Raven-1/tests/test_knowledge_deepdoc_operators.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

281 lines
10 KiB
Python

"""The transforms a rendered page passes through before the model sees it.
Resize, normalise, reorder the axes, pad to a stride. Pure arithmetic over
arrays, so none of it needs the weights -- and all of it is the kind of ported
code where a mistake is silent: the wrong mean still produces a tensor of the
right shape, and the model answers it a little worse.
The shapes are deliberately small and the numbers round, so each case says what
the transform does rather than merely that it ran.
"""
from __future__ import annotations
import numpy as np
import pytest
from PIL import Image
from raven.knowledge.parser.deepdoc._operators import (
DetResizeForTest,
KeepKeys,
LinearResize,
NormalizeImage,
PadStride,
Permute,
StandardizeImage,
ToCHWImage,
_as_array,
nms,
preprocess,
)
def page(height: int = 20, width: int = 30, value: int = 255) -> np.ndarray:
return np.full((height, width, 3), value, dtype=np.uint8)
class TestAsArray:
def test_a_pil_image_becomes_an_array(self) -> None:
out = _as_array(Image.new("RGB", (4, 3)))
assert isinstance(out, np.ndarray)
assert out.shape == (3, 4, 3), "PIL is width-first, arrays are height-first"
def test_an_array_is_handed_back_unchanged(self) -> None:
given = page()
assert _as_array(given) is given
class TestStandardizeImage:
def test_it_scales_to_the_unit_range_then_subtracts_the_mean(self) -> None:
out, _ = StandardizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])(page(value=255).astype(np.float32), {})
assert out == pytest.approx(np.ones((20, 30, 3)))
def test_without_scaling_the_raw_values_are_standardised(self) -> None:
out, _ = StandardizeImage(mean=[0.0] * 3, std=[1.0] * 3, is_scale=False)(page(value=7).astype(np.float32), {})
assert out == pytest.approx(np.full((20, 30, 3), 7.0))
def test_norm_type_none_only_scales(self) -> None:
"""The mean and std are still constructed; what changes is whether they
are applied. A page that skipped the scaling too would arrive at 255."""
out, _ = StandardizeImage(mean=[9.0] * 3, std=[9.0] * 3, norm_type="none")(
page(value=255).astype(np.float32), {}
)
assert out == pytest.approx(np.ones((20, 30, 3)))
def test_the_info_dictionary_rides_through_untouched(self) -> None:
info = {"scale_factor": np.array([1.0, 1.0])}
_, out = StandardizeImage(mean=[0.0] * 3, std=[1.0] * 3)(page().astype(np.float32), info)
assert out is info
class TestNormalizeImage:
"""Unlike `StandardizeImage` beside it, this one defaults to channel-first:
the recognition graph takes its input that way, so the statistics are
shaped (3, 1, 1) and an HWC page will not broadcast against them."""
@staticmethod
def _chw(value: int, height: int = 4, width: int = 6) -> np.ndarray:
return np.full((3, height, width), value, dtype=np.uint8)
def test_it_scales_and_standardises(self) -> None:
op = NormalizeImage(scale=1.0 / 255.0, mean=[0.5] * 3, std=[0.5] * 3)
out = op({"image": self._chw(255)})
assert out["image"] == pytest.approx(np.ones((3, 4, 6)))
def test_it_defaults_to_the_imagenet_statistics(self) -> None:
"""The published weights were trained against these; a build that
quietly used zeros would read every page slightly washed out."""
out = NormalizeImage()({"image": self._chw(0)})
assert out["image"].min() < 0, "subtracting a positive mean from black"
def test_the_last_axis_holds_the_channels_when_asked(self) -> None:
op = NormalizeImage(order="hwc", scale=1.0 / 255.0, mean=[0.5] * 3, std=[0.5] * 3)
out = op({"image": page(value=255)})
assert out["image"] == pytest.approx(np.ones((20, 30, 3)))
def test_a_scale_written_as_a_fraction_is_read_as_one(self) -> None:
"""The operator tables are JSON, where this arrives as the string
"1/255" rather than a number."""
op = NormalizeImage(scale="1/255", mean=[0.0] * 3, std=[1.0] * 3)
out = op({"image": self._chw(255)})
assert out["image"] == pytest.approx(np.ones((3, 4, 6)))
def test_a_scale_written_as_a_decimal_string_is_read_as_one(self) -> None:
op = NormalizeImage(scale="0.5", mean=[0.0] * 3, std=[1.0] * 3)
out = op({"image": self._chw(2)})
assert out["image"] == pytest.approx(np.ones((3, 4, 6)))
class TestToCHWImage:
def test_the_channel_axis_moves_to_the_front(self) -> None:
out = ToCHWImage()({"image": page(height=4, width=6)})
assert out["image"].shape == (3, 4, 6)
def test_the_pixels_are_the_same_ones(self) -> None:
given = np.arange(2 * 3 * 3, dtype=np.uint8).reshape(2, 3, 3)
out = ToCHWImage()({"image": given.copy()})
assert out["image"][0, 1, 2] == given[1, 2, 0]
class TestKeepKeys:
def test_only_the_named_keys_come_back_and_in_order(self) -> None:
out = KeepKeys(keep_keys=["b", "a"])({"a": 1, "b": 2, "c": 3})
assert out == [2, 1]
def test_an_empty_selection_answers_with_nothing(self) -> None:
assert KeepKeys(keep_keys=[])({"a": 1}) == []
class TestLinearResize:
def test_a_single_int_target_is_taken_as_a_square(self) -> None:
assert LinearResize(64).target_size == [64, 64]
def test_without_keeping_the_ratio_the_page_lands_on_the_target(self) -> None:
out, info = LinearResize([10, 20], keep_ratio=False)(page(height=40, width=40), {})
assert out.shape[:2] == (10, 20)
assert info["im_shape"] == pytest.approx(np.array([10.0, 20.0]))
def test_keeping_the_ratio_scales_both_axes_alike(self) -> None:
out, info = LinearResize([10, 10], keep_ratio=True)(page(height=40, width=20), {})
assert info["scale_factor"][0] == pytest.approx(info["scale_factor"][1])
assert max(out.shape[:2]) <= 10
def test_the_scale_factor_records_what_the_resize_did(self) -> None:
"""The boxes come back in the resized frame; this is what puts them
back on the page."""
_, info = LinearResize([10, 20], keep_ratio=False)(page(height=40, width=40), {})
assert info["scale_factor"] == pytest.approx(np.array([0.25, 0.5]))
def test_a_target_of_zero_is_refused(self) -> None:
with pytest.raises(AssertionError):
LinearResize([0, 10], keep_ratio=False)(page(), {})
class TestPermute:
def test_the_channel_axis_moves_to_the_front(self) -> None:
out, _ = Permute()(page(height=4, width=6), {})
assert out.shape == (3, 4, 6)
class TestPadStride:
def test_a_page_is_padded_up_to_a_multiple_of_the_stride(self) -> None:
"""The graph's convolutions need the input a multiple of 32; the pad is
on the right and bottom so the boxes keep their coordinates."""
given = np.ones((3, 10, 20), dtype=np.float32)
out, _ = PadStride(stride=32)(given, {})
assert out.shape == (3, 32, 32)
def test_the_original_pixels_keep_their_place(self) -> None:
given = np.ones((3, 10, 20), dtype=np.float32)
out, _ = PadStride(stride=32)(given, {})
assert out[:, :10, :20] == pytest.approx(1.0)
assert out[:, 10:, :] == pytest.approx(0.0), "padded with zeros, after the image"
def test_a_stride_of_zero_leaves_the_page_alone(self) -> None:
given = np.ones((3, 10, 20), dtype=np.float32)
out, _ = PadStride(stride=0)(given, {})
assert out.shape == given.shape
def test_a_page_already_on_the_stride_is_not_grown(self) -> None:
given = np.ones((3, 32, 64), dtype=np.float32)
out, _ = PadStride(stride=32)(given, {})
assert out.shape == (3, 32, 64)
class TestDetResizeForTest:
def test_a_fixed_target_shape_is_honoured(self) -> None:
out = DetResizeForTest(image_shape=[32, 64])({"image": page(height=10, width=20)})
assert out["image"].shape[:2] == (32, 64)
def test_the_ratio_comes_back_for_putting_boxes_home(self) -> None:
out = DetResizeForTest(image_shape=[32, 64])({"image": page(height=10, width=20)})
assert "shape" in out
assert len(out["shape"]) == 4, "height, width, and the two ratios"
def test_a_limit_side_resize_keeps_the_page_within_the_limit(self) -> None:
out = DetResizeForTest(limit_side_len=64, limit_type="max")({"image": page(height=200, width=100)})
assert max(out["image"].shape[:2]) <= 64 + 32, "within the limit, rounded up to the stride"
def test_a_resize_long_keeps_the_longest_side_near_its_target(self) -> None:
out = DetResizeForTest(resize_long=96)({"image": page(height=200, width=100)})
assert max(out["image"].shape[:2]) == pytest.approx(96, abs=32)
class TestPreprocess:
def test_the_operators_run_in_order_and_the_info_accumulates(self) -> None:
ops = [
LinearResize([16, 16], keep_ratio=False),
StandardizeImage(mean=[0.0] * 3, std=[1.0] * 3),
Permute(),
PadStride(stride=32),
]
out, info = preprocess(page(height=40, width=40), ops)
assert out.shape == (3, 32, 32)
assert info["scale_factor"] == pytest.approx(np.array([0.4, 0.4]))
def test_the_original_shape_is_recorded_before_anything_runs(self) -> None:
_, info = preprocess(page(height=40, width=60), [])
assert info["im_shape"] == pytest.approx(np.array([40.0, 60.0]))
class TestNms:
def test_two_detections_of_one_thing_keep_the_better(self) -> None:
boxes = np.array([[0.0, 0.0, 10.0, 10.0], [0.5, 0.5, 10.5, 10.5]])
scores = np.array([0.9, 0.6])
assert nms(boxes, scores, iou_thresh=0.3) == [0]
def test_detections_that_do_not_overlap_both_survive(self) -> None:
boxes = np.array([[0.0, 0.0, 10.0, 10.0], [100.0, 100.0, 110.0, 110.0]])
scores = np.array([0.9, 0.6])
assert sorted(nms(boxes, scores, iou_thresh=0.3)) == [0, 1]
def test_the_order_is_by_score_rather_than_by_position(self) -> None:
boxes = np.array([[0.0, 0.0, 10.0, 10.0], [100.0, 100.0, 110.0, 110.0]])
scores = np.array([0.1, 0.9])
assert nms(boxes, scores, iou_thresh=0.3)[0] == 1
def test_a_single_detection_is_kept(self) -> None:
assert nms(np.array([[0.0, 0.0, 10.0, 10.0]]), np.array([0.9]), iou_thresh=0.3) == [0]