vedana-backoffice 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- vedana_backoffice/Caddyfile +17 -0
- vedana_backoffice/__init__.py +0 -0
- vedana_backoffice/components/__init__.py +0 -0
- vedana_backoffice/components/etl_graph.py +132 -0
- vedana_backoffice/components/ui_chat.py +236 -0
- vedana_backoffice/graph/__init__.py +0 -0
- vedana_backoffice/graph/build.py +169 -0
- vedana_backoffice/pages/__init__.py +0 -0
- vedana_backoffice/pages/chat.py +204 -0
- vedana_backoffice/pages/etl.py +353 -0
- vedana_backoffice/pages/eval.py +1006 -0
- vedana_backoffice/pages/jims_thread_list_page.py +894 -0
- vedana_backoffice/pages/main_dashboard.py +483 -0
- vedana_backoffice/py.typed +0 -0
- vedana_backoffice/start_services.py +39 -0
- vedana_backoffice/state.py +0 -0
- vedana_backoffice/states/__init__.py +0 -0
- vedana_backoffice/states/chat.py +368 -0
- vedana_backoffice/states/common.py +66 -0
- vedana_backoffice/states/etl.py +1590 -0
- vedana_backoffice/states/eval.py +1940 -0
- vedana_backoffice/states/jims.py +508 -0
- vedana_backoffice/states/main_dashboard.py +757 -0
- vedana_backoffice/ui.py +115 -0
- vedana_backoffice/util.py +71 -0
- vedana_backoffice/vedana_backoffice.py +23 -0
- vedana_backoffice-0.1.0.dist-info/METADATA +10 -0
- vedana_backoffice-0.1.0.dist-info/RECORD +30 -0
- vedana_backoffice-0.1.0.dist-info/WHEEL +4 -0
- vedana_backoffice-0.1.0.dist-info/entry_points.txt +2 -0
|
@@ -0,0 +1,1006 @@
|
|
|
1
|
+
import reflex as rx
|
|
2
|
+
|
|
3
|
+
from vedana_backoffice.states.eval import EvalState, RunSummary
|
|
4
|
+
from vedana_backoffice.states.chat import ChatState
|
|
5
|
+
from vedana_backoffice.ui import app_header
|
|
6
|
+
|
|
7
|
+
|
|
8
|
+
def _selection_and_actions() -> rx.Component:
|
|
9
|
+
"""Selection controls and action buttons for the right panel."""
|
|
10
|
+
return rx.vstack(
|
|
11
|
+
rx.hstack(
|
|
12
|
+
rx.input(
|
|
13
|
+
placeholder="Test run name",
|
|
14
|
+
value=EvalState.test_run_name,
|
|
15
|
+
on_change=EvalState.set_test_run_name,
|
|
16
|
+
width="18em",
|
|
17
|
+
),
|
|
18
|
+
rx.button(
|
|
19
|
+
EvalState.selection_label,
|
|
20
|
+
color_scheme="blue",
|
|
21
|
+
on_click=EvalState.run_selected_tests,
|
|
22
|
+
loading=EvalState.is_running,
|
|
23
|
+
disabled=rx.cond(EvalState.can_run, False, True), # type: ignore[arg-type]
|
|
24
|
+
),
|
|
25
|
+
rx.spacer(),
|
|
26
|
+
rx.button(
|
|
27
|
+
"Reset selection",
|
|
28
|
+
variant="ghost",
|
|
29
|
+
color_scheme="gray",
|
|
30
|
+
size="1",
|
|
31
|
+
disabled=rx.cond(EvalState.selected_count > 0, False, True), # type: ignore[arg-type]
|
|
32
|
+
on_click=EvalState.reset_selection,
|
|
33
|
+
),
|
|
34
|
+
spacing="2",
|
|
35
|
+
align="center",
|
|
36
|
+
width="100%",
|
|
37
|
+
),
|
|
38
|
+
rx.button(
|
|
39
|
+
"Reload data",
|
|
40
|
+
variant="soft",
|
|
41
|
+
color_scheme="gray",
|
|
42
|
+
size="1",
|
|
43
|
+
on_click=EvalState.load_eval_data,
|
|
44
|
+
loading=EvalState.loading,
|
|
45
|
+
width="100%",
|
|
46
|
+
),
|
|
47
|
+
spacing="3",
|
|
48
|
+
width="100%",
|
|
49
|
+
)
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
def _questions_card() -> rx.Component:
|
|
53
|
+
def _expandable_text(row: dict[str, rx.Var], key: str, clamp: int = 10) -> rx.Component:
|
|
54
|
+
"""todo reuse _expandable_text in both tables"""
|
|
55
|
+
row_id = row.get("id", "")
|
|
56
|
+
return rx.table.cell(
|
|
57
|
+
rx.box(
|
|
58
|
+
rx.cond(
|
|
59
|
+
row.get("expanded", False),
|
|
60
|
+
rx.text(row.get(key, ""), size="1", white_space="pre-wrap"),
|
|
61
|
+
rx.text(
|
|
62
|
+
row.get(key, ""),
|
|
63
|
+
size="1",
|
|
64
|
+
white_space="pre-wrap",
|
|
65
|
+
style={
|
|
66
|
+
"display": "-webkit-box",
|
|
67
|
+
"WebkitLineClamp": str(clamp),
|
|
68
|
+
"WebkitBoxOrient": "vertical",
|
|
69
|
+
"overflow": "hidden",
|
|
70
|
+
"textOverflow": "ellipsis",
|
|
71
|
+
},
|
|
72
|
+
),
|
|
73
|
+
),
|
|
74
|
+
cursor="pointer",
|
|
75
|
+
on_click=EvalState.toggle_gds_row(row_id=row_id), # type: ignore[arg-type,call-arg,func-returns-value]
|
|
76
|
+
)
|
|
77
|
+
)
|
|
78
|
+
|
|
79
|
+
def _row(row: dict[str, rx.Var]) -> rx.Component:
|
|
80
|
+
row_id_var = row.get("id", "")
|
|
81
|
+
return rx.table.row(
|
|
82
|
+
rx.table.cell(
|
|
83
|
+
rx.checkbox(
|
|
84
|
+
checked=row.get("selected", False),
|
|
85
|
+
on_change=lambda checked: EvalState.toggle_question_selection(question=row_id_var, checked=checked), # type: ignore[arg-type,call-arg,func-returns-value]
|
|
86
|
+
)
|
|
87
|
+
),
|
|
88
|
+
_expandable_text(row, "gds_question"),
|
|
89
|
+
_expandable_text(row, "gds_answer"),
|
|
90
|
+
_expandable_text(row, "question_context"),
|
|
91
|
+
rx.table.cell(
|
|
92
|
+
rx.cond(
|
|
93
|
+
row.get("question_scenario", "") != "",
|
|
94
|
+
rx.badge(
|
|
95
|
+
row.get("question_scenario", ""),
|
|
96
|
+
variant="soft",
|
|
97
|
+
size="1",
|
|
98
|
+
color_scheme=row.get("scenario_color", "gray"),
|
|
99
|
+
),
|
|
100
|
+
rx.box(),
|
|
101
|
+
)
|
|
102
|
+
),
|
|
103
|
+
)
|
|
104
|
+
|
|
105
|
+
return rx.card(
|
|
106
|
+
rx.vstack(
|
|
107
|
+
rx.hstack(
|
|
108
|
+
rx.heading("Golden QA Dataset", size="4"),
|
|
109
|
+
rx.spacer(),
|
|
110
|
+
rx.hstack(
|
|
111
|
+
rx.text("Scenario", size="1", color="gray"),
|
|
112
|
+
rx.select(
|
|
113
|
+
items=EvalState.available_scenarios,
|
|
114
|
+
value=EvalState.selected_scenario,
|
|
115
|
+
on_change=EvalState.set_scenario,
|
|
116
|
+
width="12em",
|
|
117
|
+
),
|
|
118
|
+
rx.tooltip(
|
|
119
|
+
rx.button(
|
|
120
|
+
"↻",
|
|
121
|
+
variant="ghost",
|
|
122
|
+
color_scheme="gray",
|
|
123
|
+
size="1",
|
|
124
|
+
on_click=EvalState.refresh_golden_dataset,
|
|
125
|
+
loading=EvalState.is_running,
|
|
126
|
+
),
|
|
127
|
+
content="Refresh golden dataset from Grist",
|
|
128
|
+
),
|
|
129
|
+
spacing="2",
|
|
130
|
+
align="center",
|
|
131
|
+
),
|
|
132
|
+
align="center",
|
|
133
|
+
width="100%",
|
|
134
|
+
),
|
|
135
|
+
rx.scroll_area(
|
|
136
|
+
rx.table.root(
|
|
137
|
+
rx.table.header(
|
|
138
|
+
rx.table.row(
|
|
139
|
+
rx.table.column_header_cell(
|
|
140
|
+
rx.checkbox(
|
|
141
|
+
checked=EvalState.all_selected,
|
|
142
|
+
on_change=EvalState.toggle_select_all,
|
|
143
|
+
)
|
|
144
|
+
),
|
|
145
|
+
rx.table.column_header_cell("Question"),
|
|
146
|
+
rx.table.column_header_cell("Golden answer"),
|
|
147
|
+
rx.table.column_header_cell("Context"),
|
|
148
|
+
rx.table.column_header_cell("Scenario"),
|
|
149
|
+
)
|
|
150
|
+
),
|
|
151
|
+
rx.table.body(rx.foreach(EvalState.eval_gds_rows_with_selection, _row)),
|
|
152
|
+
variant="surface",
|
|
153
|
+
style={"width": "100%", "tableLayout": "fixed"},
|
|
154
|
+
),
|
|
155
|
+
type="always",
|
|
156
|
+
scrollbars="vertical",
|
|
157
|
+
style={"flex": "1", "width": "100%", "minHeight": "0"},
|
|
158
|
+
),
|
|
159
|
+
spacing="3",
|
|
160
|
+
style={"height": "100%", "display": "flex", "flexDirection": "column"},
|
|
161
|
+
),
|
|
162
|
+
padding="1em",
|
|
163
|
+
width="100%",
|
|
164
|
+
style={"maxHeight": "80vh", "display": "flex", "flexDirection": "column"},
|
|
165
|
+
)
|
|
166
|
+
|
|
167
|
+
|
|
168
|
+
def _judge_card() -> rx.Component:
|
|
169
|
+
return rx.card(
|
|
170
|
+
rx.vstack(
|
|
171
|
+
rx.heading("Judge configuration", size="4"),
|
|
172
|
+
rx.box(
|
|
173
|
+
rx.text("Judge model", weight="medium"),
|
|
174
|
+
rx.text(EvalState.judge_model, size="3"),
|
|
175
|
+
padding_bottom="0.75em",
|
|
176
|
+
),
|
|
177
|
+
rx.button(
|
|
178
|
+
"View Judge Prompt",
|
|
179
|
+
variant="soft",
|
|
180
|
+
size="1",
|
|
181
|
+
on_click=EvalState.open_judge_prompt_dialog,
|
|
182
|
+
disabled=rx.cond(EvalState.judge_prompt_id != "", False, True), # type: ignore[arg-type]
|
|
183
|
+
width="100%",
|
|
184
|
+
margin_top="0.5em",
|
|
185
|
+
),
|
|
186
|
+
spacing="1",
|
|
187
|
+
width="100%",
|
|
188
|
+
),
|
|
189
|
+
padding="1em",
|
|
190
|
+
width="100%",
|
|
191
|
+
)
|
|
192
|
+
|
|
193
|
+
|
|
194
|
+
def _pipeline_card() -> rx.Component:
|
|
195
|
+
return rx.card(
|
|
196
|
+
rx.vstack(
|
|
197
|
+
rx.heading("Pipeline config", size="4"),
|
|
198
|
+
rx.vstack(
|
|
199
|
+
rx.box(
|
|
200
|
+
rx.text("Data model", weight="medium"),
|
|
201
|
+
rx.button(
|
|
202
|
+
"View Data Model",
|
|
203
|
+
variant="soft",
|
|
204
|
+
size="1",
|
|
205
|
+
on_click=EvalState.open_data_model_dialog,
|
|
206
|
+
disabled=rx.cond(EvalState.dm_id != "", False, True), # type: ignore[arg-type]
|
|
207
|
+
width="100%",
|
|
208
|
+
margin_top="0.5em",
|
|
209
|
+
),
|
|
210
|
+
rx.button(
|
|
211
|
+
"Refresh Data Model",
|
|
212
|
+
variant="soft",
|
|
213
|
+
size="1",
|
|
214
|
+
on_click=ChatState.reload_data_model,
|
|
215
|
+
loading=ChatState.is_refreshing_dm,
|
|
216
|
+
width="100%",
|
|
217
|
+
margin_top="0.5em",
|
|
218
|
+
),
|
|
219
|
+
width="100%",
|
|
220
|
+
padding_bottom="0.75em",
|
|
221
|
+
),
|
|
222
|
+
rx.box(
|
|
223
|
+
rx.text("Pipeline model", weight="medium", width="100%"),
|
|
224
|
+
rx.hstack(
|
|
225
|
+
rx.select(
|
|
226
|
+
items=["openai", "openrouter"],
|
|
227
|
+
value=EvalState.provider,
|
|
228
|
+
on_change=EvalState.set_provider,
|
|
229
|
+
width="100%",
|
|
230
|
+
placeholder="Provider",
|
|
231
|
+
),
|
|
232
|
+
rx.cond(
|
|
233
|
+
EvalState.provider == "openrouter",
|
|
234
|
+
rx.input(
|
|
235
|
+
placeholder=rx.cond(
|
|
236
|
+
EvalState.default_openrouter_key_present,
|
|
237
|
+
"(Optional) custom OPENROUTER_API_KEY",
|
|
238
|
+
"(Required) OPENROUTER_API_KEY",
|
|
239
|
+
),
|
|
240
|
+
type="password",
|
|
241
|
+
value=EvalState.custom_openrouter_key,
|
|
242
|
+
on_change=EvalState.set_custom_openrouter_key,
|
|
243
|
+
width="100%",
|
|
244
|
+
required=rx.cond(EvalState.default_openrouter_key_present, False, True),
|
|
245
|
+
),
|
|
246
|
+
),
|
|
247
|
+
rx.select(
|
|
248
|
+
items=EvalState.available_models_view,
|
|
249
|
+
value=EvalState.pipeline_model,
|
|
250
|
+
on_change=EvalState.set_pipeline_model,
|
|
251
|
+
width="100%",
|
|
252
|
+
placeholder="Select model",
|
|
253
|
+
),
|
|
254
|
+
rx.checkbox(
|
|
255
|
+
"Filter Data Model",
|
|
256
|
+
checked=EvalState.enable_dm_filtering,
|
|
257
|
+
on_change=EvalState.set_enable_dm_filtering,
|
|
258
|
+
size="2",
|
|
259
|
+
),
|
|
260
|
+
spacing="2",
|
|
261
|
+
align="center",
|
|
262
|
+
wrap="wrap",
|
|
263
|
+
width="100%",
|
|
264
|
+
),
|
|
265
|
+
padding_bottom="0.75em",
|
|
266
|
+
width="100%",
|
|
267
|
+
),
|
|
268
|
+
rx.box(
|
|
269
|
+
rx.text("Embeddings", weight="medium"),
|
|
270
|
+
rx.text(
|
|
271
|
+
rx.cond(EvalState.embeddings_model != "", EvalState.embeddings_model, "—"),
|
|
272
|
+
size="3",
|
|
273
|
+
),
|
|
274
|
+
rx.text(
|
|
275
|
+
EvalState.embeddings_dim_label,
|
|
276
|
+
size="1",
|
|
277
|
+
color="gray",
|
|
278
|
+
),
|
|
279
|
+
),
|
|
280
|
+
spacing="1",
|
|
281
|
+
width="100%",
|
|
282
|
+
),
|
|
283
|
+
spacing="3",
|
|
284
|
+
width="100%",
|
|
285
|
+
),
|
|
286
|
+
padding="1em",
|
|
287
|
+
width="100%",
|
|
288
|
+
)
|
|
289
|
+
|
|
290
|
+
|
|
291
|
+
def _tests_card() -> rx.Component:
|
|
292
|
+
def _expandable_text(row: dict[str, rx.Var], key: str, clamp: int = 2) -> rx.Component:
|
|
293
|
+
row_id = row.get("row_id", "")
|
|
294
|
+
return rx.table.cell(
|
|
295
|
+
rx.box(
|
|
296
|
+
rx.cond(
|
|
297
|
+
row.get("expanded", False),
|
|
298
|
+
rx.text(
|
|
299
|
+
row.get(key, ""),
|
|
300
|
+
size="1",
|
|
301
|
+
white_space="pre-wrap",
|
|
302
|
+
style={"wordBreak": "break-word"},
|
|
303
|
+
),
|
|
304
|
+
rx.text(
|
|
305
|
+
row.get(key, ""),
|
|
306
|
+
size="1",
|
|
307
|
+
white_space="pre-wrap",
|
|
308
|
+
style={
|
|
309
|
+
"display": "-webkit-box",
|
|
310
|
+
"WebkitLineClamp": str(clamp),
|
|
311
|
+
"WebkitBoxOrient": "vertical",
|
|
312
|
+
"overflow": "hidden",
|
|
313
|
+
"textOverflow": "ellipsis",
|
|
314
|
+
"wordBreak": "break-word",
|
|
315
|
+
},
|
|
316
|
+
),
|
|
317
|
+
),
|
|
318
|
+
cursor="pointer",
|
|
319
|
+
on_click=EvalState.toggle_row_expand(row_id=row_id), # type: ignore[arg-type,call-arg,func-returns-value]
|
|
320
|
+
style={"minWidth": "0", "width": "100%"},
|
|
321
|
+
),
|
|
322
|
+
style={"minWidth": "0"},
|
|
323
|
+
)
|
|
324
|
+
|
|
325
|
+
def _row(row: dict[str, rx.Var]) -> rx.Component:
|
|
326
|
+
return rx.table.row(
|
|
327
|
+
rx.table.cell(rx.text(row.get("test_date", ""))),
|
|
328
|
+
_expandable_text(row, "gds_question"),
|
|
329
|
+
rx.table.cell(rx.text(row.get("pipeline_model", ""))),
|
|
330
|
+
_expandable_text(row, "llm_answer"),
|
|
331
|
+
_expandable_text(row, "gds_answer"),
|
|
332
|
+
rx.table.cell(
|
|
333
|
+
rx.badge(
|
|
334
|
+
row.get("test_status", ""),
|
|
335
|
+
color_scheme=row.get("status_color", "gray"),
|
|
336
|
+
variant="soft",
|
|
337
|
+
)
|
|
338
|
+
),
|
|
339
|
+
rx.table.cell(rx.text(row.get("eval_judge_rating", "—"))),
|
|
340
|
+
_expandable_text(row, "eval_judge_comment"),
|
|
341
|
+
)
|
|
342
|
+
|
|
343
|
+
return rx.card(
|
|
344
|
+
rx.vstack(
|
|
345
|
+
rx.hstack(
|
|
346
|
+
rx.heading("Test results", size="4"),
|
|
347
|
+
rx.spacer(),
|
|
348
|
+
rx.text(EvalState.pass_fail_summary, size="2", color="gray"),
|
|
349
|
+
rx.badge(EvalState.cost_label, color_scheme="gray", variant="soft"),
|
|
350
|
+
rx.select(
|
|
351
|
+
items=EvalState.tests_sort_options,
|
|
352
|
+
value=EvalState.selected_tests_sort,
|
|
353
|
+
placeholder="Sort",
|
|
354
|
+
on_change=EvalState.select_tests_sort,
|
|
355
|
+
width="14em",
|
|
356
|
+
),
|
|
357
|
+
rx.select(
|
|
358
|
+
items=EvalState.tests_scenario_options,
|
|
359
|
+
value=EvalState.selected_tests_scenario,
|
|
360
|
+
placeholder="Scenario (All)",
|
|
361
|
+
on_change=EvalState.select_tests_scenario,
|
|
362
|
+
width="14em",
|
|
363
|
+
),
|
|
364
|
+
rx.select(
|
|
365
|
+
items=EvalState.run_id_options,
|
|
366
|
+
value=EvalState.selected_run_id,
|
|
367
|
+
placeholder="Run id (All)",
|
|
368
|
+
on_change=EvalState.select_run,
|
|
369
|
+
width="18em",
|
|
370
|
+
),
|
|
371
|
+
align="center",
|
|
372
|
+
width="100%",
|
|
373
|
+
),
|
|
374
|
+
rx.scroll_area(
|
|
375
|
+
rx.table.root(
|
|
376
|
+
rx.table.header(
|
|
377
|
+
rx.table.row(
|
|
378
|
+
rx.table.column_header_cell("Run at"),
|
|
379
|
+
rx.table.column_header_cell("Question"),
|
|
380
|
+
rx.table.column_header_cell("Pipeline"),
|
|
381
|
+
rx.table.column_header_cell("Answer"),
|
|
382
|
+
rx.table.column_header_cell("Golden Answer"),
|
|
383
|
+
rx.table.column_header_cell("Status"),
|
|
384
|
+
rx.table.column_header_cell("Rating"),
|
|
385
|
+
rx.table.column_header_cell("Judge comment"),
|
|
386
|
+
)
|
|
387
|
+
),
|
|
388
|
+
rx.table.body(rx.foreach(EvalState.tests_rows, _row)),
|
|
389
|
+
variant="surface",
|
|
390
|
+
style={
|
|
391
|
+
"width": "100%",
|
|
392
|
+
"maxWidth": "100%",
|
|
393
|
+
"tableLayout": "fixed",
|
|
394
|
+
},
|
|
395
|
+
),
|
|
396
|
+
type="always",
|
|
397
|
+
scrollbars="vertical",
|
|
398
|
+
style={
|
|
399
|
+
"maxHeight": "80vh",
|
|
400
|
+
"width": "100%",
|
|
401
|
+
"maxWidth": "100%",
|
|
402
|
+
},
|
|
403
|
+
),
|
|
404
|
+
# Server-side pagination controls
|
|
405
|
+
rx.hstack(
|
|
406
|
+
rx.text(EvalState.tests_rows_display, size="2", color="gray"), # type: ignore[arg-type]
|
|
407
|
+
rx.spacer(),
|
|
408
|
+
rx.hstack(
|
|
409
|
+
rx.button(
|
|
410
|
+
"⏮",
|
|
411
|
+
variant="soft",
|
|
412
|
+
size="1",
|
|
413
|
+
on_click=EvalState.tests_first_page,
|
|
414
|
+
disabled=~EvalState.tests_has_prev, # type: ignore[operator]
|
|
415
|
+
),
|
|
416
|
+
rx.button(
|
|
417
|
+
"← Prev",
|
|
418
|
+
variant="soft",
|
|
419
|
+
size="1",
|
|
420
|
+
on_click=EvalState.tests_prev_page,
|
|
421
|
+
disabled=~EvalState.tests_has_prev, # type: ignore[operator]
|
|
422
|
+
),
|
|
423
|
+
rx.text(
|
|
424
|
+
EvalState.tests_page_display, # type: ignore[arg-type]
|
|
425
|
+
size="2",
|
|
426
|
+
style={"minWidth": "100px", "textAlign": "center"},
|
|
427
|
+
),
|
|
428
|
+
rx.button(
|
|
429
|
+
"Next →",
|
|
430
|
+
variant="soft",
|
|
431
|
+
size="1",
|
|
432
|
+
on_click=EvalState.tests_next_page,
|
|
433
|
+
disabled=~EvalState.tests_has_next, # type: ignore[operator]
|
|
434
|
+
),
|
|
435
|
+
rx.button(
|
|
436
|
+
"⏭",
|
|
437
|
+
variant="soft",
|
|
438
|
+
size="1",
|
|
439
|
+
on_click=EvalState.tests_last_page,
|
|
440
|
+
disabled=~EvalState.tests_has_next, # type: ignore[operator]
|
|
441
|
+
),
|
|
442
|
+
spacing="2",
|
|
443
|
+
align="center",
|
|
444
|
+
),
|
|
445
|
+
align="center",
|
|
446
|
+
width="100%",
|
|
447
|
+
),
|
|
448
|
+
spacing="3",
|
|
449
|
+
),
|
|
450
|
+
padding="1em",
|
|
451
|
+
width="100%",
|
|
452
|
+
)
|
|
453
|
+
|
|
454
|
+
|
|
455
|
+
def _compare_card() -> rx.Component:
|
|
456
|
+
return rx.card(
|
|
457
|
+
rx.vstack(
|
|
458
|
+
rx.heading("Compare runs", size="4"),
|
|
459
|
+
rx.vstack(
|
|
460
|
+
rx.select(
|
|
461
|
+
items=EvalState.run_options_only,
|
|
462
|
+
value=EvalState.compare_run_a,
|
|
463
|
+
placeholder="Run A",
|
|
464
|
+
on_change=EvalState.set_compare_run_a,
|
|
465
|
+
width="100%",
|
|
466
|
+
),
|
|
467
|
+
rx.select(
|
|
468
|
+
items=EvalState.run_options_only,
|
|
469
|
+
value=EvalState.compare_run_b,
|
|
470
|
+
placeholder="Run B",
|
|
471
|
+
on_change=EvalState.set_compare_run_b,
|
|
472
|
+
width="100%",
|
|
473
|
+
),
|
|
474
|
+
spacing="2",
|
|
475
|
+
align="center",
|
|
476
|
+
width="100%",
|
|
477
|
+
),
|
|
478
|
+
rx.button(
|
|
479
|
+
"Compare",
|
|
480
|
+
on_click=EvalState.compare_runs,
|
|
481
|
+
disabled=~EvalState.can_compare_runs, # type: ignore[operator]
|
|
482
|
+
loading=EvalState.compare_loading,
|
|
483
|
+
width="100%",
|
|
484
|
+
margin_top="0.5em",
|
|
485
|
+
),
|
|
486
|
+
rx.cond(
|
|
487
|
+
EvalState.compare_error != "",
|
|
488
|
+
rx.callout(EvalState.compare_error, color_scheme="red", variant="soft"),
|
|
489
|
+
rx.box(),
|
|
490
|
+
),
|
|
491
|
+
spacing="3",
|
|
492
|
+
width="100%",
|
|
493
|
+
),
|
|
494
|
+
padding="1em",
|
|
495
|
+
width="100%",
|
|
496
|
+
)
|
|
497
|
+
|
|
498
|
+
|
|
499
|
+
def _compare_dialog() -> rx.Component:
|
|
500
|
+
def _stat_block(label: str, summary: RunSummary) -> rx.Component:
|
|
501
|
+
avg_time = summary["avg_answer_time_sec"]
|
|
502
|
+
median_time = summary["median_answer_time_sec"]
|
|
503
|
+
return rx.card(
|
|
504
|
+
rx.vstack(
|
|
505
|
+
rx.text(label, weight="medium"),
|
|
506
|
+
rx.hstack(
|
|
507
|
+
rx.badge(
|
|
508
|
+
rx.text(
|
|
509
|
+
rx.cond(
|
|
510
|
+
summary["tests_total"],
|
|
511
|
+
f"Pass: {summary['passed']}",
|
|
512
|
+
"No tests",
|
|
513
|
+
),
|
|
514
|
+
size="1",
|
|
515
|
+
),
|
|
516
|
+
color_scheme="green",
|
|
517
|
+
variant="soft",
|
|
518
|
+
),
|
|
519
|
+
rx.badge(
|
|
520
|
+
f"Fail: {summary['failed']}",
|
|
521
|
+
color_scheme="red",
|
|
522
|
+
variant="soft",
|
|
523
|
+
),
|
|
524
|
+
rx.badge(
|
|
525
|
+
f"Rating: {summary['avg_rating']}",
|
|
526
|
+
color_scheme="blue",
|
|
527
|
+
variant="soft",
|
|
528
|
+
),
|
|
529
|
+
rx.badge(
|
|
530
|
+
f"Cost: ${summary['cost_total']:.3f}",
|
|
531
|
+
color_scheme="gray",
|
|
532
|
+
variant="soft",
|
|
533
|
+
),
|
|
534
|
+
rx.badge(
|
|
535
|
+
rx.text(
|
|
536
|
+
f"Time (avg/med): {avg_time} / {median_time}",
|
|
537
|
+
size="1",
|
|
538
|
+
),
|
|
539
|
+
color_scheme="purple",
|
|
540
|
+
variant="soft",
|
|
541
|
+
),
|
|
542
|
+
spacing="2",
|
|
543
|
+
align="center",
|
|
544
|
+
),
|
|
545
|
+
spacing="2",
|
|
546
|
+
),
|
|
547
|
+
padding="0.75em",
|
|
548
|
+
width="100%",
|
|
549
|
+
)
|
|
550
|
+
|
|
551
|
+
def _config_block(rows: list[dict[str, rx.Var]]) -> rx.Component:
|
|
552
|
+
def _row(row: dict[str, rx.Var]) -> rx.Component:
|
|
553
|
+
return rx.cond(
|
|
554
|
+
row["diff"],
|
|
555
|
+
rx.hstack(
|
|
556
|
+
rx.text(row["label"], weight="medium", size="1"),
|
|
557
|
+
rx.spacer(),
|
|
558
|
+
rx.text(row["value"], size="1"),
|
|
559
|
+
style={"padding": "0.15em 0.25em", "backgroundColor": "var(--amber-3)"},
|
|
560
|
+
),
|
|
561
|
+
rx.hstack(
|
|
562
|
+
rx.text(row["label"], weight="medium", size="1"),
|
|
563
|
+
rx.spacer(),
|
|
564
|
+
rx.text(row["value"], size="1"),
|
|
565
|
+
style={"padding": "0.15em 0.25em"},
|
|
566
|
+
),
|
|
567
|
+
)
|
|
568
|
+
|
|
569
|
+
return rx.card(
|
|
570
|
+
rx.vstack(
|
|
571
|
+
rx.text("Config", weight="medium"),
|
|
572
|
+
rx.foreach(rows, _row),
|
|
573
|
+
spacing="1",
|
|
574
|
+
),
|
|
575
|
+
padding="0.75em",
|
|
576
|
+
width="100%",
|
|
577
|
+
)
|
|
578
|
+
|
|
579
|
+
def _diff_table(title: str, rows: list[dict[str, rx.Var]]) -> rx.Component:
|
|
580
|
+
def _line(row: dict[str, rx.Var]) -> rx.Component:
|
|
581
|
+
left_border = rx.cond(
|
|
582
|
+
row.get("strong", False),
|
|
583
|
+
f"2px solid {row.get('left_color', 'inherit')}",
|
|
584
|
+
"2px solid transparent",
|
|
585
|
+
)
|
|
586
|
+
right_border = rx.cond(
|
|
587
|
+
row.get("strong", False),
|
|
588
|
+
f"2px solid {row.get('right_color', 'inherit')}",
|
|
589
|
+
"2px solid transparent",
|
|
590
|
+
)
|
|
591
|
+
return rx.hstack(
|
|
592
|
+
rx.box(
|
|
593
|
+
rx.text(
|
|
594
|
+
row.get("left", ""),
|
|
595
|
+
size="1",
|
|
596
|
+
white_space="pre-wrap",
|
|
597
|
+
weight=rx.cond(row.get("strong", False), "bold", "regular"),
|
|
598
|
+
color=row.get("left_color", "inherit"),
|
|
599
|
+
),
|
|
600
|
+
style={
|
|
601
|
+
"fontFamily": "monospace",
|
|
602
|
+
"padding": "1px 3px",
|
|
603
|
+
"borderRadius": "4px",
|
|
604
|
+
"borderLeft": left_border,
|
|
605
|
+
"width": "100%",
|
|
606
|
+
},
|
|
607
|
+
),
|
|
608
|
+
rx.box(
|
|
609
|
+
rx.text(
|
|
610
|
+
row.get("right", ""),
|
|
611
|
+
size="1",
|
|
612
|
+
white_space="pre-wrap",
|
|
613
|
+
weight=rx.cond(row.get("strong", False), "bold", "regular"),
|
|
614
|
+
color=row.get("right_color", "inherit"),
|
|
615
|
+
),
|
|
616
|
+
style={
|
|
617
|
+
"fontFamily": "monospace",
|
|
618
|
+
"padding": "1px 3px",
|
|
619
|
+
"borderRadius": "4px",
|
|
620
|
+
"borderLeft": right_border,
|
|
621
|
+
"width": "100%",
|
|
622
|
+
},
|
|
623
|
+
),
|
|
624
|
+
spacing="1",
|
|
625
|
+
width="100%",
|
|
626
|
+
)
|
|
627
|
+
|
|
628
|
+
return rx.vstack(
|
|
629
|
+
rx.text(title, weight="medium"),
|
|
630
|
+
rx.hstack(
|
|
631
|
+
rx.text(EvalState.compare_run_label_a, weight="medium", size="1"),
|
|
632
|
+
rx.spacer(),
|
|
633
|
+
rx.text(EvalState.compare_run_label_b, weight="medium", size="1"),
|
|
634
|
+
width="100%",
|
|
635
|
+
),
|
|
636
|
+
rx.scroll_area(
|
|
637
|
+
rx.vstack(
|
|
638
|
+
rx.foreach(rows, _line),
|
|
639
|
+
spacing="1",
|
|
640
|
+
width="100%",
|
|
641
|
+
),
|
|
642
|
+
type="always",
|
|
643
|
+
scrollbars="vertical",
|
|
644
|
+
style={"maxHeight": "22vh", "padding": "1px"},
|
|
645
|
+
),
|
|
646
|
+
spacing="1",
|
|
647
|
+
width="100%",
|
|
648
|
+
)
|
|
649
|
+
|
|
650
|
+
def _result_row(row: dict[str, rx.Var]) -> rx.Component:
|
|
651
|
+
def _badge_color(status: str | None) -> rx.Var:
|
|
652
|
+
"""
|
|
653
|
+
green for pass, red for fail, grey else
|
|
654
|
+
"""
|
|
655
|
+
return rx.cond(
|
|
656
|
+
status == "pass",
|
|
657
|
+
"green",
|
|
658
|
+
rx.cond(status == "fail", "red", "gray"),
|
|
659
|
+
)
|
|
660
|
+
|
|
661
|
+
def _answer_block(text: str, tool_calls: str) -> rx.Component:
|
|
662
|
+
return rx.vstack(
|
|
663
|
+
rx.text(text, size="1", white_space="pre-wrap", style={"wordBreak": "break-word"}),
|
|
664
|
+
rx.cond(
|
|
665
|
+
tool_calls != "",
|
|
666
|
+
rx.accordion.root(
|
|
667
|
+
rx.accordion.item(
|
|
668
|
+
rx.accordion.trigger("Tool calls", style={"fontSize": "12px"}),
|
|
669
|
+
rx.accordion.content(
|
|
670
|
+
rx.text(
|
|
671
|
+
tool_calls,
|
|
672
|
+
size="1",
|
|
673
|
+
color="gray",
|
|
674
|
+
white_space="pre-wrap",
|
|
675
|
+
style={"wordBreak": "break-word"},
|
|
676
|
+
)
|
|
677
|
+
),
|
|
678
|
+
value="tool-calls",
|
|
679
|
+
),
|
|
680
|
+
collapsible=True,
|
|
681
|
+
type="single",
|
|
682
|
+
default_value="",
|
|
683
|
+
variant="ghost",
|
|
684
|
+
style={"width": "100%"},
|
|
685
|
+
),
|
|
686
|
+
rx.box(),
|
|
687
|
+
),
|
|
688
|
+
spacing="1",
|
|
689
|
+
align="start",
|
|
690
|
+
width="100%",
|
|
691
|
+
)
|
|
692
|
+
|
|
693
|
+
return rx.table.row(
|
|
694
|
+
rx.table.cell(
|
|
695
|
+
rx.vstack(
|
|
696
|
+
rx.text(row["question"]),
|
|
697
|
+
rx.text(
|
|
698
|
+
row["golden_answer"],
|
|
699
|
+
size="1",
|
|
700
|
+
color="gray",
|
|
701
|
+
white_space="pre-wrap",
|
|
702
|
+
style={"wordBreak": "break-word"},
|
|
703
|
+
),
|
|
704
|
+
spacing="1",
|
|
705
|
+
align="start",
|
|
706
|
+
width="100%",
|
|
707
|
+
)
|
|
708
|
+
),
|
|
709
|
+
rx.table.cell(
|
|
710
|
+
rx.vstack(
|
|
711
|
+
rx.hstack(
|
|
712
|
+
rx.badge(
|
|
713
|
+
row["status_a"],
|
|
714
|
+
color_scheme=_badge_color(row["status_a"]), # type: ignore[arg-type]
|
|
715
|
+
variant="soft",
|
|
716
|
+
),
|
|
717
|
+
rx.text(f"Rating: {row['rating_a']}", size="1"),
|
|
718
|
+
align="center",
|
|
719
|
+
),
|
|
720
|
+
_answer_block(row["answer_a"], row["tool_calls_a"]), # type: ignore[arg-type]
|
|
721
|
+
rx.text(row["comment_a"], size="1", color="gray"),
|
|
722
|
+
spacing="1",
|
|
723
|
+
align="start",
|
|
724
|
+
)
|
|
725
|
+
),
|
|
726
|
+
rx.table.cell(
|
|
727
|
+
rx.vstack(
|
|
728
|
+
rx.hstack(
|
|
729
|
+
rx.badge(
|
|
730
|
+
row["status_b"],
|
|
731
|
+
color_scheme=_badge_color(row["status_b"]), # type: ignore[arg-type]
|
|
732
|
+
variant="soft",
|
|
733
|
+
),
|
|
734
|
+
rx.text(f"Rating: {row['rating_b']}", size="1"),
|
|
735
|
+
align="center",
|
|
736
|
+
),
|
|
737
|
+
_answer_block(row["answer_b"], row["tool_calls_b"]), # type: ignore[arg-type]
|
|
738
|
+
rx.text(row["comment_b"], size="1", color="gray"),
|
|
739
|
+
spacing="1",
|
|
740
|
+
align="start",
|
|
741
|
+
)
|
|
742
|
+
),
|
|
743
|
+
)
|
|
744
|
+
|
|
745
|
+
return rx.dialog.root(
|
|
746
|
+
rx.dialog.content(
|
|
747
|
+
rx.dialog.title("Run comparison"),
|
|
748
|
+
rx.vstack(
|
|
749
|
+
rx.hstack(
|
|
750
|
+
_stat_block(EvalState.compare_run_label_a, EvalState.compare_summary_a), # type: ignore[arg-type]
|
|
751
|
+
_stat_block(EvalState.compare_run_label_b, EvalState.compare_summary_b), # type: ignore[arg-type]
|
|
752
|
+
spacing="3",
|
|
753
|
+
width="100%",
|
|
754
|
+
),
|
|
755
|
+
rx.hstack(
|
|
756
|
+
_config_block(EvalState.compare_config_a_rows), # type: ignore[arg-type]
|
|
757
|
+
_config_block(EvalState.compare_config_b_rows), # type: ignore[arg-type]
|
|
758
|
+
spacing="3",
|
|
759
|
+
width="100%",
|
|
760
|
+
),
|
|
761
|
+
rx.cond(
|
|
762
|
+
EvalState.compare_diff_keys != [],
|
|
763
|
+
rx.card(
|
|
764
|
+
rx.vstack(
|
|
765
|
+
rx.hstack(
|
|
766
|
+
rx.text("Differences:", weight="medium", size="1"),
|
|
767
|
+
rx.box(rx.foreach(EvalState.compare_diff_keys, lambda k: rx.badge(k, variant="soft"))),
|
|
768
|
+
spacing="3",
|
|
769
|
+
align="center",
|
|
770
|
+
),
|
|
771
|
+
rx.accordion.root(
|
|
772
|
+
rx.accordion.item(
|
|
773
|
+
rx.accordion.trigger("Judge prompt diff"),
|
|
774
|
+
rx.accordion.content(
|
|
775
|
+
rx.vstack(
|
|
776
|
+
rx.checkbox(
|
|
777
|
+
"Show only changes",
|
|
778
|
+
default_checked=True,
|
|
779
|
+
checked=EvalState.compare_judge_prompt_compact,
|
|
780
|
+
on_change=EvalState.set_compare_judge_prompt_compact,
|
|
781
|
+
size="2",
|
|
782
|
+
),
|
|
783
|
+
_diff_table("Judge prompt diff", EvalState.compare_prompt_rows_view), # type: ignore[arg-type]
|
|
784
|
+
spacing="1",
|
|
785
|
+
)
|
|
786
|
+
),
|
|
787
|
+
value="prompt-diff-block",
|
|
788
|
+
),
|
|
789
|
+
rx.accordion.item(
|
|
790
|
+
rx.accordion.trigger("Data model diff"),
|
|
791
|
+
rx.accordion.content(
|
|
792
|
+
rx.vstack(
|
|
793
|
+
rx.checkbox(
|
|
794
|
+
"Show only changes",
|
|
795
|
+
default_checked=True,
|
|
796
|
+
checked=EvalState.compare_dm_compact,
|
|
797
|
+
on_change=EvalState.set_compare_dm_compact,
|
|
798
|
+
size="2",
|
|
799
|
+
),
|
|
800
|
+
_diff_table("Data model diff", EvalState.compare_dm_rows_view), # type: ignore[arg-type]
|
|
801
|
+
spacing="1",
|
|
802
|
+
)
|
|
803
|
+
),
|
|
804
|
+
value="dm-diff-block",
|
|
805
|
+
),
|
|
806
|
+
type="multiple",
|
|
807
|
+
collapsible=True,
|
|
808
|
+
variant="outline",
|
|
809
|
+
width="100%",
|
|
810
|
+
),
|
|
811
|
+
spacing="2",
|
|
812
|
+
width="100%",
|
|
813
|
+
),
|
|
814
|
+
variant="surface",
|
|
815
|
+
width="100%",
|
|
816
|
+
),
|
|
817
|
+
rx.box(),
|
|
818
|
+
),
|
|
819
|
+
rx.cond(
|
|
820
|
+
EvalState.compare_loading,
|
|
821
|
+
rx.center(rx.spinner(size="3"), height="200px"),
|
|
822
|
+
rx.scroll_area(
|
|
823
|
+
rx.table.root(
|
|
824
|
+
rx.table.header(
|
|
825
|
+
rx.table.row(
|
|
826
|
+
rx.table.column_header_cell("Question", style={"width": "25%"}),
|
|
827
|
+
rx.table.column_header_cell(
|
|
828
|
+
EvalState.compare_run_label_a, style={"width": "37.5%"}
|
|
829
|
+
),
|
|
830
|
+
rx.table.column_header_cell(
|
|
831
|
+
EvalState.compare_run_label_b, style={"width": "37.5%"}
|
|
832
|
+
),
|
|
833
|
+
)
|
|
834
|
+
),
|
|
835
|
+
rx.table.body(rx.foreach(EvalState.compare_rows, _result_row)),
|
|
836
|
+
variant="surface",
|
|
837
|
+
style={"width": "100%", "tableLayout": "fixed"},
|
|
838
|
+
),
|
|
839
|
+
type="always",
|
|
840
|
+
scrollbars="vertical",
|
|
841
|
+
style={"maxHeight": "60vh"},
|
|
842
|
+
),
|
|
843
|
+
),
|
|
844
|
+
rx.dialog.close(rx.button("Close", variant="soft")),
|
|
845
|
+
spacing="3",
|
|
846
|
+
width="100%",
|
|
847
|
+
),
|
|
848
|
+
style={"maxWidth": "92vw"},
|
|
849
|
+
),
|
|
850
|
+
open=EvalState.compare_dialog_open,
|
|
851
|
+
on_open_change=EvalState.set_compare_dialog_open,
|
|
852
|
+
)
|
|
853
|
+
|
|
854
|
+
|
|
855
|
+
def _judge_prompt_dialog() -> rx.Component:
|
|
856
|
+
return rx.dialog.root(
|
|
857
|
+
rx.dialog.content(
|
|
858
|
+
rx.dialog.title("Judge Prompt"),
|
|
859
|
+
rx.vstack(
|
|
860
|
+
rx.text(f"Prompt ID: {EvalState.judge_prompt_id}", size="2", color="gray"),
|
|
861
|
+
rx.box(
|
|
862
|
+
rx.text(
|
|
863
|
+
rx.cond(
|
|
864
|
+
EvalState.judge_prompt != "",
|
|
865
|
+
EvalState.judge_prompt,
|
|
866
|
+
"Prompt not loaded",
|
|
867
|
+
),
|
|
868
|
+
size="2",
|
|
869
|
+
),
|
|
870
|
+
padding="1em",
|
|
871
|
+
border="1px solid var(--gray-6)",
|
|
872
|
+
border_radius="8px",
|
|
873
|
+
style={"maxHeight": "60vh", "overflow": "auto", "whiteSpace": "pre-wrap"},
|
|
874
|
+
),
|
|
875
|
+
rx.dialog.close(
|
|
876
|
+
rx.button("Close", variant="soft"),
|
|
877
|
+
),
|
|
878
|
+
spacing="3",
|
|
879
|
+
width="100%",
|
|
880
|
+
),
|
|
881
|
+
style={"maxWidth": "800px"},
|
|
882
|
+
),
|
|
883
|
+
open=EvalState.judge_prompt_dialog_open,
|
|
884
|
+
on_open_change=EvalState.set_judge_prompt_dialog_open,
|
|
885
|
+
)
|
|
886
|
+
|
|
887
|
+
|
|
888
|
+
def _data_model_dialog() -> rx.Component:
|
|
889
|
+
return rx.dialog.root(
|
|
890
|
+
rx.dialog.content(
|
|
891
|
+
rx.dialog.title("Data Model"),
|
|
892
|
+
rx.vstack(
|
|
893
|
+
rx.text(f"Model ID: {EvalState.dm_id}", size="2", color="gray"),
|
|
894
|
+
rx.box(
|
|
895
|
+
rx.text(
|
|
896
|
+
rx.cond(
|
|
897
|
+
EvalState.dm_description != "",
|
|
898
|
+
EvalState.dm_description,
|
|
899
|
+
"Description not loaded",
|
|
900
|
+
),
|
|
901
|
+
size="2",
|
|
902
|
+
),
|
|
903
|
+
padding="1em",
|
|
904
|
+
border="1px solid var(--gray-6)",
|
|
905
|
+
border_radius="8px",
|
|
906
|
+
style={"maxHeight": "60vh", "overflow": "auto", "whiteSpace": "pre-wrap"},
|
|
907
|
+
),
|
|
908
|
+
rx.dialog.close(
|
|
909
|
+
rx.button("Close", variant="soft"),
|
|
910
|
+
),
|
|
911
|
+
spacing="3",
|
|
912
|
+
width="100%",
|
|
913
|
+
),
|
|
914
|
+
style={"maxWidth": "800px"},
|
|
915
|
+
),
|
|
916
|
+
open=EvalState.data_model_dialog_open,
|
|
917
|
+
on_open_change=EvalState.set_data_model_dialog_open,
|
|
918
|
+
)
|
|
919
|
+
|
|
920
|
+
|
|
921
|
+
def _status_messages() -> rx.Component:
|
|
922
|
+
return rx.vstack(
|
|
923
|
+
rx.cond(
|
|
924
|
+
EvalState.status_message != "",
|
|
925
|
+
rx.callout(
|
|
926
|
+
rx.vstack(
|
|
927
|
+
rx.text(EvalState.status_message, weight="medium"),
|
|
928
|
+
rx.cond(
|
|
929
|
+
EvalState.current_question_progress != "",
|
|
930
|
+
rx.text(EvalState.current_question_progress, size="1", color="gray"), # type: ignore[arg-type]
|
|
931
|
+
rx.box(),
|
|
932
|
+
),
|
|
933
|
+
spacing="1",
|
|
934
|
+
),
|
|
935
|
+
color_scheme="green",
|
|
936
|
+
variant="soft",
|
|
937
|
+
width="100%",
|
|
938
|
+
),
|
|
939
|
+
rx.box(),
|
|
940
|
+
),
|
|
941
|
+
rx.cond(
|
|
942
|
+
EvalState.error_message != "",
|
|
943
|
+
rx.callout(EvalState.error_message, color_scheme="red", variant="soft"),
|
|
944
|
+
rx.box(),
|
|
945
|
+
),
|
|
946
|
+
rx.cond(
|
|
947
|
+
EvalState.has_run_progress,
|
|
948
|
+
rx.card(
|
|
949
|
+
rx.vstack(
|
|
950
|
+
rx.heading("Run log", size="3"),
|
|
951
|
+
rx.scroll_area(
|
|
952
|
+
rx.vstack(
|
|
953
|
+
rx.foreach(
|
|
954
|
+
EvalState.run_progress,
|
|
955
|
+
lambda line: rx.text(line, size="1"),
|
|
956
|
+
)
|
|
957
|
+
),
|
|
958
|
+
type="always",
|
|
959
|
+
scrollbars="vertical",
|
|
960
|
+
style={"height": "160px"},
|
|
961
|
+
),
|
|
962
|
+
spacing="2",
|
|
963
|
+
),
|
|
964
|
+
padding="1em",
|
|
965
|
+
width="100%",
|
|
966
|
+
),
|
|
967
|
+
rx.box(),
|
|
968
|
+
),
|
|
969
|
+
spacing="3",
|
|
970
|
+
width="100%",
|
|
971
|
+
)
|
|
972
|
+
|
|
973
|
+
|
|
974
|
+
def page() -> rx.Component:
|
|
975
|
+
return rx.vstack(
|
|
976
|
+
app_header(),
|
|
977
|
+
rx.vstack(
|
|
978
|
+
rx.grid(
|
|
979
|
+
rx.vstack(
|
|
980
|
+
_questions_card(),
|
|
981
|
+
_tests_card(),
|
|
982
|
+
),
|
|
983
|
+
rx.vstack(
|
|
984
|
+
_judge_card(),
|
|
985
|
+
_pipeline_card(),
|
|
986
|
+
_selection_and_actions(),
|
|
987
|
+
_status_messages(),
|
|
988
|
+
_compare_card(),
|
|
989
|
+
spacing="4",
|
|
990
|
+
width="100%",
|
|
991
|
+
),
|
|
992
|
+
columns="2",
|
|
993
|
+
spacing="4",
|
|
994
|
+
width="100%",
|
|
995
|
+
style={"gridTemplateColumns": "3fr 1fr", "height": "calc(100vh - 200px)", "minHeight": "700px"},
|
|
996
|
+
),
|
|
997
|
+
spacing="4",
|
|
998
|
+
width="100%",
|
|
999
|
+
),
|
|
1000
|
+
_compare_dialog(),
|
|
1001
|
+
_judge_prompt_dialog(),
|
|
1002
|
+
_data_model_dialog(),
|
|
1003
|
+
align="start",
|
|
1004
|
+
spacing="2",
|
|
1005
|
+
padding="1.5em",
|
|
1006
|
+
)
|