Spaces:
Running
Running
File size: 11,181 Bytes
f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 | from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Iterable
from .schemas import PipelineConfig, QueryPlan, SearchHit
@dataclass(slots=True)
class ContextBudgetDecision:
hits: list[SearchHit]
used: bool
reason: str
policy: str
target_chunks: int
corpus_scale: str
retrieval_confidence: float
score_gap: float
chunks_before: int
chunks_after: int
sources_before: int
sources_after: int
chars_before: int
chars_after: int
tokens_est_before: int
tokens_est_after: int
@property
def reduction_ratio(self) -> float:
if self.chars_before <= 0:
return 0.0
return max(0.0, min(1.0, 1.0 - (self.chars_after / self.chars_before)))
def trace_fields(self) -> dict[str, object]:
return {
"context_pruning_used": self.used,
"context_pruning_reason": self.reason,
"context_budget_policy": self.policy,
"context_budget_target_chunks": self.target_chunks,
"corpus_scale": self.corpus_scale,
"retrieval_confidence": round(self.retrieval_confidence, 3),
"retrieval_score_gap": round(self.score_gap, 3),
"context_chunks_before": self.chunks_before,
"context_chunks_after": self.chunks_after,
"context_sources_before": self.sources_before,
"context_sources_after": self.sources_after,
"context_chars_before": self.chars_before,
"context_chars_after": self.chars_after,
"context_tokens_est_before": self.tokens_est_before,
"context_tokens_est_after": self.tokens_est_after,
"context_reduction_pct": round(self.reduction_ratio * 100.0, 1),
}
def estimate_tokens_from_chars(chars: int) -> int:
"""Cheap deterministic token estimate for observability, not billing."""
return int(math.ceil(max(0, int(chars)) / 4.0))
def corpus_scale_label(chunks: int, sources: int) -> str:
chunks = max(0, int(chunks))
sources = max(0, int(sources))
if chunks >= 3000 or sources >= 80:
return "very_large"
if chunks >= 1000 or sources >= 30:
return "large"
if chunks >= 250 or sources >= 10:
return "medium"
return "small"
def adaptive_retrieval_top_k(
config: PipelineConfig,
plan: QueryPlan,
*,
corpus_chunks: int,
corpus_sources: int,
) -> int:
"""Choose initial retrieval breadth from corpus scale and task breadth.
``top_k`` remains the small-corpus baseline configured by the user. Larger
corpora can retrieve a wider candidate set before the focused context
budget trims the generation context. Broad tasks retain wider evidence.
"""
baseline = max(2, min(12, int(config.top_k)))
if not getattr(config, "use_adaptive_top_k", True):
return baseline
scale = corpus_scale_label(corpus_chunks, corpus_sources)
scale_floor = {"small": baseline, "medium": max(baseline, 8), "large": max(baseline, 10), "very_large": 12}[scale]
if plan.task_type in {"overview", "insight_synthesis", "cross_document_synthesis", "comparison"}:
# Broad tasks need source breadth. Keep the absolute cap bounded by the
# schema and avoid forcing more chunks than sources can usefully supply.
broad_floor = min(12, max(scale_floor, min(max(6, corpus_sources), 12)))
return broad_floor
return min(12, scale_floor)
def _context_chars(hits: Iterable[SearchHit]) -> int:
# Include a small deterministic metadata allowance per chunk because the
# generation context contains source/page labels in addition to raw text.
return sum(len(hit.chunk.text or "") + len(hit.chunk.source or "") + 24 for hit in hits)
def _source_count(hits: Iterable[SearchHit]) -> int:
return len({hit.chunk.source for hit in hits})
def _hit_signal(hit: SearchHit) -> float:
dense = max(0.0, min(1.0, float(hit.dense_score or 0.0)))
sparse = max(0.0, min(1.0, float(hit.sparse_score or 0.0))) * 0.9
fused = max(0.0, min(1.0, float(hit.score or 0.0)))
return max(dense, sparse, fused)
def _retrieval_confidence(hits: list[SearchHit]) -> tuple[float, float]:
if not hits:
return 0.0, 0.0
signals = [_hit_signal(hit) for hit in hits[:4]]
top = signals[0]
second = signals[1] if len(signals) > 1 else 0.0
gap = max(0.0, top - second)
# Agreement between dense and sparse on the first result adds confidence,
# while a small top-two gap indicates ambiguity even with a strong top hit.
first = hits[0]
method_agreement = float(
float(first.dense_score or 0.0) >= 0.20 and float(first.sparse_score or 0.0) >= 0.05
)
confidence = 0.65 * top + 0.25 * min(1.0, gap / 0.20) + 0.10 * method_agreement
return max(0.0, min(1.0, confidence)), gap
def adaptive_context_budget(
hits: list[SearchHit],
plan: QueryPlan,
config: PipelineConfig,
*,
corpus_chunks: int = 0,
corpus_sources: int = 0,
) -> ContextBudgetDecision:
"""Adaptively shrink context for focused local lookups only.
v1.8 proved that a three-chunk safety floor could cut the demo context by
roughly half without harming source recall. v1.9 generalizes that policy:
the target is chosen from retrieval confidence, score separation and corpus
scale. Broad/synthesis tasks are never pruned by this function.
"""
before = list(hits or [])
chunks_before = len(before)
chars_before = _context_chars(before)
sources_before = _source_count(before)
scale = corpus_scale_label(corpus_chunks or chunks_before, corpus_sources or sources_before)
retrieval_confidence, score_gap = _retrieval_confidence(before)
def decision(
after: list[SearchHit],
used: bool,
reason: str,
*,
target: int,
policy: str = "adaptive_focused_budget",
) -> ContextBudgetDecision:
chars_after = _context_chars(after)
return ContextBudgetDecision(
hits=after,
used=used,
reason=reason,
policy=policy,
target_chunks=max(0, int(target)),
corpus_scale=scale,
retrieval_confidence=retrieval_confidence,
score_gap=score_gap,
chunks_before=chunks_before,
chunks_after=len(after),
sources_before=sources_before,
sources_after=_source_count(after),
chars_before=chars_before,
chars_after=chars_after,
tokens_est_before=estimate_tokens_from_chars(chars_before),
tokens_est_after=estimate_tokens_from_chars(chars_after),
)
if not before:
return decision(before, False, "no_document_evidence", target=0)
if not getattr(config, "use_context_pruning", True):
return decision(before, False, "disabled_by_user", target=chunks_before, policy="disabled")
if plan.route not in {"documents", "hybrid"}:
return decision(before, False, "non_local_route", target=chunks_before)
if plan.task_type not in {"fact_lookup", "followup"}:
return decision(before, False, "broad_or_multi_source_task", target=chunks_before)
if plan.retrieval_strategy not in {"semantic", "hierarchical"}:
return decision(before, False, "strategy_requires_breadth", target=chunks_before)
# Base safety grows with corpus scale because a larger corpus increases the
# chance that a seemingly focused question needs a runner-up source.
base = {"small": 3, "medium": 3, "large": 4, "very_large": 5}[scale]
target = base
# A clearly separated high-confidence top result can safely use two chunks
# on small/medium corpora. Avoid top-1 pruning so a second supporting chunk
# remains available for citation and accidental under-classification.
if scale in {"small", "medium"} and retrieval_confidence >= 0.78 and score_gap >= 0.15:
target = 2
# Ambiguous retrieval needs extra evidence, especially when the top results
# already span multiple sources.
top_sources = len({hit.chunk.source for hit in before[:3]})
if retrieval_confidence < 0.55 or score_gap < 0.035 or top_sources >= 3:
target = min(5, max(target, base + 1))
if plan.task_type == "followup":
target = min(5, max(target, 3))
# The initial adaptive retrieval depth may exceed config.top_k on a large
# corpus, so cap against the actual candidate count rather than config.top_k.
target = max(2, min(target, chunks_before))
if chunks_before <= target:
return decision(before, False, "already_within_adaptive_budget", target=target)
after = before[:target]
reason = f"adaptive_{scale}_top{target}"
return decision(after, True, reason, target=target)
def focused_context_budget(
hits: list[SearchHit],
plan: QueryPlan,
config: PipelineConfig,
*,
safety_floor: int = 3,
) -> ContextBudgetDecision:
"""Backward-compatible v1.8 fixed safety-floor policy.
The runtime uses :func:`adaptive_context_budget` in v1.9. Keeping this
helper preserves the documented v1.8 ablation semantics and makes release
deltas auditable instead of silently changing an old experiment.
"""
before = list(hits or [])
chunks_before = len(before)
chars_before = _context_chars(before)
sources_before = _source_count(before)
confidence, gap = _retrieval_confidence(before)
def decision(after: list[SearchHit], used: bool, reason: str) -> ContextBudgetDecision:
chars_after = _context_chars(after)
return ContextBudgetDecision(
hits=after,
used=used,
reason=reason,
policy="v1.8_fixed_safety_floor",
target_chunks=len(after),
corpus_scale=corpus_scale_label(chunks_before, sources_before),
retrieval_confidence=confidence,
score_gap=gap,
chunks_before=chunks_before,
chunks_after=len(after),
sources_before=sources_before,
sources_after=_source_count(after),
chars_before=chars_before,
chars_after=chars_after,
tokens_est_before=estimate_tokens_from_chars(chars_before),
tokens_est_after=estimate_tokens_from_chars(chars_after),
)
if not before:
return decision(before, False, "no_document_evidence")
if not getattr(config, "use_context_pruning", True):
return decision(before, False, "disabled_by_user")
if plan.route not in {"documents", "hybrid"}:
return decision(before, False, "non_local_route")
if plan.task_type not in {"fact_lookup", "followup"}:
return decision(before, False, "broad_or_multi_source_task")
if plan.retrieval_strategy not in {"semantic", "hierarchical"}:
return decision(before, False, "strategy_requires_breadth")
target = max(2, min(int(safety_floor), int(config.top_k)))
if chunks_before <= target:
return decision(before, False, "already_within_budget")
return decision(before[:target], True, "focused_lookup_top3_safety_floor")
|