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: keep ticked PDF form check boxes in knowledge bases (#15572)

Nilaydb9d4f018ee4
diff --git a/studio/backend/core/rag/parsers.py b/studio/backend/core/rag/parsers.py
index c79e409cf5f..9a0a3cafc6d 100644
--- a/studio/backend/core/rag/parsers.py
+++ b/studio/backend/core/rag/parsers.py
@@ -436,6 +436,51 @@ def _pdf_markdown(doc, pages: range | None = None) -> list[str] | None:
return [str(c.get("text") or "") for c in chunks]
+def _pdf_form_field_lines(page) -> list[str]:
+ import fitz
+
+ try:
+ widgets = list(page.widgets())
+ except Exception as exc: # noqa: BLE001 - a malformed field must not fail the whole PDF
+ logger.warning("skipping form fields on PDF page %d: %s", page.number + 1, exc)
+ return []
+ tooltip_owners: dict[str, set[str]] = {}
+ for widget in widgets:
+ tooltip = " ".join((widget.field_label or "").split())
+ tooltip_owners.setdefault(tooltip, set()).add(widget.field_name)
+ fields = {}
+ for widget in widgets:
+ shown = dict(c for c in widget.choice_values or [] if isinstance(c, (list, tuple)))
+ value = widget.field_value
+ value = ", ".join(
+ shown.get(part, part)
+ for part in (value if isinstance(value, list) else [value])
+ if isinstance(part, str)
+ )
+ radio = widget.field_type == fitz.PDF_WIDGET_TYPE_RADIOBUTTON
+ # Kids of one field share its /V; the widget's own /AS says which kid is ticked.
+ unchecked = (radio or widget.field_type == fitz.PDF_WIDGET_TYPE_CHECKBOX) and (
+ value == "Off" or page.parent.xref_get_key(widget.xref, "AS") == ("name", "/Off")
+ )
+ password = (
+ widget.field_type == fitz.PDF_WIDGET_TYPE_TEXT
+ and widget.field_flags & fitz.PDF_TX_FIELD_IS_PASSWORD
+ )
+ _, flags = page.parent.xref_get_key(widget.xref, "F")
+ hidden = widget.field_display in (1, 3) or (
+ flags.isdigit() and int(flags) & fitz.PDF_ANNOT_IS_INVISIBLE
+ )
+ if not widget.field_name or not value.strip() or unchecked or password or hidden:
+ continue
+ # A radio widget's /TU tooltip names one option, not the group's selected value.
+ tooltip = "" if radio else " ".join((widget.field_label or "").split())
+ label = tooltip or widget.field_name
+ if tooltip and len(tooltip_owners[tooltip]) > 1:
+ label = f"{tooltip} ({widget.field_name})"
+ fields[widget.field_name] = f"{label}: {value}"
+ return list(fields.values())
+
+
def _pdf(
source: str | bytes,
want_images: bool,
@@ -453,6 +498,10 @@ def _pdf(
raise ValueError("encrypted PDF requires a password")
total_pages = doc.page_count
page_numbers = range(total_pages if max_pages is None else min(total_pages, max_pages))
+ # Read before pymupdf4llm, which bakes the widgets into page content.
+ form_fields = (
+ {n: _pdf_form_field_lines(doc[n]) for n in page_numbers} if doc.is_form_pdf else {}
+ )
if not config.PDF_MARKDOWN:
md = None
elif max_pages is None:
@@ -491,6 +540,7 @@ def _pdf(
break
if needs_ocr:
text = plain
+ text = "\n".join(filter(None, [text, *form_fields.get(page_number, [])]))
pages.append(Page(text, page_number + 1, len(text), needs_ocr = needs_ocr))
if want_images:
for img in page.get_images(full = True):
diff --git a/studio/backend/tests/test_rag_parsing.py b/studio/backend/tests/test_rag_parsing.py
index 43dd881e75a..35b9b80e11d 100644
--- a/studio/backend/tests/test_rag_parsing.py
+++ b/studio/backend/tests/test_rag_parsing.py
@@ -165,6 +165,134 @@ def test_pdf_markdown_off_uses_plain_text(tmp_path, monkeypatch):
assert "#" not in text and "|" not in text
[email protected]("markdown", [True, False])
+def test_pdf_form_keeps_which_boxes_are_ticked(tmp_path, monkeypatch, markdown):
+ if markdown:
+ pytest.importorskip("pymupdf4llm")
+ import pymupdf
+
+ from core.rag import config, parsers
+
+ monkeypatch.setattr(config, "PDF_MARKDOWN", markdown)
+ doc = pymupdf.open()
+ page = doc.new_page()
+ page.insert_text((40, 60), "Smoker:")
+ for x, answer, ticked in ((120, "Yes", False), (180, "No", True)):
+ widget = pymupdf.Widget()
+ widget.field_type = pymupdf.PDF_WIDGET_TYPE_CHECKBOX
+ widget.field_name = f"smoker_{answer.lower()}"
+ widget.field_label = f"Smoker {answer}"
+ widget.rect = pymupdf.Rect(x, 50, x + 12, 62)
+ widget.field_value = ticked
+ page.add_widget(widget)
+ page.insert_text((x + 16, 60), answer)
+ pdf = tmp_path / "form.pdf"
+ doc.save(str(pdf))
+ doc.close()
+
+ text = "\n".join(p.text for p in parsers.parse(str(pdf)))
+ assert "Smoker No: Yes" in text
+ assert "Smoker Yes" not in text
+
+
[email protected]("shared_field", [True, False])
+def test_pdf_form_names_the_ticked_box_when_boxes_share_a_name_or_tooltip(tmp_path, shared_field):
+ import pymupdf
+
+ from core.rag import parsers
+
+ doc = pymupdf.open()
+ page = doc.new_page()
+ refs = []
+ for x, answer in ((120, "Yes"), (210, "No")):
+ widget = pymupdf.Widget()
+ widget.field_type = pymupdf.PDF_WIDGET_TYPE_CHECKBOX
+ widget.field_name = f"smoker_{answer.lower()}"
+ widget.field_label = f"Smoker {answer}" if shared_field else "Smoker"
+ widget.field_value = answer == ("Yes" if shared_field else "No")
+ widget.rect = pymupdf.Rect(x, 50, x + 12, 62)
+ refs.append(page.add_widget(widget).xref)
+ if shared_field:
+ # One /Btn field, two kids with export values Yes / No; both inherit /V /Yes.
+ parent = doc.get_new_xref()
+ doc.update_object(
+ parent, f"<< /FT /Btn /T (smoker) /V /Yes /Kids [{refs[0]} 0 R {refs[1]} 0 R] >>"
+ )
+ for ref in refs:
+ doc.xref_set_key(ref, "Parent", f"{parent} 0 R")
+ for key in ("T", "FT", "V"):
+ doc.xref_set_key(ref, key, "null")
+ _, on = doc.xref_get_key(refs[1], "AP/N/Yes")
+ _, off = doc.xref_get_key(refs[1], "AP/N/Off")
+ doc.xref_set_key(refs[1], "AP/N", f"<< /Off {off} /No {on} >>")
+ doc.xref_set_key(doc.pdf_catalog(), "AcroForm/Fields", f"[{parent} 0 R]")
+ pdf = tmp_path / "form.pdf"
+ doc.save(str(pdf))
+ doc.close()
+
+ with pymupdf.open(str(pdf)) as reopened:
+ lines = parsers._pdf_form_field_lines(reopened[0])
+ if shared_field:
+ # pymupdf 2.x labels the kid with its parent field ("smoker: Yes").
+ assert len(lines) == 1 and lines[0].endswith(": Yes") and "No" not in lines[0]
+ else:
+ assert lines == ["Smoker (smoker_no): Yes"]
+
+
[email protected]("flag", [1, 2, 32])
+def test_pdf_form_skips_invisible_hidden_and_noview_fields(tmp_path, flag):
+ import pymupdf
+
+ from core.rag import parsers
+
+ doc = pymupdf.open()
+ page = doc.new_page()
+ for name, y in (("shown", 80), ("masked", 120)):
+ widget = pymupdf.Widget()
+ widget.field_type = pymupdf.PDF_WIDGET_TYPE_TEXT
+ widget.field_name = name
+ widget.field_value = f"{name} value"
+ widget.rect = pymupdf.Rect(40, y, 240, y + 20)
+ xref = page.add_widget(widget).xref
+ doc.xref_set_key(xref, "F", str(flag))
+ pdf = tmp_path / "form.pdf"
+ doc.save(str(pdf))
+ doc.close()
+
+ with pymupdf.open(str(pdf)) as reopened:
+ assert parsers._pdf_form_field_lines(reopened[0]) == ["shown: shown value"]
+
+
+# PyMuPDF raises while loading either widget; main never read widgets, so it parsed these.
[email protected]("key, value", [("DA", "(garbage Tf)"), ("Parent", "self")])
[email protected]("markdown", [True, False])
+def test_pdf_with_a_malformed_form_field_still_reads_its_text(
+ tmp_path, monkeypatch, markdown, key, value
+):
+ if markdown:
+ pytest.importorskip("pymupdf4llm")
+ import pymupdf
+
+ from core.rag import config, parsers
+
+ monkeypatch.setattr(config, "PDF_MARKDOWN", markdown)
+ doc = pymupdf.open()
+ page = doc.new_page()
+ page.insert_text((40, 60), "Budget: 4200")
+ widget = pymupdf.Widget()
+ widget.field_type = pymupdf.PDF_WIDGET_TYPE_TEXT
+ widget.field_name = "notes"
+ widget.field_value = "x"
+ widget.rect = pymupdf.Rect(40, 80, 240, 100)
+ xref = page.add_widget(widget).xref
+ doc.xref_set_key(xref, key, f"{xref} 0 R" if value == "self" else value)
+ pdf = tmp_path / "damaged.pdf"
+ doc.save(str(pdf))
+ doc.close()
+
+ assert "Budget: 4200" in "\n".join(p.text for p in parsers.parse(str(pdf)))
+
+
def test_pdf_bytes_use_same_extraction_path(tmp_path, monkeypatch):
from core.rag import config, parsers