codesearch / scripts /compare_retrievers.py
CamQuestByte
M2 done-criterion: qualitative BM25 vs dense comparison
5de6364
Raw
History Blame Contribute Delete
10.3 kB
"""
Find queries that illustrate where dense beats BM25, where BM25 beats dense,
and where both nail it.
M2 done-criterion: be able to explain one query per category with a concrete
reason, not just "dense is better at semantics." This script surfaces the
candidates; the writeup is for the human.
Pipeline:
1. Load cached query vectors (no re-embed) — see scripts/cache_query_vectors.py
2. Sample to --max-queries
3. Load full corpus, build BM25 index
4. Run BM25 batch on sample
5. Run dense batch (Qdrant via httpx) on sample using cached vectors
6. For each query: rank of relevant doc in BM25 and in dense (or None if >100)
7. Classify and print top --top-each examples per category with code snippets
Usage:
uv run python scripts/cache_query_vectors.py # one-time, if not done
uv run python scripts/compare_retrievers.py
uv run python scripts/compare_retrievers.py --max-queries 5000 --top-each 5
"""
from __future__ import annotations
import argparse
import os
import pickle
import random
import sys
import textwrap
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "src"))
import httpx
from tqdm import tqdm
from codesearch.config import (
EMBEDDING_MODEL,
QDRANT_API_KEY,
QDRANT_COLLECTION,
QDRANT_URL,
)
from codesearch.data import load_codesearch
from codesearch.retrievers.bm25 import BM25Retriever
from codesearch.retrievers.bm25_index import BM25Index
_SAMPLE_SEED = 42
_TOP_K = 100
_SEARCH_BATCH = 50
CACHE_DIR = ".cache"
BM25_CACHE_DIR = ".cache/bm25"
# Classification thresholds for an "X wins" example
_RANK_GOOD = 5 # winner placed relevant doc at rank <= this
_RANK_BAD = 20 # loser placed it deeper than this (or missed it entirely)
def cache_path(model_name: str) -> str:
safe = model_name.replace("/", "_")
return os.path.join(CACHE_DIR, f"query_vectors_{safe}.pkl")
def load_cached() -> tuple[list[dict], list[list[float]]]:
path = cache_path(EMBEDDING_MODEL)
if not os.path.exists(path):
sys.exit(
f"[error] Cache not found at {path}.\n"
f" Run: uv run python scripts/cache_query_vectors.py"
)
with open(path, "rb") as f:
data = pickle.load(f)
if data["model"] != EMBEDDING_MODEL:
sys.exit(
f"[error] Cache model mismatch. Re-run scripts/cache_query_vectors.py --recompute"
)
return data["queries"], data["vectors"].tolist()
def dense_search_batch(http: httpx.Client, vectors: list[list[float]], ef: int = 128) -> list[list[str]]:
"""Returns one list of ranked doc_ids per query."""
hits_all: list[list[str]] = []
n_batches = (len(vectors) + _SEARCH_BATCH - 1) // _SEARCH_BATCH
for i in tqdm(range(0, len(vectors), _SEARCH_BATCH), total=n_batches, desc="Dense search"):
chunk = vectors[i : i + _SEARCH_BATCH]
payload = {
"searches": [
{
"query": qv,
"limit": _TOP_K,
"params": {"hnsw_ef": ef},
"with_payload": ["doc_id"],
}
for qv in chunk
]
}
r = http.post(
f"/collections/{QDRANT_COLLECTION}/points/query/batch",
json=payload,
timeout=60.0,
)
r.raise_for_status()
for resp in r.json()["result"]:
hits_all.append([p["payload"]["doc_id"] for p in resp["points"]])
return hits_all
def rank_of(hit_ids: list[str], relevant_id: str) -> int | None:
for i, h in enumerate(hit_ids, start=1):
if h == relevant_id:
return i
return None
def fmt_rank(r: int | None) -> str:
return f"#{r}" if r is not None else ">100"
def _snippet(text: str | None, n_lines: int = 4, max_chars: int = 280) -> str:
if not text:
return "(empty)"
body = "\n".join(text.strip().splitlines()[:n_lines])
if len(body) > max_chars:
body = body[:max_chars].rstrip() + " ..."
return body
def print_example(
idx: int,
category: str,
q: dict,
gt_doc: dict | None,
bm25_rank: int | None,
dense_rank: int | None,
bm25_top1_doc: dict | None,
dense_top1_doc: dict | None,
) -> None:
sep = "─" * 78
print(f"\n{sep}")
print(f" {category} — example #{idx}")
print(sep)
print(f"Query: {q['query'].strip()[:240]}")
print()
print(f" Ground-truth doc: {q['relevant_id']}")
print(f" BM25 rank: {fmt_rank(bm25_rank):>5} Dense rank: {fmt_rank(dense_rank):>5}")
if gt_doc:
print(f" GT docstring: {(gt_doc.get('docstring') or '').strip()[:160]}")
print(f" GT code:")
print(textwrap.indent(_snippet(gt_doc.get("code")), " "))
print()
if bm25_top1_doc is not None:
print(f" BM25 top-1: {bm25_top1_doc.get('id', '?')}")
print(textwrap.indent(_snippet(bm25_top1_doc.get("code")), " "))
print()
if dense_top1_doc is not None:
print(f" Dense top-1: {dense_top1_doc.get('id', '?')}")
print(textwrap.indent(_snippet(dense_top1_doc.get("code")), " "))
def main() -> None:
parser = argparse.ArgumentParser(
description="Find dense-vs-BM25 illustrative queries (M2 concept-checkpoint)."
)
parser.add_argument(
"--max-queries",
type=int,
default=2000,
help="Sample size (default: 2000 — plenty to surface clear examples).",
)
parser.add_argument(
"--top-each",
type=int,
default=3,
help="Examples to print per category (default: 3).",
)
args = parser.parse_args()
# [1/5] Cached queries + vectors
print("[1/5] Loading cached query vectors...")
queries, vectors = load_cached()
print(f" Loaded {len(queries):,} queries.")
if args.max_queries and len(queries) > args.max_queries:
random.seed(_SAMPLE_SEED)
idx = random.sample(range(len(queries)), args.max_queries)
queries = [queries[i] for i in idx]
vectors = [vectors[i] for i in idx]
print(f" Sampled {len(queries):,} queries (seed={_SAMPLE_SEED}).")
# [2/4] BM25 — from cache if available, else build from scratch
if BM25Index.exists(BM25_CACHE_DIR):
print(f"[2/4] Loading cached BM25 index from {BM25_CACHE_DIR}...")
bm25 = BM25Retriever.from_cache(BM25_CACHE_DIR)
print(f" Loaded BM25 over {len(bm25.corpus):,} docs.")
else:
print("[2/4] No BM25 cache found — building from scratch (~2-3 min).")
print(f" Tip: run scripts/cache_bm25.py to skip this next time.")
corpus, _ = load_codesearch(n=-1)
bm25 = BM25Retriever(corpus)
corpus_by_id = {doc["id"]: doc for doc in bm25.corpus}
# [3/4] BM25 search
print(f"[3/4] Running BM25 on {len(queries):,} queries...")
bm25_results = bm25.retrieve_batch([q["query"] for q in queries], top_k=_TOP_K)
bm25_hits = [[h["id"] for h in row] for row in bm25_results]
# [4/4] Dense search via Qdrant REST
print(f"[4/4] Running dense (Qdrant) on {len(queries):,} queries...")
http = httpx.Client(
base_url=QDRANT_URL,
headers={"api-key": QDRANT_API_KEY},
)
dense_hits = dense_search_batch(http, vectors)
http.close()
# Compute ranks
annotated = []
for q, b_ids, d_ids in zip(queries, bm25_hits, dense_hits):
rel = q["relevant_id"]
annotated.append(
{
"query": q,
"bm25_rank": rank_of(b_ids, rel),
"dense_rank": rank_of(d_ids, rel),
"bm25_top1": b_ids[0] if b_ids else None,
"dense_top1": d_ids[0] if d_ids else None,
}
)
# Classify
def well(r):
return r is not None and r <= _RANK_GOOD
def badly(r):
return r is None or r > _RANK_BAD
dense_wins = [r for r in annotated if well(r["dense_rank"]) and badly(r["bm25_rank"])]
bm25_wins = [r for r in annotated if well(r["bm25_rank"]) and badly(r["dense_rank"])]
ties = [r for r in annotated if r["bm25_rank"] == 1 and r["dense_rank"] == 1]
# Sort by margin so the most striking cases come first
dense_wins.sort(key=lambda r: -(r["bm25_rank"] or 9999))
bm25_wins.sort(key=lambda r: -(r["dense_rank"] or 9999))
random.seed(_SAMPLE_SEED)
if ties:
random.shuffle(ties)
# Summary
n = len(annotated)
print(
f"\nClassification on {n:,} queries "
f"(winner rank<={_RANK_GOOD}, loser rank>{_RANK_BAD} or missing):"
)
print(f" Dense wins : {len(dense_wins):>5} ({len(dense_wins)/n:.1%})")
print(f" BM25 wins : {len(bm25_wins):>5} ({len(bm25_wins)/n:.1%})")
print(f" Both at #1 : {len(ties):>5} ({len(ties)/n:.1%})")
def doc(doc_id):
return corpus_by_id.get(doc_id)
# Output
print("\n" + "=" * 78)
print(f" DENSE WINS (dense top-{_RANK_GOOD}, BM25 missed)")
print("=" * 78)
for i, r in enumerate(dense_wins[: args.top_each], start=1):
print_example(
i,
"Dense win",
r["query"],
doc(r["query"]["relevant_id"]),
r["bm25_rank"],
r["dense_rank"],
doc(r["bm25_top1"]),
doc(r["dense_top1"]),
)
print("\n\n" + "=" * 78)
print(f" BM25 WINS (BM25 top-{_RANK_GOOD}, dense missed)")
print("=" * 78)
for i, r in enumerate(bm25_wins[: args.top_each], start=1):
print_example(
i,
"BM25 win",
r["query"],
doc(r["query"]["relevant_id"]),
r["bm25_rank"],
r["dense_rank"],
doc(r["bm25_top1"]),
doc(r["dense_top1"]),
)
print("\n\n" + "=" * 78)
print(" TIES (both retrievers placed relevant doc at rank #1)")
print("=" * 78)
for i, r in enumerate(ties[: args.top_each], start=1):
print_example(
i,
"Tie",
r["query"],
doc(r["query"]["relevant_id"]),
r["bm25_rank"],
r["dense_rank"],
doc(r["bm25_top1"]),
doc(r["dense_top1"]),
)
if __name__ == "__main__":
main()