Raven-1/tests/test_knowledge_deepdoc_layout_assignment.py
Kevin Hu 82c76a82a9
feat(knowledge): read every format, show what the index holds, and let an agent search it (#489)
<img width="2318" height="1448" alt="image"
src="https://github.com/user-attachments/assets/bca71d69-4972-420c-b14e-c8cf74a17de9"
/>

## Summary

Everything a knowledge base could not read, and the pieces it cut them
into
saying what they are.

**Reading a PDF.** A PDF records glyphs at coordinates; whether they are
a
heading, a table cell or a running footer is inferred from how they
look. There
are now two readers. The geometric one infers structure from type size
and
position and needs nothing but the file. The other asks deepdoc's
models, ported
into `raven/knowledge/parser/deepdoc`: a layout model that says what
each region
is, a detector and recognizer that read a page carrying no text at all,
and a
table-structure model that rebuilds a grid from a picture of one. 
 Detecting boxes around text the file already spells
out exactly costs an inference pass per page to learn less, and
recognising it
costs accuracy.

The layout graph answers with a shape its own postprocessing
raises on, and its classes are DocLayout-YOLO's order rather than the
list
Both were established by running the model, and the label order is
pinned by a test.

**Tables go in as tables.** `<table>` is the only shape that keeps what
a grid
states -- which column a cell stood under, which cells are merged, which
row is
the header -- and it is the shape a model reads a table in, which is
what these
chunks are for. PDF, Word and PowerPoint all state their merges exactly,
and
they now survive as `colspan` and `rowspan`. The flat form this replaces
had to
repeat a header standing over two columns across both of them, because
there was
nowhere to record a span; it had no callers left afterwards and is gone.
Two
deviations from upstream, both about what the string is for: every
cell's text
is escaped, where upstream interpolates it raw and a cell holding a `<`
produces
markup that is no longer a table, and the span attributes are written
only where
there is a span, where upstream emits `<td  >` for every ordinary cell.

**A picture of where each chunk came from.** A chunk is a few hundred
words with
a page number on it, which is enough to cite and not enough to answer
whether
the parser read the page correctly. So every PDF chunk that knows where
it sits
carries one image: the regions it covers, cropped out of the rendered
pages and
stacked. here they are already structured in the parser's element spans,
so
the markers would only put text nobody wrote into the thing being
embedded.
The panel draws a chunk's tables as tables, parsed and
rebuilt as elements rather than injected: a person can rewrite any piece
from
that same panel, so what comes back from the store is user input however
it got
there.

**The page that reads them.** Upstream retired the knowledge page while
this
branch was out; it is rebuilt against the new architecture, and the half
that
never existed is the one a reader needs -- what a document actually
became once
it was cut and embedded. Documents sit in one-level folders and arrive
by four
routes: a file, a folder, a URL, a written note. Adding and indexing are
two
calls on the contract and nothing drains that queue on its own, so a
page that
stopped after the first left every upload saying Queued for good.

Opening a file takes the whole page: the document on the left, the
pieces the
index holds of it on the right. The left half is the file drawn as its
own
pages rather than framed by the browser's PDF viewer, which reads where
to open
once and answers to nothing afterwards -- and which nothing can be drawn
on.
Pages are rendered once and kept, and the region a piece was cut from is
washed
over the page it came from, placed in fractions of the page so the
picture's
own scale never enters the arithmetic. Clicking a piece takes the file
there;
double-clicking opens it for rewriting.

The pieces can be searched, filtered by state, turned on and off,
deleted,
rewritten and added to, one at a time or in a batch. Every one of those
was
already on the contract with no reader for it. A base's settings are
back and
fuller: how many pieces a search returns, how documents are cut, and how
much
prose comes with a table or a figure. The embedding model is a picker
while the
base holds no pieces and a stated fact once it does -- moving it then is
the
collection made again at the new width and every document re-cut, which
is not
a thing a panel does by being saved.

**An agent can read them.** The retrieval engine was there and nothing
could
reach it: no agent had a tool that opened a library. A reader now points
a
conversation at bases from the composer, the way they pin a working
directory,
and the selection rides in the same place -- the session's metadata,
bound
around the turn.

The tool takes a query and never a base. The scope is the reader's
choice, so a
model cannot reach material nobody offered it, and it is withheld where
nothing
was picked rather than offered and refused. A hit says which document
and page
it came from, because a hit a reader cannot place is a hit they cannot
check --
and the passages are pressable in the transcript, walking into the base,
the
document, the folder it is filed under and then the piece. A citation a
reader
cannot follow is a claim they have to take on trust, and the whole point
of
searching their own material is that they can go and look.

**The word segmenter and the resources.** `raven.core.Tokenizer` is
Infinity's
`rag_tokenizer`, ported: a forward and backward maximum match over a
dictionary
trie, with stemming and lemmatisation on the latin path. XGBoost is not
carried
over. None of the data lives in git -- the dictionary, the NLTK corpora
and the
ONNX graphs run to about a hundred megabytes, so `make fetch-resources`
on an
install and an optional Docker layer behind `RAVEN_RESOURCES` both point
at one
script. A checkout that could not reach the network still works, and
every
feature that reads this data says so when asked.

## Type

- [ ] Fix
- [x] Feature
- [ ] Docs
- [ ] CI / tooling
- [ ] Refactor
- [ ] Other

## Verification

Run on macOS against Python 3.12, with the deepdoc weights fetched (the
model
tests skip without them):

```
uv run pytest                        27450 passed, 111 skipped, 3 failed
npm run test --prefix ui-web          3276 passed
npm run test --prefix ui-tui          2106 passed
npm run type-check --prefix ui-web    clean (tsc)
npm run lint --prefix ui-web          clean (4 pre-existing hook warnings)
make lint-python                      clean (ruff check + format)
make lint-types                       clean (ty)
make lint-imports                     10 contracts kept, 0 broken
make check-source-language            clean
make check-large-files                clean
make build-ui                         built, both boot snapshots match
node ui-web/scripts/check-css.mjs     clean
node ui-web/scripts/check-class-namespace.mjs  clean
```

The three Python failures are not this branch's and were each confirmed
on the
base alone: one asks a host without LibreOffice to refuse a render and
this host
has LibreOffice installed, and two simulation cases fail on an
unmodified
checkout of the base commit.

The RPC contract was regenerated for both consumers and both match it
(`npm run gen --prefix ui-web`, `npm run gen:rpc --prefix ui-tui`, then
`gen:check`). The page's boot snapshots and region goldens were
refreshed where
the composer bar gained a button, and each records that button and
nothing else.

- [x] Relevant tests pass locally
- [x] Relevant lint / type checks pass locally
- [x] User-facing docs or screenshots are updated when needed

## Risk

A PDF's tables, and a Word or PowerPoint file's, are now indexed as HTML
rather
than as header-prefixed lines. Chunk text changes shape for those
documents, so
a base indexed before this and searched after it holds both forms until
its
documents are reindexed; retrieval keeps working either way, because
both forms
are text. The same applies to how much prose a table or a figure carries
with
it: the setting now reaches as far as it says, where before it stopped
at the
nearest paragraph, so a chunk cut before this keeps the narrower context
until
its document is rebuilt. Rolling back is reverting the branch and
reindexing.

The deepdoc models are optional. An install that never ran `make
fetch-resources` reads PDFs geometrically, exactly as before, and says
so when
the models are asked for by name. Nothing here reaches the network at
request
time.

Chunk crops and rendered pages are written at index time and on first
view, both
under the state directory rather than in the repository, and both are
deleted
with their document. A page is immutable for the life of its document
id, so
the route serving them caches; the document route does not. All three
routes
take ids and no path, and the two image routes answer under the
strictest
sandbox in the tree.

What an agent may search is the reader's choice and nothing else. The
tool takes
no base argument, so a model cannot name one, and it is absent from the
schema
entirely for a conversation that named none -- a lane nobody pointed at
a base
is unchanged by this branch. The bases a session carries are stored
beside its
working directory and travel no further.

- [x] Security impact considered
- [x] Backward compatibility considered
- [x] Rollback path is clear for risky changes

## Related Issues

N/A

---------

Co-authored-by: Claude (claude-opus-5) <noreply@anthropic.com>
2026-10-12 19:05:16 +08:00

356 lines
15 KiB
Python

"""Attaching each text box to the region it sits in.
The layout model answers with regions; this is what happens next -- every text
box on the page is given the type of the region covering it, page furniture is
dropped, and a figure carrying no text still gets a box so the crop survives.
The model's answer is supplied here rather than inferred, which is the only
part that needs the weights. Everything below it is arithmetic and ordering,
and it is where the damage is quiet: a header read as a title puts running
furniture into the index on every page of the document.
`tests/test_knowledge_deepdoc_layout.py` covers the same class against the real
weights and skips without them, which is every CI run.
"""
from __future__ import annotations
import pytest
from PIL import Image
from raven.knowledge.parser.deepdoc._layout_recognizer import LayoutRecognizer
def region(kind: str, x0: float, top: float, x1: float, bottom: float, score: float = 0.9) -> dict:
"""One region, as the model answers it: a corner box and a class name."""
return {"type": kind, "score": score, "bbox": [x0, top, x1, bottom]}
def text(content: str, x0: float, top: float, x1: float, bottom: float) -> dict:
return {"text": content, "x0": x0, "top": top, "x1": x1, "bottom": bottom}
def pages(count: int = 1, size: tuple[int, int] = (300, 900)) -> list[Image.Image]:
return [Image.new("RGB", size) for _ in range(count)]
@pytest.fixture
def recogniser(monkeypatch: pytest.MonkeyPatch):
"""A layout recogniser whose model answer is whatever a case supplies.
`__init__` opens an onnx session; nothing below reaches it.
"""
def build(answer: list[list[dict]]) -> LayoutRecognizer:
from raven.knowledge.parser.deepdoc import _recognizer
monkeypatch.setattr(_recognizer.Recognizer, "__call__", lambda self, images, thr=0.2, batch_size=16: answer)
instance = object.__new__(LayoutRecognizer)
instance.garbage_layouts = ["abandon"] # type: ignore[attr-defined]
instance.client = None # type: ignore[attr-defined]
instance.label_list = LayoutRecognizer.labels # type: ignore[attr-defined]
return instance
return build
class TestLabels:
def test_the_label_order_is_the_published_models_own(self) -> None:
"""Not RAGFlow's list, which starts with a background class: read
against that one a title comes back as `_background_` and a page
footer as a title."""
assert LayoutRecognizer.labels[:3] == ["Title", "Text", "Abandon"]
def test_page_furniture_is_one_class(self) -> None:
assert "Abandon" in LayoutRecognizer.labels
assert "Header" not in LayoutRecognizer.labels
class TestFilterGarbageLayouts:
@staticmethod
def _recogniser() -> LayoutRecognizer:
instance = object.__new__(LayoutRecognizer)
instance.garbage_layouts = ["abandon"] # type: ignore[attr-defined]
return instance
def test_a_low_scoring_furniture_region_is_dropped(self) -> None:
boxes = [{"type": "abandon", "score": 0.1}]
assert self._recogniser()._filter_garbage_layouts(boxes) == []
def test_a_confident_furniture_region_is_kept(self) -> None:
boxes = [{"type": "abandon", "score": 0.9}]
assert self._recogniser()._filter_garbage_layouts(boxes) == boxes
def test_a_low_scoring_content_region_is_kept(self) -> None:
"""The gate is about furniture, not confidence in general: a faint
table is still a table and dropping it loses the content."""
boxes = [{"type": "table", "score": 0.1}]
assert self._recogniser()._filter_garbage_layouts(boxes) == boxes
def test_the_threshold_can_be_moved(self) -> None:
boxes = [{"type": "abandon", "score": 0.5}]
assert self._recogniser()._filter_garbage_layouts(boxes, score_thr=0.6) == []
class TestForward:
def test_it_applies_the_same_furniture_gate_as_the_main_path(self, recogniser) -> None:
"""Without this the HTTP endpoint answers raw low-confidence furniture
while the parser beside it does not."""
answer = [[region("abandon", 0, 0, 10, 10, score=0.1), region("text", 0, 20, 10, 30)]]
out = recogniser(answer).forward(pages())
assert [b["type"] for b in out[0]] == ["text"]
class TestAssigningBoxesToRegions:
def test_a_box_takes_the_type_of_the_region_over_it(self, recogniser) -> None:
answer = [[region("table", 0, 0, 300, 300)]]
boxes = [text("a value", 10, 10, 50, 40)]
out, _layout = recogniser(answer)(pages(), [boxes], scale_factor=1)
assert out[0]["layout_type"] == "table"
assert out[0]["layoutno"].startswith("table-")
def test_a_box_under_no_region_is_left_untyped(self, recogniser) -> None:
answer = [[region("table", 0, 0, 10, 10)]]
boxes = [text("far away", 200, 500, 280, 540)]
out, _layout = recogniser(answer)(pages(), [boxes], scale_factor=1)
assert out[0]["layout_type"] == ""
def test_an_equation_is_filed_as_a_figure(self, recogniser) -> None:
"""One name downstream: the crop and the caption are handled the same
way for both."""
answer = [[region("equation", 0, 0, 300, 300)]]
boxes = [text("e = mc^2", 10, 10, 50, 40)]
out, _layout = recogniser(answer)(pages(), [boxes], scale_factor=1)
assert out[0]["layout_type"] == "figure"
def test_the_page_layout_comes_back_beside_the_boxes(self, recogniser) -> None:
answer = [[region("text", 0, 0, 300, 300)]]
_out, layout = recogniser(answer)(pages(), [[text("a", 10, 10, 50, 40)]], scale_factor=1)
assert len(layout) == 1
assert layout[0][0]["type"] == "text"
def test_the_scale_factor_brings_the_regions_back_to_page_coordinates(self, recogniser) -> None:
"""The model reads a page rendered at three times its size."""
answer = [[region("text", 0, 0, 300, 300)]]
_out, layout = recogniser(answer)(pages(), [[text("a", 10, 10, 50, 40)]], scale_factor=3)
assert layout[0][0]["x1"] == pytest.approx(100.0)
def test_a_cid_artefact_is_dropped(self, recogniser) -> None:
"""`(cid:12)` is what a PDF with a broken font map yields instead of a
glyph. It is not text and must not reach the index."""
answer = [[region("text", 0, 0, 300, 300)]]
boxes = [text("(cid:12)", 10, 10, 50, 40), text("real words", 10, 50, 50, 80)]
out, _layout = recogniser(answer)(pages(), [boxes], scale_factor=1)
assert [b["text"] for b in out] == ["real words"]
class TestDroppingFurniture:
"""Page furniture, and the branch here that no longer drops it.
The published model answers one `Abandon` class for headers, footers and
page numbers, and `garbage_layouts` names it. But the loop that walks the
regions still asks for the older model's three names -- footer, header,
reference -- so an `abandon` region is never matched to a text box and the
drop beside it cannot be reached. `"abandon"` appears exactly once in this
module: in that list.
Not a leak into the index, though: `_pipeline._read` skips an `abandon`
region when it builds elements, so the furniture is excluded a step later.
What is pinned here is that this path no longer does it, which is worth
knowing before someone relies on it or deletes the list as unused.
"""
def test_a_repeated_footer_reaches_the_output_today(self, recogniser) -> None:
answer = [[region("abandon", 0, 800, 300, 880)] for _ in range(2)]
per_page = [[text("Page footer", 10, 810, 100, 850)] for _ in range(2)]
out, _layout = recogniser(answer)(pages(2), per_page, scale_factor=1)
assert [b["text"] for b in out] == ["Page footer", "Page footer"]
assert all(b["layout_type"] == "" for b in out), "never matched to the abandon region"
def test_content_is_never_dropped(self, recogniser) -> None:
answer = [
[region("abandon", 0, 800, 300, 880), region("text", 0, 0, 300, 100)],
[region("abandon", 0, 800, 300, 880), region("text", 0, 0, 300, 100)],
]
per_page = [[text("Page footer", 10, 810, 100, 850), text(f"body {n}", 10, 10, 100, 50)] for n in range(2)]
out, _layout = recogniser(answer)(pages(2), per_page, scale_factor=1)
assert {"body 0", "body 1"} <= {b["text"] for b in out}
@pytest.mark.xfail(
reason="findLayout asks for footer/header/reference, so the published model's 'abandon' is never matched",
strict=True,
)
def test_this_path_no_longer_drops_a_repeated_footer(self, recogniser) -> None:
"""What `garbage_layouts` is for, pinned as dead.
`_pipeline._read` drops the region a step later, so nothing is lost
today. This says the drop here is unreachable, so that a change which
makes it reachable again is a visible one.
"""
answer = [[region("abandon", 0, 800, 300, 880)] for _ in range(2)]
per_page = [[text("Page footer", 10, 810, 100, 850)] for _ in range(2)]
out, _layout = recogniser(answer)(pages(2), per_page, scale_factor=1)
assert out == []
def test_keeping_the_furniture_is_possible(self, recogniser) -> None:
"""`drop=False` is what the page viewer asks for: it draws the whole
page, furniture included."""
answer = [[region("abandon", 0, 800, 300, 880)] for _ in range(2)]
per_page = [[text("Page footer", 10, 810, 100, 850)] for _ in range(2)]
out, _layout = recogniser(answer)(pages(2), per_page, scale_factor=1, drop=False)
assert len(out) == 2
class TestTextlessRegions:
def test_a_figure_with_no_text_still_gets_a_box(self, recogniser) -> None:
"""Otherwise the crop is never taken and the figure is lost: there is
no text box to hang it on."""
answer = [[region("figure", 0, 0, 300, 300)]]
out, _layout = recogniser(answer)(pages(), [[]], scale_factor=1)
assert len(out) == 1
assert out[0]["layout_type"] == "figure"
assert out[0]["text"] == ""
assert out[0]["layoutno"] == "figure-0"
def test_a_figure_that_did_have_text_gets_no_second_box(self, recogniser) -> None:
answer = [[region("figure", 0, 0, 300, 300)]]
boxes = [text("a caption inside", 10, 10, 50, 40)]
out, _layout = recogniser(answer)(pages(), [boxes], scale_factor=1)
assert len(out) == 1
def test_figures_and_equations_are_numbered_in_their_own_namespaces(self, recogniser) -> None:
"""A combined index under one prefix would collide with the tags
`findLayout` assigns and merge two unrelated regions."""
answer = [[region("figure", 0, 0, 100, 100), region("equation", 0, 200, 100, 300)]]
out, _layout = recogniser(answer)(pages(), [[]], scale_factor=1)
assert sorted(b["layoutno"] for b in out) == ["equation-0", "figure-0"]
#: A second detection, scored under the 0.08 this branch forces, so it is
#: filtered out and changes no assertion. Present because `np.squeeze` flattens
#: a one-detection block into a single row, which the branch cannot index -- a
#: shape the graph never answers, since it emits one per anchor.
_IGNORED_DETECTION = [900.0, 900.0, 910.0, 910.0, 0.0, 0.0]
class TestYolov10Variant:
"""The letterboxing variant.
It fits the page inside the graph's square input rather than stretching it
to fill it, so the aspect ratio survives, and records the padding it added
so the boxes can be put back on the page afterwards. Getting the padding
subtraction wrong moves every region by the width of the grey border.
"""
@staticmethod
def _variant(input_shape: tuple[int, int] = (64, 64)):
from raven.knowledge.parser.deepdoc._layout_recognizer import LayoutRecognizer4YOLOv10
instance = object.__new__(LayoutRecognizer4YOLOv10)
instance.input_names = ["images"]
instance.input_shape = input_shape
instance.label_list = LayoutRecognizer4YOLOv10.labels
instance.center = True
return instance
def test_a_page_is_fitted_inside_the_graphs_input(self) -> None:
import numpy as np
out = self._variant().preprocess([np.zeros((100, 200, 3), dtype=np.uint8)])
assert out[0]["images"].shape == (1, 3, 64, 64)
def test_the_padding_it_added_is_recorded(self) -> None:
"""A 2:1 page in a square input is padded top and bottom, and that
offset has to come back off the boxes."""
import numpy as np
out = self._variant().preprocess([np.zeros((100, 200, 3), dtype=np.uint8)])
_wx, _wy, dw, dh = out[0]["scale_factor"]
assert dw == pytest.approx(0.0)
assert dh > 0, "padded on the short axis"
def test_a_square_page_needs_no_padding(self) -> None:
import numpy as np
out = self._variant().preprocess([np.zeros((100, 100, 3), dtype=np.uint8)])
_wx, _wy, dw, dh = out[0]["scale_factor"]
assert (dw, dh) == (pytest.approx(0.0), pytest.approx(0.0))
def test_pixels_arrive_scaled_to_the_unit_range(self) -> None:
import numpy as np
out = self._variant().preprocess([np.full((64, 64, 3), 255, dtype=np.uint8)])
assert out[0]["images"].max() == pytest.approx(1.0)
def test_a_detection_comes_back_with_the_padding_taken_off(self) -> None:
import numpy as np
answer = np.array([[[10.0, 20.0, 30.0, 40.0, 0.9, 0.0], _IGNORED_DETECTION]], dtype=np.float32)
found = self._variant().postprocess(answer, {"scale_factor": [2.0, 2.0, 5.0, 10.0]}, thr=0.5)
assert len(found) == 1
assert found[0]["bbox"] == [10.0, 20.0, 50.0, 60.0], "padding off, then scaled"
def test_a_faint_detection_is_dropped(self) -> None:
import numpy as np
answer = np.array([[[10.0, 20.0, 30.0, 40.0, 0.01, 0.0], _IGNORED_DETECTION]], dtype=np.float32)
assert self._variant().postprocess(answer, {"scale_factor": [1.0, 1.0, 0.0, 0.0]}, thr=0.5) == []
def test_two_readings_of_one_region_are_thinned(self) -> None:
import numpy as np
answer = np.array([[[10.0, 20.0, 30.0, 40.0, 0.9, 0.0], [11.0, 21.0, 31.0, 41.0, 0.6, 0.0]]], dtype=np.float32)
found = self._variant().postprocess(answer, {"scale_factor": [1.0, 1.0, 0.0, 0.0]}, thr=0.5)
assert len(found) == 1
assert found[0]["score"] == pytest.approx(0.9)
def test_two_kinds_that_overlap_both_survive(self) -> None:
import numpy as np
answer = np.array([[[10.0, 20.0, 30.0, 40.0, 0.9, 0.0], [10.0, 20.0, 30.0, 40.0, 0.9, 3.0]]], dtype=np.float32)
found = self._variant().postprocess(answer, {"scale_factor": [1.0, 1.0, 0.0, 0.0]}, thr=0.5)
assert {b["type"] for b in found} == {"title", "figure"}