DocMorph / docmorph /parser.py
Shoufa Zhang
Deploy DocMorph & SeekMorph interactive Studio and 5-layer production benchmark
66227f0
Raw History Blame Contribute Delete
15 kB
"""Multi-format structure-aware document morphing and table-header-preserving chunker."""
from __future__ import annotations
import csv
import hashlib
import io
import json
import re
from dataclasses import asdict, dataclass, field
from html.parser import HTMLParser
from pathlib import Path
from typing import Any
@dataclass(frozen=True)
class DocChunk:
"""A structure-preserving document chunk with section breadcrumbs and neighbor links."""
chunk_id: str
doc_id: str
chunk_index: int
section_path: tuple[str, ...]
content: str
kind: str = "prose" # "prose" | "table" | "code"
namespace: str = "default"
metadata: dict[str, Any] = field(default_factory=dict)
@property
def breadcrumb_str(self) -> str:
"""Human-readable section breadcrumb."""
return " > ".join(self.section_path) if self.section_path else self.doc_id
def to_dict(self) -> dict[str, Any]:
"""Serialize chunk to a JSON-compatible dictionary."""
data = asdict(self)
data["section_path"] = list(self.section_path)
return data
@classmethod
def from_dict(cls, data: dict[str, Any]) -> DocChunk:
"""Deserialize chunk from a dictionary."""
return cls(
chunk_id=str(data["chunk_id"]),
doc_id=str(data["doc_id"]),
chunk_index=int(data.get("chunk_index", 0)),
section_path=tuple(data.get("section_path") or ()),
content=str(data["content"]),
kind=str(data.get("kind", "prose")),
namespace=str(data.get("namespace", "default")),
metadata=dict(data.get("metadata") or {}),
)
class _HTMLToMarkdownParser(HTMLParser):
"""Lightweight HTML-to-Markdown converter preserving headings, tables, code blocks, and lists."""
_IGNORED_TAGS = frozenset({"script", "style", "noscript", "svg", "head"})
def __init__(self) -> None:
super().__init__(convert_charrefs=True)
self._out: list[str] = []
self._ignore_depth = 0
self._in_pre = False
self._in_table = False
self._table_rows: list[list[str]] = []
self._current_row: list[str] = []
self._current_cell: list[str] = []
self._in_cell = False
def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
t = tag.lower()
if t in self._IGNORED_TAGS:
self._ignore_depth += 1
return
if self._ignore_depth > 0:
return
if t in {"h1", "h2", "h3", "h4", "h5", "h6"}:
level = int(t[1])
self._out.append(f"\n\n{'#' * level} ")
elif t == "p" or t == "div" or t == "section" or t == "article":
self._out.append("\n\n")
elif t == "br":
self._out.append("\n")
elif t == "li":
self._out.append("\n- ")
elif t == "pre":
self._in_pre = True
self._out.append("\n\n```\n")
elif t == "code" and not self._in_pre:
self._out.append("`")
elif t == "table":
self._in_table = True
self._table_rows = []
elif t == "tr" and self._in_table:
self._current_row = []
elif t in {"th", "td"} and self._in_table:
self._in_cell = True
self._current_cell = []
def handle_endtag(self, tag: str) -> None:
t = tag.lower()
if t in self._IGNORED_TAGS:
self._ignore_depth = max(0, self._ignore_depth - 1)
return
if self._ignore_depth > 0:
return
if t in {"h1", "h2", "h3", "h4", "h5", "h6"}:
self._out.append("\n\n")
elif t == "pre":
self._in_pre = False
self._out.append("\n```\n\n")
elif t == "code" and not self._in_pre:
self._out.append("`")
elif t in {"th", "td"} and self._in_table and self._in_cell:
cell_text = " ".join("".join(self._current_cell).split()).replace("|", "\\|")
self._current_row.append(cell_text)
self._in_cell = False
elif t == "tr" and self._in_table:
if any(self._current_row):
self._table_rows.append(self._current_row)
self._current_row = []
elif t == "table" and self._in_table:
self._in_table = False
if self._table_rows:
cols = max(len(r) for r in self._table_rows)
padded = [r + [""] * (cols - len(r)) for r in self._table_rows]
header = "| " + " | ".join(padded[0]) + " |"
sep = "| " + " | ".join(["---"] * cols) + " |"
body = ["| " + " | ".join(r) + " |" for r in padded[1:]]
self._out.append("\n\n" + "\n".join([header, sep, *body]) + "\n\n")
self._table_rows = []
def handle_data(self, data: str) -> None:
if self._ignore_depth > 0:
return
if self._in_cell:
self._current_cell.append(data)
elif self._in_pre:
self._out.append(data)
else:
collapsed = re.sub(r"[ \t\r\n]+", " ", data)
self._out.append(collapsed)
def get_markdown(self) -> str:
# Flush any trailing unclosed table or cell
if self._in_table:
if self._in_cell:
cell_text = " ".join("".join(self._current_cell).split()).replace("|", "\\|")
self._current_row.append(cell_text)
self._in_cell = False
if any(self._current_row):
self._table_rows.append(self._current_row)
self._current_row = []
if self._table_rows:
cols = max(len(r) for r in self._table_rows)
padded = [r + [""] * (cols - len(r)) for r in self._table_rows]
header = "| " + " | ".join(padded[0]) + " |"
sep = "| " + " | ".join(["---"] * cols) + " |"
body = ["| " + " | ".join(r) + " |" for r in padded[1:]]
self._out.append("\n\n" + "\n".join([header, sep, *body]) + "\n\n")
self._in_table = False
if self._in_pre:
self._out.append("\n```\n\n")
self._in_pre = False
raw = "".join(self._out)
return re.sub(r"\n{3,}", "\n\n", raw).strip()
def html_to_markdown(html_text: str) -> str:
"""Convert raw HTML into structured Markdown preserving headings, tables, and code blocks."""
parser = _HTMLToMarkdownParser()
parser.feed(html_text)
parser.close()
return parser.get_markdown()
def csv_to_markdown_table(csv_text: str, delimiter: str = ",") -> str:
"""Convert CSV/TSV text into a normalized Markdown table."""
reader = csv.reader(io.StringIO(csv_text.strip()), delimiter=delimiter)
rows = [[cell.strip().replace("|", "\\|") for cell in row] for row in reader if any(c.strip() for c in row)]
if not rows:
return ""
cols = max(len(r) for r in rows)
padded = [r + [""] * (cols - len(r)) for r in rows]
header = "| " + " | ".join(padded[0]) + " |"
sep = "| " + " | ".join(["---"] * cols) + " |"
body = ["| " + " | ".join(r) + " |" for r in padded[1:]]
return "\n".join([header, sep, *body])
_HEADING_RE = re.compile(r"^(#{1,6})\s+(.+?)\s*$")
_FENCE_RE = re.compile(r"^```")
_TABLE_ROW_RE = re.compile(r"^\|.*\|$")
_TABLE_SEP_RE = re.compile(r"^\|\s*:?-{2,}:?\s*(\|\s*:?-{2,}:?\s*)*\|?$")
class DocumentMorpher:
"""Structure-aware document chunker that preserves code blocks and repeats table headers across splits."""
SUPPORTED_EXTENSIONS = frozenset({".md", ".markdown", ".txt", ".html", ".htm", ".csv", ".tsv", ".json"})
def __init__(self, max_chars: int = 1000) -> None:
if max_chars < 80:
raise ValueError("max_chars must be >= 80")
self.max_chars = max_chars
def morph_text(
self,
doc_id: str,
text: str,
*,
fmt: str = "markdown",
namespace: str = "default",
metadata: dict[str, Any] | None = None,
) -> list[DocChunk]:
"""Normalize text from markdown/html/csv/json/txt and split into structure-preserving chunks."""
fmt_lower = fmt.lower().lstrip(".")
if fmt_lower in {"html", "htm"}:
normalized = html_to_markdown(text)
elif fmt_lower == "csv":
normalized = csv_to_markdown_table(text, delimiter=",")
elif fmt_lower == "tsv":
normalized = csv_to_markdown_table(text, delimiter="\t")
elif fmt_lower == "json":
try:
parsed = json.loads(text)
normalized = f"```json\n{json.dumps(parsed, ensure_ascii=False, indent=2)}\n```"
except json.JSONDecodeError:
normalized = text
else:
normalized = text
return self._chunk_markdown(doc_id, normalized, namespace=namespace, base_metadata=metadata)
def morph_file(
self,
path: str | Path,
*,
root_dir: str | Path | None = None,
namespace: str = "default",
) -> list[DocChunk]:
"""Parse and chunk a single file on disk."""
file_path = Path(path).resolve()
rel_id = str(file_path.relative_to(Path(root_dir).resolve())) if root_dir else file_path.name
ext = file_path.suffix.lower()
raw_text = file_path.read_text(encoding="utf-8", errors="replace")
return self.morph_text(
doc_id=rel_id.replace("\\", "/"),
text=raw_text,
fmt=ext,
namespace=namespace,
metadata={"source_path": str(file_path)},
)
def morph_directory(
self,
directory: str | Path,
*,
namespace: str = "default",
) -> list[DocChunk]:
"""Recursively parse and chunk all supported documents in a directory."""
root = Path(directory).resolve()
if root.is_file():
return self.morph_file(root, namespace=namespace)
all_chunks: list[DocChunk] = []
for path in sorted(root.rglob("*")):
if any(part.startswith(".") for part in path.relative_to(root).parts):
continue
if path.is_file() and path.suffix.lower() in self.SUPPORTED_EXTENSIONS:
all_chunks.extend(self.morph_file(path, root_dir=root, namespace=namespace))
return all_chunks
def _chunk_markdown(
self,
doc_id: str,
text: str,
*,
namespace: str,
base_metadata: dict[str, Any] | None,
) -> list[DocChunk]:
meta = dict(base_metadata or {})
heading_stack: list[tuple[int, str]] = []
chunks: list[DocChunk] = []
prose_buf: list[str] = []
lines = text.splitlines()
idx = 0
def current_section() -> tuple[str, ...]:
return tuple(title for _, title in heading_stack)
def emit(content: str, kind: str) -> None:
cleaned = content.strip()
if not cleaned:
return
sec = current_section()
chunk_idx = len(chunks)
digest = hashlib.sha256(
f"{namespace}|{doc_id}|{chunk_idx}|{'/'.join(sec)}|{kind}|{cleaned}".encode("utf-8")
).hexdigest()[:12]
chunks.append(
DocChunk(
chunk_id=f"{doc_id}#{chunk_idx}:{digest}",
doc_id=doc_id,
chunk_index=chunk_idx,
section_path=sec,
content=cleaned,
kind=kind,
namespace=namespace,
metadata=dict(meta),
)
)
def flush_prose() -> None:
if not prose_buf:
return
joined = "\n".join(prose_buf).strip()
prose_buf.clear()
if not joined:
return
while len(joined) > self.max_chars:
split_at = joined.rfind("\n\n", 0, self.max_chars)
if split_at <= 0:
split_at = joined.rfind("\n", 0, self.max_chars)
if split_at <= 0:
split_at = self.max_chars
emit(joined[:split_at], "prose")
joined = joined[split_at:].lstrip()
if joined:
emit(joined, "prose")
def emit_table(table_lines: list[str]) -> None:
full_table = "\n".join(table_lines).strip()
if len(full_table) <= self.max_chars or len(table_lines) <= 2:
emit(full_table, "table")
return
has_header = len(table_lines) >= 2 and bool(_TABLE_SEP_RE.match(table_lines[1].strip()))
header_prefix = table_lines[:2] if has_header else []
data_rows = table_lines[2:] if has_header else table_lines
current_batch = list(header_prefix)
current_len = sum(len(r) + 1 for r in current_batch)
for row in data_rows:
row_len = len(row) + 1
if len(current_batch) > len(header_prefix) and (current_len + row_len > self.max_chars):
emit("\n".join(current_batch), "table")
current_batch = list(header_prefix)
current_len = sum(len(r) + 1 for r in current_batch)
current_batch.append(row)
current_len += row_len
if len(current_batch) > len(header_prefix):
emit("\n".join(current_batch), "table")
while idx < len(lines):
line = lines[idx]
heading_match = _HEADING_RE.match(line)
if heading_match:
flush_prose()
level = len(heading_match.group(1))
title = heading_match.group(2).strip()
while heading_stack and heading_stack[-1][0] >= level:
heading_stack.pop()
heading_stack.append((level, title))
idx += 1
continue
if _FENCE_RE.match(line.strip()):
flush_prose()
code_lines = [line]
idx += 1
while idx < len(lines):
code_lines.append(lines[idx])
if _FENCE_RE.match(lines[idx].strip()):
idx += 1
break
idx += 1
emit("\n".join(code_lines), "code")
continue
if _TABLE_ROW_RE.match(line.strip()):
flush_prose()
table_lines = [line]
idx += 1
while idx < len(lines) and _TABLE_ROW_RE.match(lines[idx].strip()):
table_lines.append(lines[idx])
idx += 1
emit_table(table_lines)
continue
prose_buf.append(line)
idx += 1
flush_prose()
return chunks