Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio RAG: compile the docx run XPath once and remember style list lookups per document (#16343)

Daniel Han2c59e26472a2
diff --git a/studio/backend/core/rag/parsers.py b/studio/backend/core/rag/parsers.py
index 9a0a3cafc6d..cc2c42c382f 100644
--- a/studio/backend/core/rag/parsers.py
+++ b/studio/backend/core/rag/parsers.py
@@ -12,6 +12,7 @@ import os
import re
import secrets
import struct
+import threading
from dataclasses import dataclass
from html.parser import HTMLParser
@@ -986,18 +987,25 @@ def _docx_symbol(sym) -> str:
return chr(code) if 0 <= code <= 0x10FFFF and chr(code).isprintable() else ""
+_docx_xpath = threading.local()
+
+
def _docx_run_text(run) -> str:
- if run.find(f"{_W}fldChar/{_W}ffData/{_W}checkBox") is None and run.find(_W + "sym") is None:
- return run.text
+ # Per thread: run.xpath() recompiles per call; lxml locks an evaluator shared across threads.
+ parts = getattr(_docx_xpath, "run_parts", None)
+ if parts is None:
+ from lxml import etree
+ parts = _docx_xpath.run_parts = etree.XPath(
+ "w:br | w:cr | w:noBreakHyphen | w:ptab | w:t | w:tab | w:fldChar | w:sym",
+ namespaces = {"w": _W[1:-1]},
+ )
return "".join(
_docx_checkbox(e)
if e.tag == _W + "fldChar"
else _docx_symbol(e)
if e.tag == _W + "sym"
else str(e)
- for e in run.xpath(
- "w:br | w:cr | w:noBreakHyphen | w:ptab | w:t | w:tab | w:fldChar | w:sym"
- )
+ for e in parts(run)
)
@@ -1065,17 +1073,26 @@ def _docx_number_lists(document) -> None:
abstracts = {a.get(_W + "abstractNumId"): a for a in numbering.iterchildren(_W + "abstractNum")}
nums = {n.get(_W + "numId"): n for n in numbering.iterchildren(_W + "num")}
+ style_nums = {}
+
# numId and ilvl inherit through basedOn independently, like any other paragraph property.
def style_num(
style_id,
name,
depth = 0,
):
+ if depth == 0 and (style_id, name) in style_nums:
+ return style_nums[style_id, name]
style = style_by_id.get(style_id)
if style is None or depth > 20:
- return None
- value = val(style.find(f"{_W}pPr/{_W}numPr"), name)
- return style_num(val(style, "basedOn"), name, depth + 1) if value is None else value
+ value = None
+ else:
+ value = val(style.find(f"{_W}pPr/{_W}numPr"), name)
+ if value is None:
+ value = style_num(val(style, "basedOn"), name, depth + 1)
+ if depth == 0:
+ style_nums[style_id, name] = value
+ return value
def linked(abstract_id, depth = 0):
num_id = style_num(val(abstracts.get(abstract_id), "numStyleLink"), "numId")
diff --git a/studio/backend/tests/test_rag_docx_run_text.py b/studio/backend/tests/test_rag_docx_run_text.py
new file mode 100644
index 00000000000..f10a4d1e471
--- /dev/null
+++ b/studio/backend/tests/test_rag_docx_run_text.py
@@ -0,0 +1,90 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import itertools
+import sys
+import threading
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+ sys.path.insert(0, _BACKEND_DIR)
+
+pytest.importorskip("docx")
+import docx # noqa: E402
+from docx.oxml import OxmlElement # noqa: E402
+from docx.oxml.ns import qn # noqa: E402
+
+from core.rag import parsers # noqa: E402
+
+
+def _child(tag, **attrs):
+ element = OxmlElement(tag)
+ for key, value in attrs.items():
+ element.set(qn(key), value)
+ return element
+
+
+def _children():
+ text = _child("w:t")
+ text.text = "word"
+ spaced = _child("w:t")
+ spaced.text = " gap "
+ deleted = _child("w:delText")
+ deleted.text = "gone"
+ instr = _child("w:instrText")
+ instr.text = " PAGE "
+ return [
+ text,
+ spaced,
+ _child("w:tab"),
+ _child("w:br"),
+ _child("w:br", **{"w:type": "page"}),
+ _child("w:br", **{"w:type": "column"}),
+ _child("w:br", **{"w:type": "textWrapping"}),
+ _child("w:cr"),
+ _child("w:noBreakHyphen"),
+ _child("w:softHyphen"),
+ _child("w:ptab", **{"w:relativeTo": "margin", "w:alignment": "right", "w:leader": "none"}),
+ deleted,
+ instr,
+ ]
+
+
+def _run_with(parts):
+ document = docx.Document()
+ run = document.add_paragraph().add_run()
+ for part in parts:
+ run._r.append(part)
+ return run
+
+
+def test_run_text_matches_python_docx_for_every_text_child():
+ """Without checkbox fields or symbols, the run text is exactly python-docx's Run.text."""
+ kinds = len(_children())
+ for i, j in itertools.product(range(kinds), repeat = 2):
+ fresh = _children()
+ run = _run_with([fresh[i], fresh[j]])
+ assert parsers._docx_run_text(run._r) == run.text, (i, j)
+ run = _run_with(_children())
+ assert parsers._docx_run_text(run._r) == run.text
+
+
+def test_run_text_is_the_same_on_every_thread():
+ run = _run_with(_children())
+ expected = run.text
+ results = []
+
+ def read():
+ results.extend(parsers._docx_run_text(run._r) for _ in range(50))
+
+ threads = [threading.Thread(target = read) for _ in range(4)]
+ for thread in threads:
+ thread.start()
+ for thread in threads:
+ thread.join()
+ assert results == [expected] * 200