File size: 11,295 Bytes
4e1037f | 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 | """Targeted self-consistency retry on baseline failures with Mistral codestral.
For each failing question, runs the production G pipeline N times at distinct
temperatures (0.2, 0.4, 0.6, 0.8 by default), executes each candidate, and
votes via the largest fingerprint cluster (ties → highest confidence). Output
is voting-shaped for `merge_voting_rescues.py`.
Same model (Mistral codestral) — wins beyond ~1-2 are unlikely because
voting same-model against itself plateaus, but it's a free-tier sanity probe.
Usage:
uv run python scripts/run_selfcon_retry.py \
--baseline eval/reports/2026-05-13/hybrid+multi-vote+critique-v4.json \
--out eval/reports/2026-05-13/selfcon-retry.json
uv run python scripts/run_selfcon_retry.py \
--baseline eval/reports/2026-05-22/v20-kimi-k2-thinking-merged.json \
--out eval/reports/2026-05-22/selfcon-qid1399.json --only-qids 1399
"""
from __future__ import annotations
import argparse
import json
import sys
import time
from pathlib import Path
from nl_sql.agent.graph import PipelineConfig, build_pipeline, run_pipeline
from nl_sql.config import get_settings
from nl_sql.db.registry import get_default_registry
from nl_sql.eval.dataset import load_bird_mini_dev
from nl_sql.eval.metrics.execution_accuracy import compare_results
from nl_sql.eval.runner import _compose_question, _execute_gold
from nl_sql.eval.self_consistency import Candidate, vote
from nl_sql.execution.runner import execute_validated
from nl_sql.llm.cache import CachingEmbeddingProvider, CachingLLMProvider
from nl_sql.llm.providers.base import (
EmbedRequest,
EmbedResponse,
GenerateRequest,
GenerateResponse,
ProviderError,
)
from nl_sql.llm.providers.mistral import MistralProvider
from nl_sql.schema_index.indexer import SchemaIndex
class RotatingMistralProvider:
"""Round-robin wrapper across N MistralProvider instances (different API keys).
On a 429 / rate-limit error, advances to the next key and retries. After a
full rotation without success, applies escalating backoff (5s * extra-attempts)
and keeps trying up to 2*N attempts before surrendering.
"""
name = "mistral"
def __init__(self, providers: list[MistralProvider]) -> None:
if not providers:
raise ProviderError("RotatingMistralProvider requires >=1 provider")
self._providers = providers
self._idx = 0
self.model = providers[0].model
self.embed_model = providers[0].embed_model
def _is_rate_limit(self, err: Exception) -> bool:
msg = str(err)
return "429" in msg or "Rate limit" in msg or "rate_limited" in msg
def _advance(self) -> None:
self._idx = (self._idx + 1) % len(self._providers)
def generate(self, req: GenerateRequest) -> GenerateResponse:
n = len(self._providers)
last_err: Exception | None = None
for attempt in range(n * 2):
prov = self._providers[self._idx]
try:
return prov.generate(req)
except ProviderError as exc:
if not self._is_rate_limit(exc):
raise
last_err = exc
self._advance()
if attempt >= n - 1:
time.sleep(5.0 * (attempt - n + 2))
raise ProviderError(f"all {n} keys rate-limited: {last_err}")
def embed(self, req: EmbedRequest) -> EmbedResponse:
n = len(self._providers)
last_err: Exception | None = None
for _ in range(n):
try:
return self._providers[self._idx].embed(req)
except ProviderError as exc:
if not self._is_rate_limit(exc):
raise
last_err = exc
self._advance()
raise ProviderError(f"all {n} keys rate-limited for embed: {last_err}")
def main() -> int:
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--baseline", type=Path, required=True)
p.add_argument("--bird-root", type=Path, default=Path("data/bird_mini_dev/MINIDEV"))
p.add_argument(
"--only-qids",
default="",
help="comma-separated baseline failure qids to retry exactly, preserving argument order",
)
p.add_argument("--temperatures", nargs="+", type=float, default=[0.2, 0.4, 0.6, 0.8])
p.add_argument("--gen-model", default="codestral-latest", help="Mistral model id")
p.add_argument(
"--sleep-between",
type=float,
default=0.0,
help="seconds between pipeline calls (use for mistral-large rate limits)",
)
p.add_argument(
"--api-keys",
default=None,
help="CSV of Mistral API keys for round-robin rotation. Default: settings.mistral_api_key.",
)
p.add_argument("--out", type=Path, required=True)
args = p.parse_args()
baseline = json.loads(args.baseline.read_text(encoding="utf-8"))
fails = [r for r in baseline["records"] if not r.get("match")]
try:
only_qids = [int(x) for x in args.only_qids.split(",") if x.strip()]
except ValueError:
print("[error] invalid --only-qids: expected comma-separated integers", file=sys.stderr)
return 3
if only_qids:
fails_by_qid = {int(r["question_id"]): r for r in fails}
missing_qids = [qid for qid in only_qids if qid not in fails_by_qid]
if missing_qids:
print(f"[error] qids not found in baseline failures: {missing_qids}", file=sys.stderr)
return 3
fails = [fails_by_qid[qid] for qid in only_qids]
settings = get_settings()
if args.api_keys:
keys = [k.strip() for k in args.api_keys.split(",") if k.strip()]
else:
keys = [settings.mistral_api_key]
if not keys or not keys[0]:
print("[error] no Mistral API keys provided", file=sys.stderr)
return 1
print(
f"[info] {len(fails)} failures, temps={args.temperatures}, model={args.gen_model}, keys={len(keys)}",
file=sys.stderr,
)
examples = {e.question_id: e for e in load_bird_mini_dev(args.bird_root)}
registry = get_default_registry()
gen_providers = [MistralProvider(api_key=k, gen_model=args.gen_model) for k in keys]
mistral = RotatingMistralProvider(gen_providers) if len(keys) > 1 else gen_providers[0]
sql_prov = CachingLLMProvider(mistral, cache_dir=settings.llm_cache_dir)
embed_providers = [MistralProvider(api_key=k) for k in keys]
emb_base = RotatingMistralProvider(embed_providers) if len(keys) > 1 else embed_providers[0]
emb = CachingEmbeddingProvider(emb_base, cache_dir=settings.llm_cache_dir)
idx = SchemaIndex(persist_dir="chroma_data", embedder=emb)
pipelines = [
build_pipeline(
PipelineConfig(
sql_provider=sql_prov,
explain_provider=sql_prov,
schema_index=idx,
registry=registry,
fewshot_top_k=3,
sort_schema_block=True,
cross_db_fewshot=True,
verify_retry_on_empty=False,
sql_temperature=t,
)
)
for t in args.temperatures
]
records = []
rescued = 0
regressed = 0
same = 0
for i, br in enumerate(fails, 1):
qid = br["question_id"]
ex = examples.get(qid)
if ex is None:
continue
spec = registry.get(ex.registry_db_id)
engine = spec.make_engine()
try:
t0 = time.perf_counter()
candidates = []
for pipeline, temp in zip(pipelines, args.temperatures, strict=True):
try:
r = run_pipeline(
pipeline,
question=_compose_question(ex),
db_id=ex.registry_db_id,
dialect="sqlite",
)
candidates.append(Candidate(result=r, temperature=temp))
except Exception as exc:
print(f"[{i:3d}/{len(fails)}] qid={qid} T={temp} EXC: {exc}", file=sys.stderr)
if args.sleep_between > 0:
time.sleep(args.sleep_between)
if not candidates:
continue
winner = vote(candidates)
elapsed = (time.perf_counter() - t0) * 1000.0
alt_sql = winner.result.sql or ""
try:
outcome = execute_validated(
engine,
alt_sql,
dialect="sqlite",
statement_timeout_ms=30_000,
row_cap=10_000,
)
alt_rows = list(outcome.result.rows) if outcome.result else []
except Exception:
alt_rows = []
try:
gold_rows, _ = _execute_gold(
engine, ex.sql, statement_timeout_ms=30_000, row_cap=10_000
)
except Exception:
gold_rows = []
alt_cmp = compare_results(gold_rows, alt_rows, gold_sql=ex.sql)
alt_match = bool(alt_cmp.match)
if alt_match and not br.get("match"):
rescued += 1
tag = "RESCUE"
elif br.get("match") and not alt_match:
regressed += 1
tag = "regression"
else:
same += 1
tag = "same"
records.append(
{
"question_id": qid,
"db_id": ex.db_id,
"difficulty": ex.difficulty,
"question": ex.question,
"gold_sql": ex.sql,
"baseline_pred": br["pred_sql"],
"alt_pred": alt_sql,
"alt_confidence": getattr(winner.result, "confidence", None),
"winner_temperature": winner.temperature,
"baseline_match": bool(br.get("match")),
"alt_match": alt_match,
"vote_match": alt_match,
"vote_source": "self-consistency",
"elapsed_ms": elapsed,
}
)
print(
f"[{i:3d}/{len(fails)}] qid={qid} {ex.difficulty:11s} {tag} T_win={winner.temperature:.1f} ({elapsed:.0f}ms)",
file=sys.stderr,
)
finally:
engine.dispose()
print("\n=== self-consistency retry summary ===", file=sys.stderr)
print(f" cases: {len(records)}", file=sys.stderr)
print(f" rescued: {rescued}", file=sys.stderr)
print(f" regressed: {regressed}", file=sys.stderr)
print(f" same: {same}", file=sys.stderr)
args.out.parent.mkdir(parents=True, exist_ok=True)
args.out.write_text(
json.dumps(
{
"alt_model": f"{args.gen_model}+self-consistency",
"temperatures": list(args.temperatures),
"summary": {"voted_better": rescued, "voted_worse": regressed, "voted_same": same},
"records": records,
},
indent=2,
),
encoding="utf-8",
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|