File size: 20,125 Bytes
547ce6e | 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 | """Turn whatever the user brought into pages of text.
One shape out β :class:`ExtractedDocument` β regardless of what went in, so everything
downstream (chunking, retrieval, reranking, citation verification) is the same code that
serves the law corpus. The formats differ only here.
PDF with a text layer PyMuPDF, the same reader the corpus uses
PDF without one rendered per page and sent to OCR
DOCX python-docx, headings kept as structure
TXT / MD decoded, headings inferred from markdown
HTML / URL fetched, stripped, with SSRF defences below
Images OCR
**OCR is deliberately two-tier.** Tesseract always works and costs nothing; Claude's
vision model is markedly better on a photographed page and costs a call. That is the same
split the rest of the project already makes between `offline-extractive` and
`Claude Β· grounded`, and it is labelled the same way, so a demo without a key degrades
visibly rather than silently.
Nothing here trusts its input. Every reader is wrapped so a malformed file returns a
message a user can act on instead of a 500, and the URL fetcher treats the URL as hostile
by default β see :func:`_assert_public_url`.
"""
from __future__ import annotations
import io
import ipaddress
import logging
import re
import socket
import unicodedata
from dataclasses import dataclass, field
from html.parser import HTMLParser
from typing import Any, Final
from urllib.parse import urlparse
from app.core.settings import Settings, get_settings
logger = logging.getLogger(__name__)
PDF_MAGIC: Final = b"%PDF"
DOCX_MAGIC: Final = b"PK\x03\x04"
IMAGE_SUFFIXES: Final = frozenset({".png", ".jpg", ".jpeg", ".webp", ".tif", ".tiff", ".bmp"})
TEXT_SUFFIXES: Final = frozenset({".txt", ".md", ".markdown", ".rst", ".csv", ".log"})
HTML_SUFFIXES: Final = frozenset({".html", ".htm"})
# A scanned page usually yields a handful of stray glyphs rather than nothing at all, so
# "has a text layer" cannot be `text != ""`. Below this many characters per page the PDF
# is treated as an image of a document.
TEXT_LAYER_CHARS_PER_PAGE: Final = 40
class ExtractionError(Exception):
"""Input that could not be read. The message is shown to the user verbatim."""
@dataclass(frozen=True, slots=True)
class ExtractedPage:
page_no: int
text: str
@dataclass(frozen=True, slots=True)
class ExtractedDocument:
title: str
source: str
kind: str
pages: tuple[ExtractedPage, ...]
ocr_engine: str | None = None
diagnostics: dict[str, Any] = field(default_factory=dict)
@property
def text(self) -> str:
return "\n\n".join(page.text for page in self.pages if page.text.strip())
# ββ text hygiene βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def normalise(text: str) -> str:
"""Fold the encoding damage that arrives with real documents.
NFKC collapses the ligatures and full-width forms that PDF producers emit; the corpus
parser fixes the same class of defect (AUDIT.md records a broken ligature that reached
the BM25 index). Zero-width characters are dropped outright: they are invisible, they
survive into a chunk, and they silently break exact-term matching.
"""
cleaned = unicodedata.normalize("NFKC", text)
cleaned = cleaned.replace("\u200b", "").replace("ο»Ώ", "").replace("\xad", "")
cleaned = cleaned.replace("\r\n", "\n").replace("\r", "\n")
cleaned = re.sub(r"[ \t]+", " ", cleaned)
return re.sub(r"\n{3,}", "\n\n", cleaned).strip()
# ββ HTML βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class _Readable(HTMLParser):
"""Strip a page to readable text.
Written against the standard library rather than adding a parser dependency: the job
is to drop script/style/nav and keep block structure, which does not need a full DOM.
"""
_SKIP: Final = frozenset({"script", "style", "noscript", "svg", "head", "nav", "footer"})
_BLOCK: Final = frozenset(
{"p", "div", "br", "li", "tr", "section", "article", "h1", "h2", "h3", "h4", "h5", "h6"}
)
def __init__(self) -> None:
super().__init__(convert_charrefs=True)
self.parts: list[str] = []
self.title: str = ""
self._skip_depth = 0
self._in_title = False
def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
del attrs # the readable-text pass needs tag structure, not attributes
if tag in self._SKIP:
self._skip_depth += 1
elif tag == "title":
self._in_title = True
elif tag in self._BLOCK:
self.parts.append("\n")
def handle_endtag(self, tag: str) -> None:
if tag in self._SKIP:
self._skip_depth = max(0, self._skip_depth - 1)
elif tag == "title":
self._in_title = False
elif tag in self._BLOCK:
self.parts.append("\n")
def handle_data(self, data: str) -> None:
if self._in_title:
self.title += data
elif self._skip_depth == 0:
self.parts.append(data)
def html_to_text(html: str) -> tuple[str, str]:
"""Return ``(title, text)``."""
parser = _Readable()
parser.feed(html)
return parser.title.strip(), normalise("".join(parser.parts))
# ββ URL fetching βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _assert_public_url(url: str) -> None:
"""Refuse anything that could reach the host's own network.
The server fetches a URL chosen by a stranger, which is a server-side request forgery
primitive unless it is constrained. Every address the hostname resolves to is checked,
not merely the first: a name that returns one public and one loopback address would
otherwise pass this and then connect to the loopback one.
This is defence in depth and not a complete answer β it cannot stop a DNS record that
changes between this check and the connection. The size cap and the redirect ceiling
in :func:`fetch_url` bound what a successful attempt could achieve.
"""
parsed = urlparse(url)
if parsed.scheme not in {"http", "https"}:
raise ExtractionError(f"only http and https URLs are accepted, not {parsed.scheme!r}")
host = parsed.hostname
if not host:
raise ExtractionError("that URL has no host")
try:
resolved = socket.getaddrinfo(
host, parsed.port or (443 if parsed.scheme == "https" else 80)
)
except socket.gaierror as exc:
raise ExtractionError(f"could not resolve {host}") from exc
for info in resolved:
address = ipaddress.ip_address(str(info[4][0]))
if (
address.is_private
or address.is_loopback
or address.is_link_local # 169.254.169.254 is the cloud metadata endpoint
or address.is_reserved
or address.is_multicast
or address.is_unspecified
):
raise ExtractionError(
f"{host} resolves to {address}, which is not a public address. "
"Only public URLs can be fetched."
)
def fetch_url(url: str, settings: Settings | None = None) -> ExtractedDocument:
import httpx
cfg = settings or get_settings()
limit = cfg.workspace_max_bytes
_assert_public_url(url)
try:
with httpx.Client(
follow_redirects=False,
timeout=cfg.workspace_fetch_timeout_s,
headers={"User-Agent": "Lexora/1.0 (+document ingest)"},
) as client:
response = client.get(url)
# Redirects are followed by hand so each hop is re-validated. `follow_redirects`
# would check only the URL the user supplied and then happily chase a 302 into
# the private network the check above exists to keep it out of.
for _ in range(cfg.workspace_max_redirects):
if response.status_code not in {301, 302, 303, 307, 308}:
break
location = response.headers.get("location", "")
if not location:
break
nxt = str(httpx.URL(url).join(location))
_assert_public_url(nxt)
url, response = nxt, client.get(nxt)
response.raise_for_status()
except httpx.HTTPStatusError as exc:
raise ExtractionError(f"{url} returned HTTP {exc.response.status_code}") from exc
except httpx.HTTPError as exc:
raise ExtractionError(f"could not fetch {url}: {exc}") from exc
body = response.content[: limit + 1]
if len(body) > limit:
raise ExtractionError(f"{url} is larger than the {limit // 1_000_000} MB limit")
content_type = response.headers.get("content-type", "").split(";")[0].strip()
if content_type == "application/pdf" or body[:4] == PDF_MAGIC:
document = extract_bytes(body, filename=url.rsplit("/", 1)[-1] or "page.pdf", settings=cfg)
return ExtractedDocument(
title=document.title,
source=url,
kind="pdf",
pages=document.pages,
ocr_engine=document.ocr_engine,
diagnostics=document.diagnostics,
)
charset = response.encoding or "utf-8"
decoded = body.decode(charset, errors="replace")
if "html" in content_type or decoded.lstrip()[:1] == "<":
title, text = html_to_text(decoded)
else:
title, text = "", normalise(decoded)
if not text.strip():
raise ExtractionError(f"no readable text found at {url}")
return ExtractedDocument(
title=title or url,
source=url,
kind="html",
pages=(ExtractedPage(1, text),),
diagnostics={"content_type": content_type, "bytes": len(body)},
)
# ββ OCR ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _ocr_tesseract(image_bytes: bytes) -> str | None:
"""Return text, or ``None`` when Tesseract is not installed in this image."""
try:
import pytesseract
from PIL import Image
except ImportError:
return None
try:
with Image.open(io.BytesIO(image_bytes)) as image:
return str(pytesseract.image_to_string(image))
except Exception as exc:
logger.warning("tesseract failed: %s", exc)
return None
def _ocr_claude(image_bytes: bytes, media_type: str, settings: Settings) -> str | None:
"""Transcribe with Claude's vision model. ``None`` when no key is configured."""
if not settings.use_anthropic:
return None
import base64
from app.core.claude import get_client
try:
client = get_client(settings)
if client is None:
return None
message = client.messages.create(
model=settings.answer_model,
max_tokens=4096,
temperature=0,
messages=[
{
"role": "user",
"content": [
{
"type": "image",
"source": {
"type": "base64",
"media_type": media_type,
"data": base64.b64encode(image_bytes).decode(),
},
},
{
"type": "text",
# Transcribe, do not summarise: this text becomes a citation,
# and a citation that paraphrases its source is worse than no
# citation at all.
"text": (
"Transcribe every word of this document image exactly as "
"written, preserving headings, numbering and paragraph "
"breaks. Do not summarise, explain, or add commentary. "
"If the image contains no legible text, reply with "
"exactly: NO_TEXT"
),
},
],
}
],
)
parts = [block.text for block in message.content if getattr(block, "type", "") == "text"]
text = "\n".join(parts).strip()
except Exception as exc:
logger.warning("claude vision OCR failed: %s", exc)
return None
# NO_TEXT is the model's own signal that the image is illegible. Returning None lets
# the caller try Tesseract, which occasionally reads what the model declined to.
return None if text == "NO_TEXT" else text
def ocr(image_bytes: bytes, media_type: str, settings: Settings) -> tuple[str, str]:
"""Return ``(text, engine)``. Claude when a key is present, else Tesseract."""
text = _ocr_claude(image_bytes, media_type, settings)
if text:
return normalise(text), "claude-vision"
text = _ocr_tesseract(image_bytes)
if text:
return normalise(text), "tesseract"
raise ExtractionError(
"no text could be read from that image. Claude vision needs "
"LEXORA_ANTHROPIC_API_KEY, and Tesseract is not installed in this container."
)
# ββ PDF ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _extract_pdf(
data: bytes, settings: Settings
) -> tuple[list[ExtractedPage], str | None, dict[str, Any]]:
import fitz
try:
document = fitz.open(stream=data, filetype="pdf")
except Exception as exc:
raise ExtractionError(f"that PDF could not be opened: {exc}") from exc
with document:
if document.needs_pass:
raise ExtractionError("that PDF is password protected")
page_count = min(document.page_count, settings.workspace_max_pages)
raw = [(index + 1, document[index].get_text("text")) for index in range(page_count)]
total_chars = sum(len(text) for _, text in raw)
# A scanned PDF has pages but effectively no text layer. Render and OCR instead of
# indexing a handful of stray glyphs and calling the document empty.
if page_count and total_chars < TEXT_LAYER_CHARS_PER_PAGE * page_count:
pages: list[ExtractedPage] = []
engine: str | None = None
for index in range(page_count):
pixmap = document[index].get_pixmap(dpi=200)
text, engine = ocr(pixmap.tobytes("png"), "image/png", settings)
pages.append(ExtractedPage(index + 1, text))
return (
pages,
engine,
{
"pages": page_count,
"text_layer": False,
"truncated": document.page_count > page_count,
},
)
return (
[ExtractedPage(no, normalise(text)) for no, text in raw],
None,
{
"pages": page_count,
"text_layer": True,
"truncated": document.page_count > page_count,
},
)
# ββ DOCX βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _extract_docx(data: bytes) -> list[ExtractedPage]:
try:
import docx
except ImportError as exc: # pragma: no cover - declared dependency
raise ExtractionError("DOCX support is not installed in this container") from exc
try:
document = docx.Document(io.BytesIO(data))
except Exception as exc:
raise ExtractionError(f"that DOCX could not be opened: {exc}") from exc
lines: list[str] = []
for paragraph in document.paragraphs:
text = paragraph.text.strip()
if not text:
continue
# Word headings become markdown so the chunker's heading detection sees the
# structure the author intended rather than an undifferentiated wall of text.
style = (paragraph.style.name or "").lower() if paragraph.style else ""
if style.startswith("heading"):
depth = "".join(ch for ch in style if ch.isdigit()) or "1"
lines.append(f"{'#' * min(int(depth), 6)} {text}")
else:
lines.append(text)
for table in document.tables:
for row in table.rows:
cells = [cell.text.strip() for cell in row.cells]
if any(cells):
lines.append(" | ".join(cells))
text = normalise("\n\n".join(lines))
if not text:
raise ExtractionError("that DOCX contains no readable text")
# Word has no fixed pages without rendering; one page keeps citations honest rather
# than inventing page numbers that do not correspond to anything.
return [ExtractedPage(1, text)]
# ββ entry point ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def extract_bytes(
data: bytes,
filename: str,
content_type: str | None = None,
settings: Settings | None = None,
) -> ExtractedDocument:
"""Read an uploaded file. Dispatch is by magic bytes first, extension second."""
cfg = settings or get_settings()
if not data:
raise ExtractionError("that file is empty")
if len(data) > cfg.workspace_max_bytes:
raise ExtractionError(
f"that file is larger than the {cfg.workspace_max_bytes // 1_000_000} MB limit"
)
name = filename.strip() or "document"
suffix = ("." + name.rsplit(".", 1)[-1].lower()) if "." in name else ""
title = name.rsplit("/", 1)[-1]
# Content sniffing before the extension: a `.txt` that is really a PDF should be read
# as a PDF, and an extension is a claim by whoever uploaded the file.
if data[:4] == PDF_MAGIC:
pages, engine, diagnostics = _extract_pdf(data, cfg)
document = ExtractedDocument(title, name, "pdf", tuple(pages), engine, diagnostics)
elif data[:4] == DOCX_MAGIC and suffix == ".docx":
document = ExtractedDocument(title, name, "docx", tuple(_extract_docx(data)))
elif suffix in IMAGE_SUFFIXES or (content_type or "").startswith("image/"):
media = content_type if content_type and content_type.startswith("image/") else "image/png"
text, engine = ocr(data, media, cfg)
document = ExtractedDocument(title, name, "image", (ExtractedPage(1, text),), engine)
elif suffix in HTML_SUFFIXES:
html_title, text = html_to_text(data.decode("utf-8", errors="replace"))
document = ExtractedDocument(html_title or title, name, "html", (ExtractedPage(1, text),))
elif suffix in TEXT_SUFFIXES or not suffix:
text = normalise(data.decode("utf-8", errors="replace"))
document = ExtractedDocument(title, name, "text", (ExtractedPage(1, text),))
else:
raise ExtractionError(
f"{suffix or 'that file type'} is not supported. Upload a PDF, DOCX, image, "
"HTML or plain-text file, or paste a link."
)
if not document.text.strip():
raise ExtractionError(
f"no readable text was found in {title}. If it is a scan, OCR produced nothing."
)
return document
|