File size: 16,880 Bytes
aca22b8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 | from __future__ import annotations
import argparse
import json
import os
import time
from dataclasses import dataclass
from typing import List, Dict, Sequence
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from crom_efficientllm.budget_packer.packer import budget_pack, Chunk
from crom_efficientllm.rerank_engine.rerank import hybrid_rerank
try:
from sentence_transformers import SentenceTransformer
except Exception: # pragma: no cover
SentenceTransformer = None # type: ignore
# Optional plugins are imported lazily when flags are set
@dataclass
class Doc:
id: str
text: str
def load_jsonl(path: str) -> List[Dict]:
with open(path, "r", encoding="utf-8") as f:
return [json.loads(line) for line in f]
def build_corpus(path: str) -> List[Doc]:
rows = load_jsonl(path)
return [Doc(id=str(r.get("id", i)), text=str(r["text"])) for i, r in enumerate(rows)]
def sparse_retrieval(query: str, corpus: Sequence[Doc], k: int = 100) -> List[Dict]:
texts = [d.text for d in corpus]
vect = TfidfVectorizer(ngram_range=(1, 2)).fit(texts)
D = vect.transform(texts)
Q = vect.transform([query])
sims = cosine_similarity(Q, D).ravel()
order = np.argsort(-sims)[:k]
return [{"id": corpus[i].id, "text": corpus[i].text, "score_sparse": float(sims[i])} for i in order]
def dense_embed_model(name: str):
if SentenceTransformer is None:
raise RuntimeError("sentence-transformers not installed. Install with `pip install -e .`.")
return SentenceTransformer(name)
def _apply_flashrank(query: str, docs: List[Dict], model_name: str) -> List[Dict]:
try:
from crom_efficientllm.plugins.flashrank_reranker import flashrank_rerank
except Exception as e: # pragma: no cover
raise RuntimeError("FlashRank plugin not available. Install extras: pip install .[plugins]") from e
ranked = flashrank_rerank(query, docs, model_name=model_name)
# Normalize plugin score to 0..1 and put into score_final
scores = np.array([d.get("score_flashrank", 0.0) for d in ranked], dtype=np.float32)
if scores.size and float(scores.max() - scores.min()) > 1e-12:
s = (scores - scores.min()) / (scores.max() - scores.min())
else:
s = np.zeros_like(scores)
for i, d in enumerate(ranked):
d["score_final"] = float(s[i])
return ranked
def _apply_llmlingua(text: str, ratio: float) -> str:
try:
from crom_efficientllm.plugins.llmlingua_compressor import compress_prompt
except Exception as e: # pragma: no cover
raise RuntimeError("LLMLingua plugin not available. Install extras: pip install .[plugins]") from e
return compress_prompt(text, target_ratio=ratio)
def _save_evidently_report(all_embs: List[List[float]], out_html: str) -> None:
try:
from crom_efficientllm.plugins.evidently_drift import drift_report
except Exception as e: # pragma: no cover
raise RuntimeError("Evidently plugin not available. Install extras: pip install .[plugins]") from e
n = len(all_embs)
if n < 4:
return
ref = all_embs[: n // 2]
cur = all_embs[n // 2 :]
rep = drift_report(ref, cur)
rep.save_html(out_html)
def mock_llm_generate(prompt: str) -> str:
time.sleep(0.005) # simulate small latency
return "[MOCK] " + prompt[:160]
def e2e(args: argparse.Namespace) -> None:
corpus = build_corpus(args.corpus)
queries = [r["query"] for r in load_jsonl(args.queries)]
embed = dense_embed_model(args.model)
all_embs: List[List[float]] = []
t0 = time.perf_counter()
all_rows = []
for q in queries:
t_s = time.perf_counter()
cands = sparse_retrieval(q, corpus, k=args.k)
t_sparse = (time.perf_counter() - t_s) * 1000
t_r = time.perf_counter()
if args.use_flashrank:
reranked = _apply_flashrank(q, cands, args.flashrank_model)
else:
reranked = hybrid_rerank(q, cands, embed, alpha=args.alpha)
t_rerank = (time.perf_counter() - t_r) * 1000
# token heuristic + budget pack
chunks = [
Chunk(text=d["text"], score=d.get("score_final", d.get("score_sparse", 0.0)), tokens=max(1, len(d["text"]) // 4))
for d in reranked
]
budget_tokens = int(sum(c.tokens for c in chunks) * args.budget)
t_p = time.perf_counter()
packed = budget_pack(chunks, budget=budget_tokens)
t_pack = (time.perf_counter() - t_p) * 1000
prompt = "\n\n".join(c.text for c in packed) + f"\n\nQ: {q}\nA:"
if args.use_llmlingua:
prompt = _apply_llmlingua(prompt, ratio=args.compress_ratio)
# collect embeddings for drift snapshot (mean-pooled)
with np.errstate(all="ignore"):
if len(packed) > 0:
doc_embs = embed.encode([c.text for c in packed], convert_to_numpy=True)
vec = np.mean(doc_embs, axis=0).tolist()
all_embs.append(vec)
t_l = time.perf_counter()
_ = mock_llm_generate(prompt)
t_llm = (time.perf_counter() - t_l) * 1000
total = (time.perf_counter() - t_s) * 1000
all_rows.append({
"query": q,
"sparse_ms": t_sparse,
"rerank_ms": t_rerank,
"pack_ms": t_pack,
"llm_ms": t_llm,
"total_ms": total,
"packed_tokens": sum(c.tokens for c in packed),
"orig_tokens": sum(c.tokens for c in chunks),
"save_ratio": 1 - (sum(c.tokens for c in packed) / max(1, sum(c.tokens for c in chunks))),
"used_flashrank": bool(args.use_flashrank),
"used_llmlingua": bool(args.use_llmlingua),
})
elapsed = (time.perf_counter() - t0) * 1000
os.makedirs(args.out_dir, exist_ok=True)
out_path = os.path.join(args.out_dir, "e2e_results.jsonl")
with open(out_path, "w", encoding="utf-8") as f:
for r in all_rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
print(f"saved results -> {out_path} ({len(all_rows)} queries) ; elapsed={elapsed:.2f}ms")
if args.use_evidently and all_embs:
html_path = os.path.join(args.out_dir, "evidently_report.html")
_save_evidently_report(all_embs, html_path)
print(f"evidently report -> {html_path}")
def budget_sweep(args: argparse.Namespace) -> None:
import itertools
corpus = build_corpus(args.corpus)
queries = [r["query"] for r in load_jsonl(args.queries)][: args.max_q]
embed = dense_embed_model(args.model)
budgets = [b / 100.0 for b in range(args.b_min, args.b_max + 1, args.b_step)]
rows = []
for q, b in itertools.product(queries, budgets):
cands = sparse_retrieval(q, corpus, k=args.k)
reranked = hybrid_rerank(q, cands, embed, alpha=args.alpha)
chunks = [Chunk(text=d["text"], score=d["score_final"], tokens=max(1, len(d["text"]) // 4)) for d in reranked]
budget_tokens = int(sum(c.tokens for c in chunks) * b)
packed = budget_pack(chunks, budget=budget_tokens)
rows.append({
"query": q,
"budget": b,
"packed_tokens": sum(c.tokens for c in packed),
"orig_tokens": sum(c.tokens for c in chunks),
"save_ratio": 1 - (sum(c.tokens for c in packed) / max(1, sum(c.tokens for c in chunks))),
"avg_score": float(np.mean([c.score for c in packed])) if packed else 0.0,
})
os.makedirs(args.out_dir, exist_ok=True)
out_path = os.path.join(args.out_dir, "budget_sweep.jsonl")
with open(out_path, "w", encoding="utf-8") as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
print(f"saved results -> {out_path} ; points={len(rows)}")
if args.save_plots:
try:
import matplotlib.pyplot as plt # noqa: F401
import matplotlib.pyplot as _plt
except Exception:
print("[warn] matplotlib not installed; install dev extras: pip install -e .[dev]")
else:
# Aggregate by budget
import collections
agg = collections.defaultdict(list)
for r in rows:
agg[r["budget"]].append(r)
budgets_sorted = sorted(agg.keys())
avg_save = [float(np.mean([x["save_ratio"] for x in agg[b]])) for b in budgets_sorted]
avg_score = [float(np.mean([x["avg_score"] for x in agg[b]])) for b in budgets_sorted]
_plt.figure()
_plt.plot([b * 100 for b in budgets_sorted], [s * 100 for s in avg_save], marker="o")
_plt.xlabel("Budget (%)")
_plt.ylabel("Avg Save Ratio (%)")
_plt.title("Budget Sweep: Save Ratio vs Budget")
_plt.grid(True)
_plt.tight_layout()
_plt.savefig(os.path.join(args.out_dir, "budget_sweep.png"))
_plt.figure()
_plt.plot([s * 100 for s in avg_save], avg_score, marker="o")
_plt.xlabel("Save Ratio (%)")
_plt.ylabel("Avg Score (packed)")
_plt.title("Pareto: Quality vs Savings")
_plt.grid(True)
_plt.tight_layout()
_plt.savefig(os.path.join(args.out_dir, "budget_pareto.png"))
print("plots ->", os.path.join(args.out_dir, "budget_sweep.png"), ",", os.path.join(args.out_dir, "budget_pareto.png"))
def scaling(args: argparse.Namespace) -> None:
def make_synth(n: int, seed: int = 42):
rng = np.random.default_rng(seed)
tokens = np.clip(rng.lognormal(4.0, 0.6, n).astype(int), 5, 2000)
score = rng.normal(0, 1, n)
return [Chunk(text="x" * int(t * 4), score=float(s), tokens=int(t)) for s, t in zip(score, tokens)]
for n in [1000, 5000, 10000, 20000, 50000, 100000]:
if n > args.n_max:
break
chunks = make_synth(n)
budget = int(sum(c.tokens for c in chunks) * args.budget)
t0 = time.perf_counter()
_ = budget_pack(chunks, budget)
ms = (time.perf_counter() - t0) * 1000
print(f"n={n:6d} budget={args.budget:.0%} time={ms:8.2f} ms")
def dp_curve(args: argparse.Namespace) -> None:
def make_synth(n: int, seed: int = 123, corr: float = 0.6):
rng = np.random.default_rng(seed)
true_rel = rng.normal(0, 1, n)
noise = rng.normal(0, 1, n) * np.sqrt(1 - corr**2)
score = corr * true_rel + noise
tokens = np.clip(rng.lognormal(4.0, 0.6, n).astype(int), 5, 2000)
chunks = [Chunk(text="x" * int(t * 4), score=float(s), tokens=int(t)) for s, t in zip(score, tokens)]
return chunks, true_rel
def optimal(chunks: Sequence[Chunk], values: np.ndarray, budget: int) -> float:
B = budget
dp = np.zeros(B + 1, dtype=np.float32)
for i, ch in enumerate(chunks):
wt = ch.tokens
val = max(0.0, float(values[i]))
for b in range(B, wt - 1, -1):
dp[b] = max(dp[b], dp[b - wt] + val)
return float(dp[B])
chunks, true_rel = make_synth(args.n)
total = sum(c.tokens for c in chunks)
budgets = [int(total * b / 100.0) for b in range(args.b_min, args.b_max + 1, args.b_step)]
out_rows = []
for B in budgets:
sel = budget_pack(chunks, B)
idx_map = {id(c): i for i, c in enumerate(chunks)}
rel_bp = float(np.sum([max(0.0, true_rel[idx_map[id(c)]]) for c in sel]))
rel_opt = optimal(chunks[: args.n_opt], true_rel[: args.n_opt], min(B, sum(c.tokens for c in chunks[: args.n_opt])))
pct = rel_bp / max(rel_opt, 1e-9)
out_rows.append({"budget": B, "pct": pct, "rel_bp": rel_bp, "rel_opt": rel_opt})
print(f"budget={B:8d} rel_bp={rel_bp:8.3f} rel_opt≈{rel_opt:8.3f} pct≈{pct*100:5.1f}% (subset n={args.n_opt})")
if args.save_plots:
try:
import matplotlib.pyplot as plt # noqa: F401
import matplotlib.pyplot as _plt
except Exception:
print("[warn] matplotlib not installed; install dev extras: pip install -e .[dev]")
else:
_plt.figure()
xs = [r["budget"] * 100.0 / total for r in out_rows]
ys = [r["pct"] * 100 for r in out_rows]
_plt.plot(xs, ys, marker="o")
_plt.xlabel("Budget (%)")
_plt.ylabel("% of optimal (subset)")
_plt.title("DP Curve: Greedy vs Optimal")
_plt.grid(True)
_plt.tight_layout()
os.makedirs(args.out_dir, exist_ok=True)
_plt.savefig(os.path.join(args.out_dir, "dp_curve.png"))
print("plot ->", os.path.join(args.out_dir, "dp_curve.png"))
def compare_haystack(args: argparse.Namespace) -> None:
try:
from haystack.nodes import BM25Retriever, SentenceTransformersRetriever
from haystack.document_stores import InMemoryDocumentStore
except Exception as e: # pragma: no cover
raise RuntimeError("Install extras: pip install .[haystack]") from e
corpus = build_corpus(args.corpus)
docs = [{"content": d.text, "meta": {"id": d.id}} for d in corpus]
store = InMemoryDocumentStore(use_bm25=True)
store.write_documents(docs)
bm25 = BM25Retriever(document_store=store)
dretr = SentenceTransformersRetriever(document_store=store, model_name_or_path=args.model)
queries = [r["query"] for r in load_jsonl(args.queries)][: args.max_q]
for q in queries:
t0 = time.perf_counter()
bm = bm25.retrieve(q, top_k=args.k)
dn = dretr.retrieve(q, top_k=args.k)
ms = (time.perf_counter() - t0) * 1000
print(f"{q[:40]:40s} bm25={len(bm):3d} dense={len(dn):3d} time={ms:7.2f} ms")
def main() -> None:
ap = argparse.ArgumentParser(prog="crom-bench")
sub = ap.add_subparsers(dest="cmd", required=True)
p = sub.add_parser("e2e", help="end-to-end: retrieval → rerank → pack → mock LLM")
p.add_argument("--corpus", default="examples/corpus/sample_docs.jsonl")
p.add_argument("--queries", default="examples/corpus/sample_queries.jsonl")
p.add_argument("--model", default="sentence-transformers/all-MiniLM-L6-v2")
p.add_argument("--k", type=int, default=200)
p.add_argument("--alpha", type=float, default=0.5)
p.add_argument("--budget", type=float, default=0.3)
# plugins
p.add_argument("--use-flashrank", action="store_true")
p.add_argument("--flashrank-model", default="ms-marco-TinyBERT-L-2-v2")
p.add_argument("--use-llmlingua", action="store_true")
p.add_argument("--compress-ratio", type=float, default=0.6)
p.add_argument("--use-evidently", action="store_true")
p.add_argument("--out-dir", default="benchmarks/out")
p.set_defaults(func=e2e)
p2 = sub.add_parser("sweep", help="budget sweep + Pareto csv")
p2.add_argument("--corpus", default="examples/corpus/sample_docs.jsonl")
p2.add_argument("--queries", default="examples/corpus/sample_queries.jsonl")
p2.add_argument("--model", default="sentence-transformers/all-MiniLM-L6-v2")
p2.add_argument("--k", type=int, default=200)
p2.add_argument("--alpha", type=float, default=0.5)
p2.add_argument("--b-min", type=int, default=10)
p2.add_argument("--b-max", type=int, default=90)
p2.add_argument("--b-step", type=int, default=10)
p2.add_argument("--max-q", type=int, default=20)
p2.add_argument("--out-dir", default="benchmarks/out")
p2.add_argument("--save-plots", action="store_true")
p2.set_defaults(func=budget_sweep)
p3 = sub.add_parser("scale", help="scaling runtime with synthetic data")
p3.add_argument("--n-max", type=int, default=100000)
p3.add_argument("--budget", type=float, default=0.3)
p3.set_defaults(func=scaling)
p4 = sub.add_parser("dp-curve", help="% of optimal vs budget (synthetic)")
p4.add_argument("--n", type=int, default=2000)
p4.add_argument("--n-opt", type=int, default=200)
p4.add_argument("--b-min", type=int, default=10)
p4.add_argument("--b-max", type=int, default=90)
p4.add_argument("--b-step", type=int, default=10)
p4.add_argument("--out-dir", default="benchmarks/out")
p4.add_argument("--save-plots", action="store_true")
p4.set_defaults(func=dp_curve)
p5 = sub.add_parser("haystack-compare", help="compare BM25 vs dense retrievers (Haystack)")
p5.add_argument("--corpus", default="examples/corpus/sample_docs.jsonl")
p5.add_argument("--queries", default="examples/corpus/sample_queries.jsonl")
p5.add_argument("--model", default="sentence-transformers/all-MiniLM-L6-v2")
p5.add_argument("--k", type=int, default=50)
p5.add_argument("--max-q", type=int, default=10)
p5.set_defaults(func=compare_haystack)
args = ap.parse_args()
args.func(args)
if __name__ == "__main__":
main()
|