Spaces:
Running
Running
File size: 16,856 Bytes
077e61e 9e4dc47 077e61e 9e4dc47 077e61e 9e4dc47 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 2a62ebe 077e61e 9e4dc47 077e61e 9e4dc47 077e61e b654948 077e61e b654948 077e61e b654948 077e61e 9e4dc47 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e 9e4dc47 077e61e b654948 077e61e 9e4dc47 077e61e b654948 077e61e b654948 077e61e b654948 077e61e 2a62ebe 077e61e 2a62ebe 077e61e 2a62ebe 077e61e 2a62ebe 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e b654948 077e61e 9e4dc47 077e61e b654948 077e61e b654948 077e61e b654948 077e61e 9e4dc47 077e61e 6c965a0 077e61e | 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 | """
Narada: Core environment logic (one instance per WebSocket session).
All state is per-instance. Never shared between sessions.
Graph is a module-level singleton loaded once at startup.
"""
from __future__ import annotations
import logging
import math
import uuid
from typing import Any, Dict, List, Optional, Set
from ..case_generator import MAX_STEPS, PatientCase, generate_case
from ..graph import NaradaGraph, get_graph
from ..models import (
NaradaAction,
NaradaObservation,
NaradaState,
GraphNode,
StepResult,
Variant,
)
logger = logging.getLogger(__name__)
# Validator requires returned scores strictly between 0 and 1 (exclusive).
# Internally we keep signed raw rewards so penalties stay meaningful.
_SCORE_MIN = 0.01
_SCORE_MAX = 0.99
_RAW_SCORE_SCALE = 0.45
def _clamp_score(value: float, default: float = 0.5) -> float:
if not math.isfinite(value):
return default
return float(max(_SCORE_MIN, min(_SCORE_MAX, value)))
def _to_score(raw_reward: float, default: float = 0.5) -> float:
"""Map signed raw reward to OpenEnv's required score interval."""
if not math.isfinite(raw_reward):
return default
return _clamp_score(0.5 + raw_reward * _RAW_SCORE_SCALE, default=default)
# ββ Step-level reward constants βββββββββββββββββββββββββββββββββββββββββββββββ
R_RELEVANT_HOP = 0.15
R_IRRELEVANT_HOP = -0.05
R_PER_STEP = -0.01
R_LAB_PENALTY = -0.10
R_BACKTRACK_RECOVERY = 0.05
R_TERMINAL_CORRECT = 1.0
R_TERMINAL_PARTIAL = 0.5 # non-terminal bonus per correct oligogenic flag
R_TERMINAL_WRONG = -0.5
R_TIMING_BONUS = 0.2 # correct flag before the tier's early-step cutoff
R_CAUSAL_GENE_MILESTONE = 0.10 # first-time visit to the actual causal gene node
OVERSEER_MIN = 0.0
OVERSEER_MAX = 0.3
class NaradaEnvironment:
"""
Stateful environment for one WebSocket session.
reset() β step()* β (done)
"""
def __init__(self) -> None:
self._graph: NaradaGraph = get_graph()
self._episode_id: str = ""
self._case: Optional[PatientCase] = None
self._step: int = 0
self._max_steps: int = 0
self._done: bool = False
# Navigation state
self._current_node_id: str = ""
self._trail: List[str] = [] # node IDs in visit order
self._trail_set: Set[str] = set()
# Reward tracking
self._step_rewards: List[float] = []
self._raw_cumulative_reward: float = 0.0
self._cumulative_reward: float = 0.0
# Flagging state
self._flagged_allele_ids: List[str] = []
# Overseer inputs
self._hallucinated_hops: int = 0 # hops to nodes not in graph edges
self._reasoning_log: List[str] = []
# ββ Public API ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def reset(self, task_type: Optional[str] = None, seed: Optional[int] = None) -> StepResult:
task_type = (task_type or "monogenic").strip().lower()
if task_type not in ("monogenic", "oligogenic", "phenotype_mismatch"):
raise ValueError(f"Unknown task_type: {task_type!r}")
self._episode_id = str(uuid.uuid4())
self._case = generate_case(self._graph, task_type, seed=seed)
self._step = 0
self._max_steps = MAX_STEPS[task_type]
self._done = False
self._current_node_id = self._case.starting_node_id
self._trail = [self._current_node_id]
self._trail_set = {self._current_node_id}
self._step_rewards = []
self._raw_cumulative_reward = 0.0
self._cumulative_reward = 0.0
self._flagged_allele_ids = []
self._hallucinated_hops = 0
self._reasoning_log = []
logger.info(
"Episode %s | task=%s disease=%s genes=%s",
self._episode_id[:8], task_type,
self._case.disease_name, self._case.causal_genes,
)
# Reset itself has no reward signal; use the neutral 0.5 score so the
# value stays strictly within the (0.01, 0.99) range OpenEnv requires
# for every StepResult.
neutral_score = _to_score(0.0)
obs = self._build_observation(step_reward=neutral_score)
return StepResult(
observation=obs,
reward=neutral_score,
done=False,
info={"episode_id": self._episode_id},
)
def step(self, action: NaradaAction) -> StepResult:
if self._done:
raise RuntimeError("Episode is done. Call reset() first.")
if self._case is None:
raise RuntimeError("Call reset() before step().")
self._step += 1
if action.reasoning:
self._reasoning_log.append(action.reasoning[:200])
# Route to action handler. Values are signed raw rewards until returned.
step_reward, terminal_reward, terminal = self._dispatch_action(action)
# Per-step efficiency penalty
step_reward += R_PER_STEP
# Terminal condition: timeout
if self._step >= self._max_steps and not terminal:
terminal = True
terminal_reward = self._compute_terminal_reward()
step_score = _to_score(step_reward)
self._step_rewards.append(step_score)
self._raw_cumulative_reward += step_reward
# Expose a true running mean of the per-step OpenEnv scores. This is
# what an agent can reason about monotonically; it is NOT the mapped
# sum of raw rewards (that would be misleading).
self._cumulative_reward = sum(self._step_rewards) / len(self._step_rewards)
if terminal:
# Only successful terminal outcomes receive overseer shaping.
# Wrong flags and timeouts must remain clearly worse than neutral.
overseer_score = self._overseer_score() if terminal_reward > 0 else 0.0
final_reward = _to_score(terminal_reward + overseer_score)
self._done = True
else:
final_reward = step_score
obs = self._build_observation(step_reward=step_score)
return StepResult(
observation=obs,
reward=final_reward if terminal else step_score,
done=self._done,
info={
"episode_id": self._episode_id,
"action_type": action.action_type,
"terminal": terminal,
},
)
def state(self) -> NaradaState:
return NaradaState(
episode_id=self._episode_id,
task_type=self._case.task_type if self._case else "unknown",
case_id=self._case.case_id if self._case else "",
step_count=self._step,
max_steps=self._max_steps,
cumulative_reward=self._cumulative_reward,
done=self._done,
flagged_variants=[f"VAR:{aid}" for aid in self._flagged_allele_ids],
ground_truth_variants=(
self._case.ground_truth_variant_ids
if self._case and self._done else []
),
)
# ββ Action dispatch βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _dispatch_action(
self, action: NaradaAction
) -> tuple[float, float, bool]:
"""Returns (step_reward, terminal_reward, is_terminal)."""
atype = action.action_type.lower()
if atype == "hop":
return self._action_hop(action.node_id or ""), 0.0, False
if atype == "flag_causal":
return self._action_flag(action.variant_id or "")
if atype == "request_lab":
return R_LAB_PENALTY, 0.0, False
if atype == "backtrack":
return self._action_backtrack(), 0.0, False
if atype == "summarise_trail":
return 0.0, 0.0, False # neutral, just informational
# Unknown action β treat as no-op with small penalty
return -0.02, 0.0, False
def _action_hop(self, target_node_id: str) -> float:
if not target_node_id:
return R_IRRELEVANT_HOP
neighbors = self._graph.get_neighbors(self._current_node_id)
# Hallucination check: node exists but is not connected
if target_node_id not in neighbors:
if target_node_id in self._graph.nodes:
self._hallucinated_hops += 1
return R_IRRELEVANT_HOP - 0.05
is_new_visit = target_node_id not in self._trail_set
self._current_node_id = target_node_id
if is_new_visit:
self._trail.append(target_node_id)
self._trail_set.add(target_node_id)
# Milestone: one-time bonus for first visit to the actual causal gene node.
# Gives the model an intermediate signal partway through phenotypeβgeneβvariant.
milestone = 0.0
if is_new_visit and any(
target_node_id == f"GENE:{g}" for g in self._case.causal_genes
):
milestone = R_CAUSAL_GENE_MILESTONE
is_relevant = target_node_id in self._case.relevant_node_ids
return (R_RELEVANT_HOP if is_relevant else R_IRRELEVANT_HOP) + milestone
def _action_backtrack(self) -> float:
if len(self._trail) < 2:
return R_IRRELEVANT_HOP
# Reward backtrack only if last hop was irrelevant
prev_node = self._current_node_id
self._trail.pop()
self._current_node_id = self._trail[-1]
was_irrelevant = prev_node not in self._case.relevant_node_ids
return R_BACKTRACK_RECOVERY if was_irrelevant else R_IRRELEVANT_HOP
def _action_flag(self, variant_id: str) -> tuple[float, float, bool]:
"""Returns (step_reward, terminal_reward, is_terminal)."""
case = self._case
# Parse allele_id from variant_id (format: "VAR:12345")
allele_id = variant_id.replace("VAR:", "").strip()
# Check if variant is in candidate list at all
candidate_ids = {v.allele_id for v in case.candidate_variants}
if allele_id not in candidate_ids:
return R_TERMINAL_WRONG, R_TERMINAL_WRONG, True
if allele_id in self._flagged_allele_ids:
return R_IRRELEVANT_HOP, 0.0, False
self._flagged_allele_ids.append(allele_id)
if case.task_type == "oligogenic":
ground_truth = set(case.causal_allele_ids)
flagged = set(self._flagged_allele_ids)
if allele_id not in ground_truth:
return R_TERMINAL_WRONG, R_TERMINAL_WRONG, True
if ground_truth.issubset(flagged):
terminal_reward = self._compute_terminal_reward()
return R_TERMINAL_PARTIAL, terminal_reward, True
progress_reward = R_TERMINAL_PARTIAL / max(1, len(ground_truth))
return progress_reward, 0.0, False
terminal_reward = self._compute_terminal_reward()
return terminal_reward * 0.1, terminal_reward, True
# ββ Terminal reward ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _compute_terminal_reward(self) -> float:
case = self._case
ground_truth = set(case.causal_allele_ids)
flagged = set(self._flagged_allele_ids)
if not flagged:
# Timed out without flagging β partial credit based on graph exploration
exploration_bonus = min(0.2, len(self._trail_set) / max(1, self._max_steps) * 0.25)
return -0.25 + exploration_bonus
# Check for decoy flag (phenotype_mismatch task)
decoy_gene = case.decoy_gene
if decoy_gene:
decoy_allele_ids = {
v["allele_id"]
for v in self._graph.get_variants_for_gene(decoy_gene)
}
if flagged & decoy_allele_ids:
return R_TERMINAL_WRONG # Flagged the decoy β maximum penalty
correct = flagged & ground_truth
wrong = flagged - ground_truth
if case.task_type == "monogenic":
if correct:
base = R_TERMINAL_CORRECT
if wrong:
base -= 0.3 * len(wrong)
timing_bonus = R_TIMING_BONUS if self._step < 10 else 0.0
return base + timing_bonus
return R_TERMINAL_WRONG
if case.task_type == "oligogenic":
n_correct = len(correct)
n_total = len(ground_truth)
# Scale oligogenic reward to the same 1.0 ceiling as monogenic so a
# fully correct diagnosis is not penalised by the tier. Partial
# credit remains linear in coverage.
coverage = (n_correct / n_total) if n_total > 0 else 0.0
partial = coverage * R_TERMINAL_CORRECT
wrong_penalty = 0.2 * len(wrong)
timing_bonus = R_TIMING_BONUS if (self._step < 15 and n_correct == n_total) else 0.0
return partial - wrong_penalty + timing_bonus
if case.task_type == "phenotype_mismatch":
if correct:
timing_bonus = R_TIMING_BONUS if self._step < 12 else 0.0
return R_TERMINAL_CORRECT + timing_bonus
return R_TERMINAL_WRONG
return 0.0
# ββ Overseer ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _overseer_score(self) -> float:
"""
Additive quality score 0.0β0.3 for the Overseer agent.
Computed locally (no LLM call in this implementation).
Full Overseer LLM call can be added in inference.py.
"""
score = OVERSEER_MAX
# Penalise hallucinated hops.
score -= self._hallucinated_hops * 0.05
# Penalise very short exploration (< 3 unique nodes = no reasoning).
unique_visited = len(self._trail_set)
if unique_visited < 3:
score -= 0.1
# Reward visiting each causal gene (oligogenic rewards both, capped).
case = self._case
gene_bonus = 0.0
for gene in case.causal_genes:
if f"GENE:{gene}" in self._trail_set:
gene_bonus += 0.05
score += min(0.10, gene_bonus)
return min(OVERSEER_MAX, max(OVERSEER_MIN, score))
# ββ Observation builder βββββββββββββββββββββββββββββββββββββββββββββββββββ
def _build_observation(self, step_reward: float) -> NaradaObservation:
case = self._case
graph = self._graph
current_node = self._node_to_model(self._current_node_id)
trail_nodes = [self._node_to_model(nid) for nid in self._trail[-10:]] # last 10
info: Dict[str, Any] = {
"task_type": case.task_type,
"episode_id": self._episode_id,
"flagged_variants": [f"VAR:{aid}" for aid in self._flagged_allele_ids],
}
if self._done:
info["disease_name"] = case.disease_name
info["ground_truth_hint"] = case.causal_genes # revealed post-episode
info["ground_truth_variants"] = case.ground_truth_variant_ids
return NaradaObservation(
step=self._step,
max_steps=self._max_steps,
task_type=case.task_type,
current_node=current_node,
trail=trail_nodes,
patient_phenotypes=case.patient_hpo_ids,
phenotype_names=case.patient_phenotype_names,
phenotypes_absent=case.absent_hpo_ids,
phenotype_absent_names=case.absent_phenotype_names,
candidate_variants=case.candidate_variants,
step_reward=round(step_reward, 4),
cumulative_reward=round(self._cumulative_reward, 4),
done=self._done,
info=info,
)
def _node_to_model(self, node_id: str) -> GraphNode:
nd = self._graph.get_node(node_id)
if nd is None:
return GraphNode(
id=node_id, type="unknown", name=node_id,
description="", connected_node_ids=[],
)
neighbors = self._graph.get_neighbors(node_id)
return GraphNode(
id=nd["id"],
type=nd["type"],
name=nd["name"],
description=nd["description"],
connected_node_ids=neighbors[:30],
metadata=nd["metadata"],
)
|