Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(grounding): preserve explicitly sourced document amounts and units

Haozhe Wud37b2e7f1183
diff --git a/agent/src/agent/grounding/figures.py b/agent/src/agent/grounding/figures.py
index 15536b3e2..1bb259423 100644
--- a/agent/src/agent/grounding/figures.py
+++ b/agent/src/agent/grounding/figures.py
@@ -97,7 +97,7 @@ _CURRENCY_CODES = frozenset(
_PERCENT_CHARS = "%%"
# Glued percentage-point and basis-point marks, and the percent each one is.
-_POINT_MARKS = (("bps", 0.01), ("pp", 1.0), ("bp", 0.01))
+_POINT_MARKS = (("个百分点", 1.0), ("bps", 0.01), ("pp", 1.0), ("bp", 0.01))
# Magnitude marks glued to a figure ("24.6M", "2.4万"): a symbol set like the
# currency marks. They scale the comparison with evidence and never decide shape.
diff --git a/agent/src/agent/grounding/policies.py b/agent/src/agent/grounding/policies.py
index 6d6129490..a2ad289cb 100644
--- a/agent/src/agent/grounding/policies.py
+++ b/agent/src/agent/grounding/policies.py
@@ -170,7 +170,7 @@ def _explicit_sign(text: str) -> int:
def _is_percent_field(name: str) -> bool:
"""Whether a field name says its value is already in percentage points."""
lowered = name.lower()
- return "percent" in lowered or "pct" in lowered or lowered.endswith("%")
+ return name == "document.ratio" or "percent" in lowered or "pct" in lowered or lowered.endswith("%")
def _is_plain_count(figure: Figure) -> bool:
@@ -881,9 +881,19 @@ class _PolicyMixin:
record
for record in records
if not _is_price_kind(record) and not _is_metadata_count_leaf(record.field)
+ and (record.source != "document" or record.field == "document.ratio")
]
elif figure is not None and figure.currency:
- records = [record for record in records if _is_price_kind(record)]
+ # Documents often put the unit in a table header. An explicitly
+ # referenced numeric cell can be restated with that unit; it is
+ # still matched to this exact source and written precision. Do not
+ # promote it to a price globally or admit ratios as amounts. Known
+ # currency identity is checked by _check_observed below.
+ records = [
+ record for record in records
+ if _is_price_kind(record)
+ or (record.source == "document" and record.field == "document.figure")
+ ]
metrics = []
return records, metrics
@@ -1977,10 +1987,22 @@ class _PolicyMixin:
anchors.extend(
float(record.value)
for record in scoped[0]
- if (not money or _is_price_kind(record))
+ if (
+ not money or _is_price_kind(record)
+ or (record.source == "document" and record.field == "document.figure")
+ )
and (currency_code is None or record.currency == currency_code)
and not (unattributed and record.symbol)
)
+ # A literal percent in a formula is parsed as its fraction, while
+ # document.ratio records retain the percent as printed. This
+ # conversion belongs only to the explicitly named document scope.
+ anchors.extend(
+ float(record.value) / 100.0
+ for record in scoped[0]
+ if record.source == "document" and record.field == "document.ratio"
+ and not money and currency_code is None
+ )
if not money and not unattributed:
anchors.extend(scoped[1])
if not anchors:
@@ -2118,13 +2140,14 @@ class _PolicyMixin:
``(low − high) / high`` or as the drop, unless the prose wrote a sign.
"""
# The normalized reading, so "0,666" is three decimals and "−5,13%" is signed.
- half_unit = _written_half_unit(figure.digits or figure.text)
+ scale = float(figure.scale or 1.0)
+ half_unit = _written_half_unit(figure.digits or figure.text) * scale
if figure.percent:
targets = {result} if points else {result * 100.0}
else:
targets = {result, result * 100.0}
sign = _explicit_sign(figure.sign or figure.text)
- value = abs(figure.value)
+ value = abs(figure.value) * scale
return any(
abs(value - abs(target)) <= max(abs(target) * _TOLERANCE, half_unit, 1e-9)
for target in targets
diff --git a/agent/src/tools/doc_reader_tool.py b/agent/src/tools/doc_reader_tool.py
index 250f64367..95efa0d89 100644
--- a/agent/src/tools/doc_reader_tool.py
+++ b/agent/src/tools/doc_reader_tool.py
@@ -475,7 +475,10 @@ class DocReaderTool(BaseTool):
parameters = {
"type": "object",
"properties": {
- "file_path": {"type": "string", "description": "Absolute path to the file."},
+ "file_path": {
+ "type": "string",
+ "description": "Absolute file path or the uploads/<name> handle returned by upload.",
+ },
"pages": {
"type": "string",
"description": "PDF only: page range (e.g. '1-10', '5', '1,3,5-8'). Ignored for other formats.",
@@ -492,8 +495,11 @@ class DocReaderTool(BaseTool):
repeatable = True
def execute(self, **kwargs: Any) -> str:
+ file_path = kwargs.get("file_path")
+ if not isinstance(file_path, str) or not file_path.strip():
+ return _err("file_path is required: pass the uploaded uploads/<name> handle or an absolute file path.")
return read_document(
- kwargs["file_path"],
+ file_path,
kwargs.get("pages", ""),
min_text_per_page=kwargs.get("min_text_per_page", _MIN_TEXT_PER_PAGE),
)
diff --git a/agent/tests/test_doc_reader_security.py b/agent/tests/test_doc_reader_security.py
index 56338f34c..f145d4da4 100644
--- a/agent/tests/test_doc_reader_security.py
+++ b/agent/tests/test_doc_reader_security.py
@@ -6,7 +6,7 @@ import json
import pytest
-from src.tools.doc_reader_tool import read_document
+from src.tools.doc_reader_tool import DocReaderTool, read_document
from src.tools.path_utils import safe_document_path, safe_user_path
@@ -46,3 +46,22 @@ def test_read_document_allows_configured_import_root(
assert result["status"] == "ok"
assert result["text"] == "VT_DOC_OK"
assert safe_document_path(str(doc)) == doc.resolve()
+
+
[email protected]("arguments", [{}, {"file_path": ""}, {"file_path": None}])
+def test_missing_document_path_names_the_required_argument(arguments) -> None:
+ result = _read_json(DocReaderTool().execute(**arguments))
+ assert result["status"] == "error"
+ assert "file_path is required" in result["error"]
+
+
+def test_uploaded_handle_uses_upload_root_even_when_a_run_dir_is_supplied(tmp_path, monkeypatch) -> None:
+ from src.config.paths import get_uploads_dir
+
+ monkeypatch.setenv("VIBE_TRADING_HOME", str(tmp_path))
+ doc = get_uploads_dir() / "report.txt"
+ doc.parent.mkdir(parents=True, exist_ok=True)
+ doc.write_text("UPLOADED_REPORT", encoding="utf-8")
+ result = _read_json(DocReaderTool().execute(file_path="uploads/report.txt", run_dir=str(tmp_path / "runs/r1")))
+ assert result["status"] == "ok"
+ assert result["text"] == "UPLOADED_REPORT"
diff --git a/agent/tests/test_grounding_document_figures.py b/agent/tests/test_grounding_document_figures.py
index f21870ae9..f311d89db 100644
--- a/agent/tests/test_grounding_document_figures.py
+++ b/agent/tests/test_grounding_document_figures.py
@@ -22,6 +22,8 @@ from __future__ import annotations
import json
from pathlib import Path
+import pytest
+
from src.agent.grounding import GroundingLedger
REPORT = """第一节 经营回顾
@@ -165,3 +167,69 @@ def test_an_error_envelope_records_nothing(tmp_path: Path) -> None:
)
assert not result.valid
+
+
[email protected]("ref", ["doc-1", "read_document", "ar2025.pdf", "doc-1::document.figure"])
+ ("value", "written"),
+ [("763,154,445.01", "763,154,445.01 元"), ("191,586,814.48", "1.92亿元")],
+)
+def test_document_table_amount_can_be_restated_with_a_unit(
+ tmp_path: Path, ref: str, value: str, written: str,
+) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", f"单位:元\n| 项目 | 金额 |\n| 营业收入 | {value} |")
+ result = ledger.validate_final_answer(
+ _declared(f"营业收入为 {written}。", f"{written} | observed | 营业收入 | {ref}")
+ )
+ assert result.valid, result.issues
+
+
[email protected]("written", ["863,154,445.01 元", "763,154,445.01 USD", "763,154,445.01%"])
+def test_document_unit_restatement_does_not_invent_value_currency_or_ratio(
+ tmp_path: Path, written: str,
+) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", "| 营业收入 | 763,154,445.01 |")
+ result = ledger.validate_final_answer(
+ _declared(f"营业收入为 {written}。", f"{written} | observed | 营业收入 | doc-1")
+ )
+ assert not result.valid
+
+
+def test_document_unit_restatement_keeps_exact_source_binding(tmp_path: Path) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", "| 营业收入 | 763,154,445.01 |")
+ _read_document(ledger, document.with_name("other.pdf"), "doc-2", "| 营业收入 | 863,154,445.01 |")
+ result = ledger.validate_final_answer(
+ _declared("营业收入为 863,154,445.01 元。", "863,154,445.01 元 | observed | 营业收入 | doc-1")
+ )
+ assert not result.valid
+
+
[email protected]("formula", ["39.66% - 38.96%", "39.66 - 38.96"])
+def test_document_percentage_points_can_be_declared_as_written(tmp_path: Path, formula: str) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", "毛利率分别为 39.66% 和 38.96%。")
+ result = ledger.validate_final_answer(
+ _declared("毛利率上升 0.70个百分点。", f"0.70个百分点 | derived | {formula} | doc-1")
+ )
+ assert result.valid, result.issues
+
+
+def test_document_percentage_point_derivation_still_needs_both_operands(tmp_path: Path) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", "毛利率为 39.66%。")
+ result = ledger.validate_final_answer(
+ _declared("毛利率上升 0.70个百分点。", "0.70个百分点 | derived | 39.66% - 38.96% | doc-1")
+ )
+ assert not result.valid
+
+
+def test_money_can_be_derived_from_explicitly_referenced_document_cells(tmp_path: Path) -> None:
+ ledger, document = _ledger_with_upload(tmp_path)
+ _read_document(ledger, document, "doc-1", "| 本期 | 上期 |\n| 763,154,445.01 | 663,154,445.01 |")
+ result = ledger.validate_final_answer(
+ _declared("营业收入增加 1.00亿元。", "1.00亿元 | derived | 763154445.01 - 663154445.01 | doc-1")
+ )
+ assert result.valid, result.issues