Datasets:
Tasks:
Document Question Answering
Size:
n<1K
Tags:
benchmark
document-ai
information-extraction
structured-extraction
key-information-extraction
ocr
License:
File size: 18,809 Bytes
8143f06 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 | """run a single document through OpenAI GPT with the source file and JSON schema.
Uploads the document, sends it to the Responses API with Structured Outputs, and writes
the same result envelope used by the benchmark:
python scripts/run_gpt.py <document_path> <schemas/doc_id.json> <output.json>
Set OPENAI_API_KEY. Override OPENAI_MODEL to change the model; the default is gpt-5.5.
LLM/API/schema failures are intentionally written as result files with status="failed"
and data={}, so score_all.py counts every labeled field as an error instead of treating
the run as missing.
"""
from __future__ import annotations
import base64
import json
import mimetypes
import os
import re
import sys
import time
from io import BytesIO
from pathlib import Path
from typing import Any
import requests
OPENAI_API_BASE = os.environ.get("OPENAI_API_BASE", "https://api.openai.com/v1")
DEFAULT_OPENAI_MODEL = "gpt-5.5"
DEFAULT_TIMEOUT_SEC = 600
SCHEMA_MODE = "openai_strict_nullable_v1"
DROP_SCHEMA_KEYS = {"$schema", "examples", "default", "title"}
PRIMITIVE_TYPES = {"string", "number", "integer", "boolean"}
IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".webp", ".gif"}
TIFF_EXTENSIONS = {".tif", ".tiff"}
class ExtractionFailure(Exception):
"""A per-document extraction failure that should be tracked as data={}."""
def __init__(self, kind: str, message: str, meta: dict[str, Any] | None = None):
super().__init__(message)
self.kind = kind
self.message = message
self.meta = meta or {}
def api_key() -> str:
key = os.environ.get("OPENAI_API_KEY")
if not key:
raise RuntimeError("OPENAI_API_KEY not set")
return key
def openai_model() -> str:
return os.environ.get("OPENAI_MODEL", DEFAULT_OPENAI_MODEL)
def load_env_file(path: Path) -> None:
if not path.exists():
return
with open(path, encoding="utf-8") as f:
for line in f:
stripped = line.strip()
if not stripped or stripped.startswith("#") or "=" not in stripped:
continue
if stripped.startswith("export "):
stripped = stripped[len("export "):].strip()
name, value = stripped.split("=", 1)
name = name.strip()
value = value.strip().strip("'\"")
if name and name not in os.environ:
os.environ[name] = value
def headers() -> dict[str, str]:
return {"Authorization": f"Bearer {api_key()}"}
def load_json(path: Path) -> dict[str, Any]:
with open(path, encoding="utf-8") as f:
return json.load(f)
def write_json(path: Path, payload: dict[str, Any]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with open(path, "w", encoding="utf-8") as f:
json.dump(payload, f, ensure_ascii=False, indent=2)
def nullable_type(type_value: str | list[str], nullable: bool) -> str | list[str]:
types = [type_value] if isinstance(type_value, str) else list(type_value)
if nullable and "null" not in types:
types.append("null")
return types[0] if len(types) == 1 else types
def sanitize_schema_literal(value: Any) -> Any:
"""Remove characters OpenAI strict schema rejects from metadata/enum literals."""
if isinstance(value, str):
return value.replace('"', "")
if isinstance(value, list):
return [sanitize_schema_literal(item) for item in value]
if isinstance(value, dict):
return {key: sanitize_schema_literal(item) for key, item in value.items()}
return value
def normalize_schema_node(spec: Any, *, nullable: bool = True) -> dict[str, Any]:
"""Convert the benchmark schema to the strict JSON schema subset OpenAI accepts.
Structured Outputs requires every object property to be listed in "required" and
rejects unknown object fields. Since benchmark labels use null for many fields even
when the source schema says "string"/"number", this normalizer preserves field shape
while making object properties nullable.
"""
if not isinstance(spec, dict):
return {"type": nullable_type("string", nullable)}
out: dict[str, Any] = {}
for key, value in spec.items():
if key in DROP_SCHEMA_KEYS or key.startswith("x_"):
continue
if key == "format":
continue
if key in {"properties", "items", "required", "additionalProperties"}:
continue
out[key] = sanitize_schema_literal(value)
type_value = spec.get("type", "object" if "properties" in spec else "string")
type_list = [type_value] if isinstance(type_value, str) else list(type_value or [])
if "object" in type_list:
properties = spec.get("properties") or {}
normalized_props = {
name: normalize_schema_node(child, nullable=True)
for name, child in properties.items()
}
out["type"] = nullable_type("object", nullable)
out["properties"] = normalized_props
out["required"] = list(normalized_props.keys())
out["additionalProperties"] = False
return out
if "array" in type_list:
out["type"] = nullable_type("array", nullable)
out["items"] = normalize_schema_node(spec.get("items", {"type": "string"}), nullable=True)
return out
primitive = next((t for t in type_list if t in PRIMITIVE_TYPES), "string")
out["type"] = nullable_type(primitive, nullable)
if "enum" in out and nullable and None not in out["enum"]:
out["enum"] = list(out["enum"]) + [None]
return out
def normalize_output_schema(schema: dict[str, Any]) -> dict[str, Any]:
return normalize_schema_node(schema, nullable=False)
def schema_name(doc_id: str) -> str:
name = re.sub(r"[^a-zA-Z0-9_-]+", "_", f"docubench_{doc_id}")
return name[:64] or "docubench_schema"
def check_response(resp: requests.Response, action: str) -> dict[str, Any]:
if resp.status_code < 300:
return resp.json()
try:
body = resp.json()
except ValueError:
body = {"error": {"message": resp.text[:800]}}
err = body.get("error") if isinstance(body, dict) else None
message = err.get("message") if isinstance(err, dict) else str(body)[:800]
raise ExtractionFailure("api_error", f"{action} failed {resp.status_code}: {message}", {"http_status": resp.status_code})
def content_type(path: Path) -> str:
guessed, _ = mimetypes.guess_type(path.name)
return guessed or "application/octet-stream"
def upload_file(path: Path) -> str:
with open(path, "rb") as f:
resp = requests.post(
f"{OPENAI_API_BASE}/files",
headers=headers(),
data={"purpose": "user_data"},
files={"file": (path.name, f, content_type(path))},
timeout=180,
)
return check_response(resp, "file upload")["id"]
def image_part_from_bytes(data: bytes, media_type: str = "image/png") -> dict[str, Any]:
encoded = base64.b64encode(data).decode("ascii")
return {"type": "input_image", "image_url": f"data:{media_type};base64,{encoded}"}
def tiff_png_parts(path: Path) -> tuple[list[dict[str, Any]], dict[str, Any]]:
try:
from PIL import Image, ImageSequence
except ImportError as exc:
raise ExtractionFailure("dependency_missing", "Pillow is required to convert TIFF inputs to PNG") from exc
parts: list[dict[str, Any]] = []
with Image.open(path) as image:
for index, frame in enumerate(ImageSequence.Iterator(image), start=1):
png = BytesIO()
page = frame.copy()
if page.mode not in {"RGB", "RGBA", "L"}:
page = page.convert("RGB")
page.save(png, format="PNG")
parts.append({"type": "input_text", "text": f"TIFF page {index}, converted to PNG:"})
parts.append(image_part_from_bytes(png.getvalue()))
if not parts:
raise ExtractionFailure("invalid_tiff", "TIFF contained no frames")
return parts, {"input_mode": "tiff_png_sequence", "tiff_pages": len(parts) // 2}
def input_parts_for_file(path: Path) -> tuple[list[dict[str, Any]], str | None, dict[str, Any]]:
if path.suffix.lower() in TIFF_EXTENSIONS:
parts, meta = tiff_png_parts(path)
return parts, None, meta
if path.suffix.lower() in IMAGE_EXTENSIONS:
with open(path, "rb") as f:
return [image_part_from_bytes(f.read(), content_type(path))], None, {"input_mode": "image"}
file_id = upload_file(path)
return [{"type": "input_file", "file_id": file_id}], file_id, {"input_mode": "file"}
PROMPT_TEMPLATE_PATH = Path(__file__).resolve().parent.parent / "prompts" / "extraction_prompt.txt"
DEFAULT_PROMPT_TEMPLATE = (
"Extract the document into the supplied JSON schema. "
"Use only information present in the document. "
"Return null for fields that are not printed or cannot be determined. "
"Preserve table rows as arrays and preserve the document language for values. "
"Document id: {doc_id}."
)
def load_prompt_template() -> str:
"""Load the canonical extraction prompt committed in prompts/extraction_prompt.txt.
The committed file is the single source of truth for the prompt that produced the
baseline result sets; fall back to the inline default only if it is missing.
"""
try:
return PROMPT_TEMPLATE_PATH.read_text(encoding="utf-8").strip()
except OSError:
return DEFAULT_PROMPT_TEMPLATE
def build_prompt(doc_id: str) -> str:
return load_prompt_template().format(doc_id=doc_id)
def response_payload(doc_id: str, document_parts: list[dict[str, Any]], schema: dict[str, Any]) -> dict[str, Any]:
payload: dict[str, Any] = {
"model": openai_model(),
"store": False,
"input": [
{
"role": "user",
"content": [
*document_parts,
{"type": "input_text", "text": build_prompt(doc_id)},
],
}
],
"text": {
"format": {
"type": "json_schema",
"name": schema_name(doc_id),
"strict": True,
"schema": schema,
}
},
}
max_output_tokens = os.environ.get("OPENAI_MAX_OUTPUT_TOKENS")
if max_output_tokens:
payload["max_output_tokens"] = int(max_output_tokens)
reasoning_effort = os.environ.get("OPENAI_REASONING_EFFORT")
if reasoning_effort:
payload["reasoning"] = {"effort": reasoning_effort}
return payload
def create_response(payload: dict[str, Any]) -> dict[str, Any]:
resp = requests.post(
f"{OPENAI_API_BASE}/responses",
headers={**headers(), "Content-Type": "application/json"},
json=payload,
timeout=DEFAULT_TIMEOUT_SEC,
)
return check_response(resp, "response creation")
def extract_text(response: dict[str, Any]) -> str:
if response.get("output_text"):
return response["output_text"]
chunks: list[str] = []
for item in response.get("output") or []:
for content in item.get("content") or []:
ctype = content.get("type")
if ctype in {"output_text", "text"} and content.get("text") is not None:
chunks.append(content["text"])
if ctype == "refusal":
raise ExtractionFailure("refusal", content.get("refusal") or "model refused the request")
return "".join(chunks).strip()
def first_type(schema: dict[str, Any]) -> str:
type_value = schema.get("type")
if isinstance(type_value, list):
return next((t for t in type_value if t != "null"), "null")
return type_value or "string"
def allows_null(schema: dict[str, Any]) -> bool:
type_value = schema.get("type")
return type_value == "null" or (isinstance(type_value, list) and "null" in type_value)
def validate_value(value: Any, schema: dict[str, Any], path: str = "$") -> list[str]:
if value is None and allows_null(schema):
return []
expected = first_type(schema)
errors: list[str] = []
if "enum" in schema and value not in schema["enum"]:
errors.append(f"{path}: {value!r} is not in enum")
if expected == "object":
if not isinstance(value, dict):
return [f"{path}: expected object"]
properties = schema.get("properties") or {}
required = schema.get("required") or []
for name in required:
if name not in value:
errors.append(f"{path}.{name}: missing required field")
if schema.get("additionalProperties") is False:
for name in value:
if name not in properties:
errors.append(f"{path}.{name}: unexpected field")
for name, child_schema in properties.items():
if name in value:
errors.extend(validate_value(value[name], child_schema, f"{path}.{name}"))
return errors
if expected == "array":
if not isinstance(value, list):
return [f"{path}: expected array"]
item_schema = schema.get("items") or {}
for index, item in enumerate(value):
errors.extend(validate_value(item, item_schema, f"{path}[{index}]"))
return errors
if expected == "string" and not isinstance(value, str):
errors.append(f"{path}: expected string")
elif expected == "number" and (not isinstance(value, (int, float)) or isinstance(value, bool)):
errors.append(f"{path}: expected number")
elif expected == "integer" and (not isinstance(value, int) or isinstance(value, bool)):
errors.append(f"{path}: expected integer")
elif expected == "boolean" and not isinstance(value, bool):
errors.append(f"{path}: expected boolean")
return errors
def parse_response_data(response: dict[str, Any], output_schema: dict[str, Any]) -> dict[str, Any]:
status = response.get("status")
if status and status != "completed":
raise ExtractionFailure(
"incomplete_response",
f"response status was {status}",
{"incomplete_details": response.get("incomplete_details")},
)
text = extract_text(response)
if not text:
raise ExtractionFailure("empty_response", "response contained no output text")
try:
data = json.loads(text)
except json.JSONDecodeError as exc:
raise ExtractionFailure("invalid_json", f"response was not valid JSON: {exc}") from exc
errors = validate_value(data, output_schema)
if errors:
raise ExtractionFailure("schema_mismatch", "; ".join(errors[:20]), {"validation_errors": errors[:200]})
return data
def estimate_cost(usage: dict[str, Any]) -> float | None:
"""Optional cost estimate using env-provided per-million token prices."""
try:
input_rate = float(os.environ["OPENAI_INPUT_USD_PER_1M"])
output_rate = float(os.environ["OPENAI_OUTPUT_USD_PER_1M"])
except (KeyError, ValueError):
return None
input_tokens = usage.get("input_tokens", 0) or 0
output_tokens = usage.get("output_tokens", 0) or 0
return (input_tokens * input_rate + output_tokens * output_rate) / 1_000_000
def failure_result(
kind: str,
message: str,
*,
started_at: float,
doc_id: str,
file_id: str | None = None,
response: dict[str, Any] | None = None,
extra_meta: dict[str, Any] | None = None,
) -> dict[str, Any]:
usage = (response or {}).get("usage") or {}
meta = {
"provider": "openai",
"model": openai_model(),
"doc_id": doc_id,
"file_id": file_id,
"response_id": (response or {}).get("id"),
"usage": usage,
"schema_mode": SCHEMA_MODE,
}
if extra_meta:
meta.update(extra_meta)
return {
"status": "failed",
"error": {"type": kind, "message": message},
"cost": estimate_cost(usage),
"time_sec": time.time() - started_at,
"data": {},
"meta": meta,
}
def run(doc_id: str, file_path: Path, json_schema: dict[str, Any]) -> dict[str, Any]:
started_at = time.time()
file_id: str | None = None
input_meta: dict[str, Any] = {}
response: dict[str, Any] | None = None
output_schema = normalize_output_schema(json_schema)
try:
document_parts, file_id, input_meta = input_parts_for_file(file_path)
response = create_response(response_payload(doc_id, document_parts, output_schema))
data = parse_response_data(response, output_schema)
except ExtractionFailure as exc:
return failure_result(
exc.kind,
exc.message,
started_at=started_at,
doc_id=doc_id,
file_id=file_id,
response=response,
extra_meta={**input_meta, **exc.meta},
)
except requests.RequestException as exc:
return failure_result(
"request_error",
str(exc),
started_at=started_at,
doc_id=doc_id,
file_id=file_id,
response=response,
extra_meta=input_meta,
)
except Exception as exc:
return failure_result(
exc.__class__.__name__,
str(exc),
started_at=started_at,
doc_id=doc_id,
file_id=file_id,
response=response,
extra_meta=input_meta,
)
usage = response.get("usage") or {}
return {
"status": "ok",
"cost": estimate_cost(usage),
"time_sec": time.time() - started_at,
"data": data,
"meta": {
"provider": "openai",
"model": openai_model(),
"doc_id": doc_id,
"file_id": file_id,
"response_id": response.get("id"),
"usage": usage,
"schema_mode": SCHEMA_MODE,
**input_meta,
},
}
def main() -> int:
if len(sys.argv) != 4:
print("usage: python3 scripts/run_gpt.py <document_path> <schemas/doc_id.json> <output.json>")
return 1
file_path = Path(sys.argv[1])
schema_path = Path(sys.argv[2])
output_path = Path(sys.argv[3])
repo_root = Path(__file__).resolve().parent.parent
load_env_file(repo_root / ".env")
load_env_file(Path.cwd() / ".env")
json_schema = load_json(schema_path)
doc_id = schema_path.stem
result = run(doc_id=doc_id, file_path=file_path, json_schema=json_schema)
write_json(output_path, result)
if result.get("status") == "ok":
print(f"wrote {output_path}")
else:
error = result.get("error") or {}
print(f"wrote failed result {output_path}: {error.get('type')}: {error.get('message')}")
return 0
if __name__ == "__main__":
sys.exit(main())
|