SurfSense/surfsense_evals/scripts/analyze_failure_timing.py
DESKTOP-RTLN3BA\$punk 9bcd50164d feat(evals): publish multimodal_doc parser_compare benchmark + n=171 report
Adds the full parser_compare experiment for the multimodal_doc suite:
six arms compared on 30 PDFs / 171 questions from MMLongBench-Doc with
anthropic/claude-sonnet-4.5 across the board.

Source code:
- core/parsers/{azure_di,llamacloud,pdf_pages}.py: direct parser SDK
  callers (Azure Document Intelligence prebuilt-read/layout, LlamaParse
  parse_page_with_llm/parse_page_with_agent) used by the LC arms,
  bypassing the SurfSense backend so each (basic/premium) extraction
  is a clean A/B independent of backend ETL routing.
- suites/multimodal_doc/parser_compare/{ingest,runner,prompt}.py:
  six-arm benchmark (native_pdf, azure_basic_lc, azure_premium_lc,
  llamacloud_basic_lc, llamacloud_premium_lc, surfsense_agentic) with
  byte-identical prompts per question, deterministic grader, Wilson
  CIs, and the per-page preprocessing tariff cost overlay.

Reproducibility:
- pyproject.toml + uv.lock pin pypdf, azure-ai-documentintelligence,
  llama-cloud-services as new deps.
- .env.example documents the AZURE_DI_* and LLAMA_CLOUD_API_KEY env
  vars now required for parser_compare.
- 12 analysis scripts under scripts/: retry pass with exponential
  backoff, post-retry accuracy merge, McNemar / latency / per-PDF
  stats, context-overflow hypothesis test, etc. Each produces one
  number cited by the blog report.

Citation surface:
- reports/blog/multimodal_doc_parser_compare_n171_report.md: 1219-line
  technical writeup (16 sections) covering headline accuracy, per-format
  accuracy, McNemar pairwise significance, latency / token / per-PDF
  distributions, error analysis, retry experiment, post-retry final
  accuracy, cost amortization model with closed-form derivation, threats
  to validity, and reproducibility appendix.
- data/multimodal_doc/runs/2026-05-14T00-53-19Z/parser_compare/{raw,
  raw_retries,raw_post_retry}.jsonl + run_artifact.json + retry summary
  whitelisted via data/.gitignore as the verifiable numbers source.

Gitignore:
- ignore logs_*.txt + retry_run.log; structured artifacts cover the
  citation surface, debug logs are noise.
- data/.gitignore default-ignores everything, whitelists the n=171 run
  artifacts only (parser manifest left ignored to avoid leaking local
  Windows usernames in absolute paths; manifest is fully regenerable
  via 'ingest multimodal_doc parser_compare').
- reports/.gitignore now whitelists hand-curated reports/blog/.

Also retires the abandoned CRAG Task 3 implementation (download script,
streaming Task 3 ingest, CragTask3Benchmark + tests) and trims the
runner / ingest module APIs to match.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-05-14 19:54:41 -07:00

125 lines
4.8 KiB
Python

"""Were the SSL failures clustered in time (network blip) or evenly
distributed (sustained limit)? Group failures by 1-min buckets using
the run start time and the per-row latency_ms / answer order.
Also: for the one *real* intrinsic failure — the 30MB Anthropic limit
on 2405.09818v1.pdf::Q007 — print the full error message + raw payload
sizes so the blog has a clean root cause.
"""
from __future__ import annotations
import json
from collections import Counter, defaultdict
from pathlib import Path
REPO = Path(__file__).resolve().parents[1]
RUN = REPO / "data" / "multimodal_doc" / "runs" / "2026-05-14T00-53-19Z" / "parser_compare"
RAW = RUN / "raw.jsonl"
PDFS = REPO / "data" / "multimodal_doc" / "mmlongbench" / "pdfs"
def main() -> None:
rows = [
json.loads(line) for line in RAW.read_text(encoding="utf-8").splitlines()
if line.strip()
]
# 1) SSL clustering: failures by question index per arm
by_arm_idx: dict[str, list[tuple[int, str]]] = defaultdict(list)
qid_order: dict[str, int] = {}
arm_seen_count: dict[str, int] = defaultdict(int)
for row in rows:
arm = row["arm"]
idx = arm_seen_count[arm]
arm_seen_count[arm] += 1
qid_order[f"{arm}::{row['qid']}"] = idx
err = row.get("error") or ""
cluster = "ssl" if "SSLError" in err else (
"empty" if not (row.get("raw_text") or "").strip() and not err else (
"5xx" if "502" in err or "503" in err else (
"size_limit" if "exceeds" in err.lower() and "limit" in err.lower() else (
"other_err" if err else "ok"
)
)
)
)
if cluster != "ok":
by_arm_idx[arm].append((idx, cluster))
print("=" * 80)
print("SSL/network-error indices per arm (each arm processes 171 questions in")
print("order; index = sequential position within that arm). Tight clustering")
print("in time = transient blip, even spread = sustained limit.")
print("=" * 80)
for arm in sorted(by_arm_idx):
items = by_arm_idx[arm]
if not items:
continue
idxs = sorted(set(i for i, _ in items))
print(f"\n{arm}: {len(items)} failures at indices {idxs}")
# show clusters
cluster_runs = []
cur = [idxs[0]]
for i in idxs[1:]:
if i - cur[-1] <= 5: # within 5 questions = same time window
cur.append(i)
else:
cluster_runs.append(cur)
cur = [i]
cluster_runs.append(cur)
print(f" clusters (gap<=5): {len(cluster_runs)}: {cluster_runs}")
# 2) The 30MB intrinsic failure — full details
print()
print("=" * 80)
print("Intrinsic failure: 30MB Anthropic input limit on 2405.09818v1.pdf::Q007")
print("=" * 80)
for row in rows:
if row["qid"] == "2405.09818v1.pdf::Q007" and row["arm"] == "native_pdf":
err = row.get("error") or ""
print(f" qid: {row['qid']}")
print(f" doc: {row['doc_id']}, pages: {row.get('pages')}")
pdf_path = PDFS / row["doc_id"]
if pdf_path.exists():
size_mb = pdf_path.stat().st_size / (1024 * 1024)
print(f" PDF size on disk: {size_mb:.1f} MB")
# base64 inflates ~33%
est_b64 = size_mb * 1.33
print(f" estimated base64 wire size: {est_b64:.1f} MB")
print(f" full error: {err[:600]}")
break
# 3) Per-PDF: which PDFs are pathological?
print()
print("=" * 80)
print("Per-PDF failure breakdown across all 6 arms (only PDFs with failures)")
print("=" * 80)
by_pdf: dict[str, list[dict]] = defaultdict(list)
for row in rows:
err = row.get("error") or ""
empty = not (row.get("raw_text") or "").strip()
if err or empty:
by_pdf[row["doc_id"]].append({
"arm": row["arm"],
"qid": row["qid"],
"err_kind": (
"ssl" if "SSLError" in err
else "size_limit" if "exceeds" in err.lower() and "limit" in err.lower()
else "5xx" if "502" in err or "503" in err
else "json_decode" if "JSONDecodeError" in err
else "empty" if empty and not err
else "other"
),
"pages": row.get("pages"),
})
for doc, items in sorted(by_pdf.items(), key=lambda x: (-len(x[1]), x[0])):
kinds = Counter(i["err_kind"] for i in items)
arms = sorted({i["arm"] for i in items})
pages = items[0]["pages"]
print(f" {doc} pages={pages} failures={len(items)} arms={arms}")
print(f" kinds: {dict(kinds)}")
if __name__ == "__main__":
main()