samwjya commited on
Commit
0ccd8ab
·
1 Parent(s): 9430c26
Files changed (1) hide show
  1. reranker.py +83 -0
reranker.py ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from sentence_transformers import CrossEncoder
2
+
3
+
4
+ DEFAULT_MODEL = "BAAI/bge-reranker-v2-m3"
5
+
6
+
7
+ def load_model(model_name: str = DEFAULT_MODEL) -> CrossEncoder:
8
+ """Load a CrossEncoder model from HuggingFace (cached after first download)."""
9
+ print(f"Loading reranker model: {model_name} ...")
10
+ model = CrossEncoder(model_name, max_length=512)
11
+ print("Model loaded.\n")
12
+ return model
13
+
14
+
15
+ def build_pairs(query: str, papers: list[dict]) -> list[tuple[str, str]]:
16
+ """Build (query, document) pairs for CrossEncoder scoring.
17
+
18
+ The document is title + abstract concatenated — both fields carry
19
+ relevance signal and the CrossEncoder attends over the full pair.
20
+ """
21
+ pairs = []
22
+ for paper in papers:
23
+ title = paper.get("title") or ""
24
+ abstract = paper.get("abstract") or ""
25
+ document = f"{title}. {abstract}".strip()
26
+ pairs.append((query, document))
27
+ return pairs
28
+
29
+
30
+ def rerank(query: str, papers: list[dict], model: CrossEncoder, batch_size: int = 16) -> list[dict]:
31
+ """Score every (query, paper) pair with the CrossEncoder and return papers sorted by score.
32
+
33
+ Args:
34
+ query: The original search query string.
35
+ papers: List of paper dicts (already BM25-ranked; order does not matter here).
36
+ model: A loaded CrossEncoder instance.
37
+ batch_size: Number of pairs to score in one forward pass.
38
+
39
+ Returns:
40
+ The same paper dicts, each annotated with `ce_score` and `ce_rank`,
41
+ sorted by CrossEncoder score descending.
42
+ """
43
+ pairs = build_pairs(query, papers)
44
+ scores = model.predict(pairs, batch_size=batch_size, show_progress_bar=True)
45
+
46
+ ranked = sorted(
47
+ zip(papers, scores),
48
+ key=lambda x: x[1],
49
+ reverse=True,
50
+ )
51
+ results = []
52
+ for rank, (paper, score) in enumerate(ranked, start=1):
53
+ paper = dict(paper)
54
+ paper["ce_score"] = round(float(score), 4)
55
+ paper["ce_rank"] = rank
56
+ results.append(paper)
57
+ return results
58
+
59
+
60
+ if __name__ == "__main__":
61
+ import argparse
62
+ from retriever import retrieve
63
+
64
+ parser = argparse.ArgumentParser(description="Rerank BM25 candidates with a CrossEncoder")
65
+ parser.add_argument("query", type=str, help="Search query")
66
+ parser.add_argument("--model", type=str, default=DEFAULT_MODEL, help="HuggingFace CrossEncoder model ID")
67
+ parser.add_argument("--limit", type=int, default=50, help="Number of BM25 candidates to fetch")
68
+ parser.add_argument("--top", type=int, default=10, help="Number of reranked results to display")
69
+ args = parser.parse_args()
70
+
71
+ papers = retrieve(args.query, limit=args.limit)
72
+ model = load_model(args.model)
73
+ reranked = rerank(args.query, papers, model)
74
+
75
+ print(f"\nQuery: {args.query!r}")
76
+ print(f"Showing top {args.top} after CrossEncoder reranking:\n")
77
+ for paper in reranked[:args.top]:
78
+ authors = ", ".join(a["name"] for a in paper.get("authors", [])[:2])
79
+ print(f"[CE #{paper['ce_rank']}] ce_score={paper['ce_score']} (was BM25 #{paper['bm25_rank']})")
80
+ print(f" Title : {paper['title']}")
81
+ print(f" Year : {paper.get('year', 'N/A')} | Citations: {paper.get('citationCount', 'N/A')}")
82
+ print(f" Authors: {authors}")
83
+ print()