact, so it should be treated as an exploratory row, not a main comparison.
- English and Gaming currently have BM25-rerank results but not complete dense/RRF/rerank rows. Gaming exact dense retrieval is large enough on this MacBook Air that it is not worth blocking the first paper.
- MPS produced non-finite embeddings for some long/chunked candidates. The experiment repairs those rows via CPU fallback before scoring, so quality metrics are usable; runtime numbers should be reported only as local diagnostics.