File size: 26,450 Bytes
4dd67cd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 | """
GraphRAG Novelties Engine
==========================
Six cutting-edge techniques from 2024-2025 GraphRAG literature,
implemented as modular components that plug into the orchestration layer.
1. PPR Confidence-Weighted Retrieval (CatRAG, 2602.01965)
2. Graph Token Budget Controller (TERAG, 2509.18667)
3. Flow-Pruned Path Serializer (PathRAG, 2502.14902)
4. Spreading Activation Context Scorer (SA-RAG, 2512.15922)
5. PolyG Hybrid Retrieval Router (RAGRouter-Bench, 2602.00296)
6. Incremental Graph Updater (TG-RAG, 2510.13590)
"""
import logging
import math
import re
from collections import defaultdict
from typing import Any, Dict, List, Optional, Set, Tuple
logger = logging.getLogger(__name__)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 1. PPR CONFIDENCE-WEIGHTED RETRIEVAL
# Paper: CatRAG (2602.01965), HippoRAG 2 (2502.14802)
# Key idea: Personalized PageRank from seed entities with
# query-aware dynamic edge weights. PPR score = confidence.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class PPRConfidenceScorer:
"""
Runs Personalized PageRank from query-matched seed entities.
Each node's PPR score becomes its confidence weight for context ranking.
Formula (TG-RAG, 2510.13590):
s(edge) = s(v1) + s(v2) # sum of incident node PPR scores
s(chunk) = w(c) Γ Ξ£ s(edges_in_c) # weighted by query-chunk similarity
"""
def __init__(self, damping: float = 0.85, max_iterations: int = 20,
convergence: float = 1e-6):
self.damping = damping
self.max_iter = max_iterations
self.convergence = convergence
def compute_ppr(
self,
adjacency: Dict[str, List[Tuple[str, float]]], # node β [(neighbor, weight)]
seed_nodes: List[str],
seed_weights: Optional[Dict[str, float]] = None,
) -> Dict[str, float]:
"""
Compute Personalized PageRank scores.
Returns: {node_id: ppr_score}
"""
all_nodes = set(adjacency.keys())
for neighbors in adjacency.values():
for n, _ in neighbors:
all_nodes.add(n)
n = len(all_nodes)
if n == 0:
return {}
# Personalization vector (seed distribution)
personalization: Dict[str, float] = {}
if seed_weights:
total = sum(seed_weights.values())
personalization = {k: v / total for k, v in seed_weights.items()} if total > 0 else {}
else:
for s in seed_nodes:
personalization[s] = 1.0 / len(seed_nodes) if seed_nodes else 0
# Initialize scores
scores = {node: 1.0 / n for node in all_nodes}
# Power iteration
for iteration in range(self.max_iter):
new_scores: Dict[str, float] = {}
for node in all_nodes:
# Teleport component
teleport = personalization.get(node, 0.0)
# Random walk component
walk_score = 0.0
for source, neighbors in adjacency.items():
out_weight = sum(w for _, w in neighbors)
if out_weight > 0:
for target, weight in neighbors:
if target == node:
walk_score += scores[source] * (weight / out_weight)
new_scores[node] = (1 - self.damping) * teleport + self.damping * walk_score
# Check convergence
diff = sum(abs(new_scores[n] - scores[n]) for n in all_nodes)
scores = new_scores
if diff < self.convergence:
logger.debug(f"PPR converged at iteration {iteration + 1}")
break
return scores
def score_contexts(
self,
ppr_scores: Dict[str, float],
entity_to_chunks: Dict[str, List[str]], # entity_id β [chunk_ids]
chunk_texts: Dict[str, str],
query_similarity: Optional[Dict[str, float]] = None, # chunk_id β sim score
) -> List[Tuple[str, str, float]]:
"""
Score and rank chunks using PPR scores of their entities.
Returns: [(chunk_id, text, confidence_score)] sorted by score desc.
"""
chunk_scores: Dict[str, float] = defaultdict(float)
for entity_id, chunk_ids in entity_to_chunks.items():
entity_ppr = ppr_scores.get(entity_id, 0.0)
for cid in chunk_ids:
chunk_scores[cid] += entity_ppr
# Multiply by query-chunk similarity if available
if query_similarity:
for cid in chunk_scores:
sim = query_similarity.get(cid, 0.5)
chunk_scores[cid] *= (1 + sim)
results = [
(cid, chunk_texts.get(cid, ""), score)
for cid, score in chunk_scores.items()
if cid in chunk_texts
]
results.sort(key=lambda x: x[2], reverse=True)
return results
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 2. GRAPH TOKEN BUDGET CONTROLLER
# Paper: TERAG (2509.18667)
# Key idea: Cap context by token budget. Prioritize by
# concept_frequency Γ semantic_relevance. 97% reduction possible.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class TokenBudgetController:
"""
Controls the token budget for graph-retrieved context.
Prioritizes high-value content within a fixed token limit.
TERAG insight: 3-11% of LightRAG's token cost retains 80%+ accuracy.
"""
def __init__(self, max_tokens: int = 2000, chars_per_token: float = 4.0):
self.max_tokens = max_tokens
self.chars_per_token = chars_per_token
def estimate_tokens(self, text: str) -> int:
return max(1, int(len(text) / self.chars_per_token))
def prune_context(
self,
scored_items: List[Tuple[str, float]], # [(text, score)]
budget: Optional[int] = None,
) -> Tuple[List[str], Dict[str, Any]]:
"""
Select highest-scored items within token budget.
Returns: (selected_texts, stats)
"""
limit = budget or self.max_tokens
selected: List[str] = []
total_tokens = 0
total_available = sum(self.estimate_tokens(t) for t, _ in scored_items)
items_considered = 0
# Sort by score descending
sorted_items = sorted(scored_items, key=lambda x: x[1], reverse=True)
for text, score in sorted_items:
tokens = self.estimate_tokens(text)
items_considered += 1
if total_tokens + tokens <= limit:
selected.append(text)
total_tokens += tokens
elif total_tokens == 0:
# At least include one item (truncated)
truncated = text[:int(limit * self.chars_per_token)]
selected.append(truncated)
total_tokens = limit
break
stats = {
"budget_tokens": limit,
"used_tokens": total_tokens,
"utilization_pct": round(total_tokens / limit * 100, 1) if limit > 0 else 0,
"items_selected": len(selected),
"items_available": len(scored_items),
"tokens_saved": total_available - total_tokens,
"reduction_pct": round((1 - total_tokens / max(total_available, 1)) * 100, 1),
}
return selected, stats
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 3. FLOW-PRUNED PATH SERIALIZER
# Paper: PathRAG (2502.14902)
# Key idea: Retrieve key relational paths between entities,
# prune low-flow paths, serialize for LLM consumption.
# Exploits "lost-in-the-middle" by placing best paths first.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class PathPruner:
"""
Extracts and prunes reasoning paths between entities.
High-reliability paths placed FIRST in context (recency bias exploit).
PathRAG result: 62-65% win rate vs LightRAG in comprehensiveness.
"""
def find_paths(
self,
adjacency: Dict[str, List[Tuple[str, str, float]]], # node β [(neighbor, relation, weight)]
source: str,
target: str,
max_depth: int = 3,
max_paths: int = 5,
) -> List[List[Tuple[str, str, str]]]:
"""
Find top paths between source and target using DFS.
Returns: [[(entity, relation, next_entity), ...], ...]
"""
paths: List[List[Tuple[str, str, str]]] = []
def dfs(current: str, target: str, path: List[Tuple[str, str, str]],
visited: Set[str], depth: int):
if depth > max_depth or len(paths) >= max_paths * 3:
return
if current == target and path:
paths.append(list(path))
return
visited.add(current)
for neighbor, relation, weight in adjacency.get(current, []):
if neighbor not in visited:
path.append((current, relation, neighbor))
dfs(neighbor, target, path, visited, depth + 1)
path.pop()
visited.discard(current)
dfs(source, target, [], set(), 0)
return paths
def score_and_prune(
self,
paths: List[List[Tuple[str, str, str]]],
edge_weights: Dict[Tuple[str, str], float],
threshold: float = 0.1,
) -> List[Tuple[List[Tuple[str, str, str]], float]]:
"""
Score paths by accumulated edge weight, prune below threshold.
Returns: [(path, score)] sorted by score desc.
"""
scored = []
for path in paths:
score = 1.0
for src, rel, tgt in path:
w = edge_weights.get((src, tgt), edge_weights.get((tgt, src), 0.5))
score *= w
if score >= threshold:
scored.append((path, score))
scored.sort(key=lambda x: x[1], reverse=True)
return scored
def serialize_paths(
self,
scored_paths: List[Tuple[List[Tuple[str, str, str]], float]],
max_paths: int = 5,
) -> str:
"""
Serialize paths into LLM-friendly text.
HIGH-reliability paths placed FIRST (exploits lost-in-the-middle bias).
"""
lines = ["### Reasoning Paths (ranked by reliability):"]
for i, (path, score) in enumerate(scored_paths[:max_paths]):
chain = " β ".join(
[path[0][0]] + [f"--[{rel}]--> {tgt}" for _, rel, tgt in path]
)
lines.append(f" Path {i+1} (confidence: {score:.3f}): {chain}")
return "\n".join(lines)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 4. SPREADING ACTIVATION CONTEXT SCORER
# Paper: SA-RAG (2512.15922)
# Key idea: Activate seed nodes, propagate activation through
# graph edges with decay. Activation score = retrieval priority.
# Result: +39% answer correctness on MuSiQue.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class SpreadingActivation:
"""
Spreading Activation from seed entities through the knowledge graph.
Nodes with high activation are most relevant to the query.
"""
def __init__(self, decay_factor: float = 0.7, threshold: float = 0.01,
max_steps: int = 3):
self.decay = decay_factor
self.threshold = threshold
self.max_steps = max_steps
def activate(
self,
adjacency: Dict[str, List[Tuple[str, float]]],
seed_activations: Dict[str, float],
) -> Dict[str, float]:
"""
Spread activation from seeds through the graph.
Returns: {node_id: activation_level}
"""
activations = dict(seed_activations)
frontier = set(seed_activations.keys())
for step in range(self.max_steps):
new_activations: Dict[str, float] = {}
next_frontier: Set[str] = set()
for node in frontier:
current_activation = activations.get(node, 0.0)
if current_activation < self.threshold:
continue
for neighbor, weight in adjacency.get(node, []):
spread = current_activation * weight * self.decay
if spread >= self.threshold:
existing = new_activations.get(neighbor, 0.0)
new_activations[neighbor] = max(existing, spread)
next_frontier.add(neighbor)
# Merge new activations (keep max)
for node, act in new_activations.items():
activations[node] = max(activations.get(node, 0.0), act)
frontier = next_frontier
if not frontier:
break
return activations
def rank_contexts(
self,
activations: Dict[str, float],
entity_to_chunks: Dict[str, List[str]],
chunk_texts: Dict[str, str],
) -> List[Tuple[str, str, float]]:
"""Rank chunks by sum of their entities' activation levels."""
chunk_scores: Dict[str, float] = defaultdict(float)
for entity_id, chunk_ids in entity_to_chunks.items():
act = activations.get(entity_id, 0.0)
for cid in chunk_ids:
chunk_scores[cid] += act
results = [
(cid, chunk_texts.get(cid, ""), score)
for cid, score in chunk_scores.items()
if score > 0 and cid in chunk_texts
]
results.sort(key=lambda x: x[2], reverse=True)
return results
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 5. POLYG HYBRID RETRIEVAL ROUTER
# Papers: RAGRouter-Bench (2602.00296), PolyG (2504.02112)
# Key idea: 4-class query taxonomy determines retrieval strategy.
# No single paradigm wins everywhere β route adaptively.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class PolyGRouter:
"""
Enhanced hybrid router using PolyG's 4-class query taxonomy.
Routes queries to optimal retrieval strategy:
- entity_centric β Graph 1-hop lookup
- relation_lookup β Vector semantic search
- multi_hop β Graph traversal (PPR + paths)
- summarization β Community summaries
- hybrid β Both vector + graph (dual channel)
"""
# Regex patterns for query classification
ENTITY_PATTERNS = [
r"^(what|who|where) (is|are|was|were) ",
r"^tell me about ",
r"^describe ",
r"^define ",
]
RELATION_PATTERNS = [
r"(what|which) .* (did|does|do) .* (do|make|create|write|direct)",
r"(what|which) .* (position|role|job|title)",
r"how (did|does|do) .* (relate|connect)",
]
MULTI_HOP_PATTERNS = [
r"(same|both|compare|difference|which.*first|who.*born.*first)",
r"(what|who) .* (the|a) .* (that|which|who) ",
r"(capital|director|author|founder) .* (of|for) .* (the|a) .* (that|which)",
]
SUMMARIZATION_PATTERNS = [
r"^(summarize|overview|main themes|what are the)",
r"(overall|in general|broadly)",
]
def classify_query(self, query: str) -> Dict[str, Any]:
"""
Classify query into retrieval strategy.
Returns: {strategy, confidence, query_type, reasoning}
"""
q = query.lower().strip()
# Score each category
scores = {
"entity_centric": 0.0,
"relation_lookup": 0.0,
"multi_hop": 0.0,
"summarization": 0.0,
}
for pattern in self.ENTITY_PATTERNS:
if re.search(pattern, q):
scores["entity_centric"] += 0.4
for pattern in self.RELATION_PATTERNS:
if re.search(pattern, q):
scores["relation_lookup"] += 0.4
for pattern in self.MULTI_HOP_PATTERNS:
if re.search(pattern, q):
scores["multi_hop"] += 0.5
for pattern in self.SUMMARIZATION_PATTERNS:
if re.search(pattern, q):
scores["summarization"] += 0.4
# Structural signals
question_marks = q.count("?")
word_count = len(q.split())
has_comparison = any(w in q for w in ["same", "both", "compare", "difference", "versus", "vs"])
has_chain = any(w in q for w in ["that", "which", "who", "where", "whose"])
entity_count = sum(1 for word in q.split() if word[0:1].isupper()) if q else 0
if has_comparison:
scores["multi_hop"] += 0.3
if has_chain:
scores["multi_hop"] += 0.2
if word_count > 15:
scores["multi_hop"] += 0.1
if word_count < 8 and entity_count <= 1:
scores["entity_centric"] += 0.2
if entity_count >= 2:
scores["multi_hop"] += 0.15
# Determine winner
best_type = max(scores, key=scores.get) # type: ignore
best_score = scores[best_type]
# Map to strategy
strategy_map = {
"entity_centric": "graph_lookup",
"relation_lookup": "vector_search",
"multi_hop": "graph_traversal",
"summarization": "community_summary",
}
# If no strong signal, use hybrid
if best_score < 0.2:
strategy = "hybrid"
best_type = "ambiguous"
else:
strategy = strategy_map[best_type]
return {
"strategy": strategy,
"query_type": best_type,
"confidence": round(min(best_score, 1.0), 3),
"scores": {k: round(v, 3) for k, v in scores.items()},
"use_graph": strategy in ["graph_lookup", "graph_traversal", "hybrid"],
"use_vector": strategy in ["vector_search", "hybrid"],
"use_community": strategy == "community_summary",
"reasoning": f"Classified as '{best_type}' (score={best_score:.2f}) β {strategy}",
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 6. INCREMENTAL GRAPH UPDATER
# Papers: TG-RAG (2510.13590), LightRAG (2410.05779)
# Key idea: Add new documents without rebuilding the entire graph.
# Only recompute communities for affected subgraph.
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class IncrementalGraphUpdater:
"""
Supports incremental document ingestion without full graph rebuild.
New entities merge with existing by embedding similarity.
Community re-detection scoped to affected neighborhoods only.
"""
def __init__(self, merge_threshold: float = 0.85):
self.merge_threshold = merge_threshold
def find_merge_candidates(
self,
new_entity: Dict[str, Any], # {name, type, embedding}
existing_entities: List[Dict[str, Any]],
similarity_fn=None,
) -> Optional[str]:
"""
Find existing entity to merge with (deduplication).
Returns existing entity_id if merge candidate found, else None.
"""
if not similarity_fn:
from .graph_layer import cosine_similarity
similarity_fn = cosine_similarity
new_emb = new_entity.get("embedding", [])
if not new_emb:
return None
best_sim = 0.0
best_id = None
for existing in existing_entities:
existing_emb = existing.get("embedding", [])
if not existing_emb:
continue
sim = similarity_fn(new_emb, existing_emb)
if sim > best_sim:
best_sim = sim
best_id = existing.get("entity_id")
if best_sim >= self.merge_threshold and best_id:
logger.info(f"Merge: '{new_entity.get('name')}' β existing '{best_id}' (sim={best_sim:.3f})")
return best_id
return None
def compute_affected_scope(
self,
new_entity_ids: Set[str],
adjacency: Dict[str, List[str]],
scope_hops: int = 2,
) -> Set[str]:
"""
Find entities affected by new additions (for scoped community re-detection).
Returns set of entity_ids within scope_hops of new entities.
"""
affected = set(new_entity_ids)
frontier = set(new_entity_ids)
for _ in range(scope_hops):
next_frontier: Set[str] = set()
for node in frontier:
for neighbor in adjacency.get(node, []):
if neighbor not in affected:
affected.add(neighbor)
next_frontier.add(neighbor)
frontier = next_frontier
return affected
def plan_update(
self,
new_entities: List[Dict[str, Any]],
new_relations: List[Dict[str, Any]],
existing_entity_count: int,
) -> Dict[str, Any]:
"""
Plan the incremental update (for logging/dashboard display).
Returns update plan with estimated savings.
"""
return {
"new_entities": len(new_entities),
"new_relations": len(new_relations),
"existing_entities": existing_entity_count,
"merge_candidates_to_check": min(len(new_entities) * 10, existing_entity_count),
"community_redetection_scope": "affected_subgraph_only",
"estimated_llm_calls_saved": max(0, existing_entity_count - len(new_entities) * 2),
"vs_full_rebuild_savings_pct": round(
(1 - len(new_entities) / max(existing_entity_count, 1)) * 100, 1
) if existing_entity_count > 0 else 0,
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# NOVELTY ORCHESTRATOR β Combines all 6 techniques
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class NoveltyEngine:
"""
Orchestrates all 6 novelty techniques into a single pipeline.
Used by the main orchestration layer to enhance GraphRAG retrieval.
"""
def __init__(self, token_budget: int = 2000):
self.ppr = PPRConfidenceScorer()
self.budget = TokenBudgetController(max_tokens=token_budget)
self.paths = PathPruner()
self.activation = SpreadingActivation()
self.router = PolyGRouter()
self.updater = IncrementalGraphUpdater()
def enhanced_retrieve(
self,
query: str,
adjacency: Dict[str, List[Tuple[str, float]]],
seed_entities: List[str],
entity_to_chunks: Dict[str, List[str]],
chunk_texts: Dict[str, str],
seed_weights: Optional[Dict[str, float]] = None,
) -> Dict[str, Any]:
"""
Full novelty-enhanced retrieval pipeline:
1. Route query β determine strategy
2. PPR scoring from seeds
3. Spreading activation for expanded context
4. Token budget pruning
5. Return ranked, pruned context with metadata
"""
# Step 1: Route
routing = self.router.classify_query(query)
# Step 2: PPR
ppr_scores = self.ppr.compute_ppr(adjacency, seed_entities, seed_weights)
# Step 3: Spreading Activation
seed_acts = {s: 1.0 for s in seed_entities}
activations = self.activation.activate(adjacency, seed_acts)
# Step 4: Combined scoring (PPR + activation)
combined_chunks: Dict[str, float] = defaultdict(float)
for entity_id, chunk_ids in entity_to_chunks.items():
ppr_s = ppr_scores.get(entity_id, 0.0)
act_s = activations.get(entity_id, 0.0)
combined = 0.6 * ppr_s + 0.4 * act_s # weighted blend
for cid in chunk_ids:
combined_chunks[cid] = max(combined_chunks[cid], combined)
# Step 5: Token budget pruning
scored_items = [
(chunk_texts.get(cid, ""), score)
for cid, score in combined_chunks.items()
if cid in chunk_texts
]
selected_texts, budget_stats = self.budget.prune_context(scored_items)
return {
"contexts": selected_texts,
"routing": routing,
"budget_stats": budget_stats,
"ppr_top_entities": sorted(ppr_scores.items(), key=lambda x: x[1], reverse=True)[:10],
"activation_spread": len([v for v in activations.values() if v > 0.01]),
"technique_chain": [
f"PolyG Router β {routing['strategy']}",
f"PPR Scoring (damping={self.ppr.damping})",
f"Spreading Activation (decay={self.activation.decay})",
f"Token Budget ({budget_stats['used_tokens']}/{budget_stats['budget_tokens']} tokens)",
f"Reduction: {budget_stats['reduction_pct']}%",
],
}
|