File size: 11,645 Bytes
6bef416 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 | from __future__ import annotations
import pandas as pd
import streamlit as st
from src.analytics import (
MIN_CONFIG_N,
RETRIEVAL_OK_THRESHOLD,
demand_coverage,
make_decision_brief,
overview_metrics,
retrieval_outcomes,
risk_slices,
)
from src.app_state import DashboardContext, FilterState
from src.data import DataBundle, filter_eval, filter_retrieval_events, load_bundle, option_values
from src.formatting import fmt_int, fmt_latency_ms, fmt_money, fmt_pct
from src.models import APP_TITLE, AppSettings
from src.ui import badge, decision_strip, hero, inject_css, metric_card
from src.views import (
ConfigComparisonViewMixin,
ExportCenterViewMixin,
OverviewViewMixin,
PolicySimulatorViewMixin,
QualityMapViewMixin,
RetrievalLabViewMixin,
RiskBoardViewMixin,
TraceExplorerViewMixin,
)
@st.cache_data(show_spinner="Loading RAG QA logs and corpus tables...")
def cached_load_bundle(data_dir: str, docs_dir: str) -> DataBundle:
return load_bundle(data_dir, docs_dir)
class CommandCenterApp(
OverviewViewMixin,
QualityMapViewMixin,
RetrievalLabViewMixin,
RiskBoardViewMixin,
ConfigComparisonViewMixin,
PolicySimulatorViewMixin,
TraceExplorerViewMixin,
ExportCenterViewMixin,
):
"""Thin Streamlit controller for the RAG QA Command Center.
The controller owns application state and page orchestration. Individual tabs live
in focused view mixins under ``src/views`` so UI concerns stay separated from
data loading, analytics, chart construction, and formatting helpers.
"""
def __init__(self, settings: AppSettings | None = None) -> None:
self.settings = settings or AppSettings()
self.bundle: DataBundle | None = None
self.filter_state: FilterState | None = None
self.context: DashboardContext | None = None
def run(self) -> None:
st.set_page_config(
page_title=APP_TITLE,
page_icon=self.settings.page_icon,
layout="wide",
initial_sidebar_state="expanded",
)
inject_css()
try:
self.bundle = cached_load_bundle(self.settings.data_dir, self.settings.docs_dir)
except Exception as exc: # pragma: no cover - rendered in Streamlit runtime
st.error(str(exc))
st.stop()
self.filter_state = self._render_sidebar(self.bundle.eval_runs)
self.context = self._build_context(self.bundle, self.filter_state)
self._render_header(self.context, self.filter_state)
if self.context.filtered_eval.empty:
st.warning("No rows match the current filters. Relax the sidebar filters to continue.")
st.stop()
self._render_kpis(self.context.metrics)
self._render_pages()
def _render_sidebar(self, eval_df: pd.DataFrame) -> FilterState:
st.sidebar.markdown("### Filters")
preset = st.sidebar.selectbox(
"Focus preset",
["All", "Error review", "Hallucination risk", "Retrieval failure", "High latency", "High cost"],
index=0,
help="Quickly narrow the evaluation rows to a common review queue.",
)
domains = st.sidebar.multiselect("Domain", option_values(eval_df, "domain"))
difficulties = st.sidebar.multiselect("Difficulty", option_values(eval_df, "difficulty"))
scenario_types = st.sidebar.multiselect("Scenario type", option_values(eval_df, "scenario_type"))
retrievers = st.sidebar.multiselect("Retriever", option_values(eval_df, "retrieval_strategy"))
generators = st.sidebar.multiselect("Generator", option_values(eval_df, "generator_model"))
splits = st.sidebar.multiselect("Split", option_values(eval_df, "split"))
min_slice_n = st.sidebar.slider(
"Minimum rows per risk slice",
min_value=10,
max_value=200,
value=30,
step=5,
help="Higher values reduce noisy small-sample slices.",
)
min_config_n = st.sidebar.slider(
"Minimum rows per config",
min_value=20,
max_value=250,
value=MIN_CONFIG_N,
step=5,
help="Higher values make configuration comparisons more reliable.",
)
return FilterState(
preset=preset,
domains=domains,
difficulties=difficulties,
scenario_types=scenario_types,
retrievers=retrievers,
generators=generators,
splits=splits,
min_slice_n=min_slice_n,
min_config_n=min_config_n,
)
def _build_context(self, bundle: DataBundle, filters: FilterState) -> DashboardContext:
filtered_eval = filter_eval(
bundle.eval_runs,
domains=filters.domains,
difficulties=filters.difficulties,
scenario_types=filters.scenario_types,
retrievers=filters.retrievers,
generators=filters.generators,
splits=filters.splits,
)
filtered_eval = self._apply_preset(filtered_eval, filters.preset)
filtered_retrieval = filter_retrieval_events(bundle.retrieval_events, filtered_eval)
metrics = overview_metrics(filtered_eval, bundle.documents, bundle.chunks, filtered_retrieval)
risk_table = risk_slices(filtered_eval, min_n=filters.min_slice_n)
retrieval_table = retrieval_outcomes(filtered_eval)
brief = make_decision_brief(
filtered_eval,
bundle.documents,
bundle.chunks,
filtered_retrieval,
min_slice_n=filters.min_slice_n,
min_config_n=filters.min_config_n,
risk_table=risk_table,
retrieval_table=retrieval_table,
)
return DashboardContext(
filtered_eval=filtered_eval,
filtered_retrieval=filtered_retrieval,
metrics=metrics,
brief=brief,
risk_slices=risk_table,
retrieval_outcomes=retrieval_table,
demand_coverage=demand_coverage(filtered_eval, bundle.documents),
)
@staticmethod
def _apply_preset(df: pd.DataFrame, preset: str) -> pd.DataFrame:
out = df.copy()
if preset == "Error review" and "is_correct" in out.columns:
return out[pd.to_numeric(out["is_correct"], errors="coerce").fillna(0.0) < 0.5]
if preset == "Hallucination risk" and "hallucination_flag" in out.columns:
return out[pd.to_numeric(out["hallucination_flag"], errors="coerce").fillna(0.0) >= 0.5]
if preset == "Retrieval failure" and "recall_at_10" in out.columns:
return out[pd.to_numeric(out["recall_at_10"], errors="coerce").fillna(0.0) < RETRIEVAL_OK_THRESHOLD]
if preset == "High latency" and "total_latency_ms" in out.columns:
cutoff = pd.to_numeric(out["total_latency_ms"], errors="coerce").quantile(0.90)
return out[pd.to_numeric(out["total_latency_ms"], errors="coerce") >= cutoff]
if preset == "High cost" and "total_cost_usd" in out.columns:
cutoff = pd.to_numeric(out["total_cost_usd"], errors="coerce").quantile(0.90)
return out[pd.to_numeric(out["total_cost_usd"], errors="coerce") >= cutoff]
return out
def _render_header(self, context: DashboardContext, filters: FilterState) -> None:
posture_kind = "good" if context.brief.posture == "Stable" else "bad" if context.brief.posture == "High Risk" else "warn"
hero(
APP_TITLE,
"Inspect offline RAG QA evaluation logs across quality, retrieval behavior, hallucination exposure, latency, cost, review thresholds, and example-level evidence.",
badges=[
badge(f"Posture: {context.brief.posture}", posture_kind),
badge(f"Preset: {filters.preset}", "info"),
badge(f"Evaluations: {fmt_int(context.metrics['evaluations'])}", "info"),
badge(f"Retrieval events: {fmt_int(context.metrics['retrieval_events'])}", "info"),
badge("Offline evaluation dataset", "warn"),
],
)
decision_strip(
[
("Quality posture", context.brief.posture, context.brief.posture_reason),
("Main driver", context.brief.main_driver, "Dominant non-healthy issue pattern under the active filters."),
("Highest-risk slice", context.brief.worst_slice, "Uses the current minimum risk-slice size."),
("Best config", context.brief.best_config, "Uses the current minimum config sample size."),
("Recommended action", context.brief.recommended_action, "Suggested next step from the filtered evidence."),
]
)
@staticmethod
def _render_kpis(metrics: dict[str, float]) -> None:
cols = st.columns(6)
card_specs = [
("Answer correctness", fmt_pct(metrics["correct_rate"]), "Share of examples judged correct.", "stable" if metrics["correct_rate"] >= 0.75 else "watch"),
("Hallucination rate", fmt_pct(metrics["hallucination_rate"]), "Lower is better; elevated segments should be reviewed.", "risk" if metrics["hallucination_rate"] >= 0.18 else "stable"),
("Recall@10", fmt_pct(metrics["recall_at_10"]), "Retrieval coverage signal at top 10 chunks.", "stable" if metrics["recall_at_10"] >= 0.65 else "watch"),
("MRR@10", fmt_pct(metrics["mrr_at_10"]), "Rank-sensitive retrieval quality signal.", "stable" if metrics["mrr_at_10"] >= 0.45 else "watch"),
("P95 latency", fmt_latency_ms(metrics["p95_latency_ms"]), "Tail latency for selected evaluation rows.", "watch"),
("Avg cost", fmt_money(metrics["avg_cost_usd"]), "Average estimated cost per run.", "info"),
]
for col, spec in zip(cols, card_specs, strict=False):
with col:
metric_card(*spec)
def _render_pages(self) -> None:
tabs = st.tabs(
[
"Overview",
"Quality Map",
"Retrieval Lab",
"Risk Board",
"Config Comparison",
"Policy Simulator",
"Trace Explorer",
"Export Center",
]
)
with tabs[0]:
self._page_overview()
with tabs[1]:
self._page_quality_map()
with tabs[2]:
self._page_retrieval_lab()
with tabs[3]:
self._page_risk_board()
with tabs[4]:
self._page_config_comparison()
with tabs[5]:
self._page_policy_simulator()
with tabs[6]:
self._page_trace_explorer()
with tabs[7]:
self._page_export_center()
def _plot(self, fig, key: str) -> None:
st.plotly_chart(
fig,
use_container_width=True,
key=f"plot_{key}",
config={"displayModeBar": False, "responsive": True},
)
@property
def ctx(self) -> DashboardContext:
if self.context is None:
raise RuntimeError("Dashboard context has not been built.")
return self.context
@property
def data(self) -> DataBundle:
if self.bundle is None:
raise RuntimeError("Data bundle has not been loaded.")
return self.bundle
@property
def filters(self) -> FilterState:
if self.filter_state is None:
raise RuntimeError("Filter state has not been collected.")
return self.filter_state
|