File size: 36,404 Bytes
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b401c21
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
af29724
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
 
 
 
 
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
f23deb1
 
 
 
 
ac224ce
 
 
 
 
 
f23deb1
 
ac224ce
f23deb1
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
b401c21
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
 
 
 
ac224ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
 
ac224ce
f23deb1
ac224ce
f23deb1
 
 
 
 
ac224ce
 
 
f23deb1
 
 
 
ac224ce
 
 
 
 
 
 
 
f23deb1
 
 
ac224ce
 
 
f23deb1
 
 
 
ac224ce
 
 
f23deb1
 
 
ac224ce
 
 
f23deb1
 
 
 
 
 
ac224ce
 
 
 
f23deb1
ac224ce
f23deb1
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
ac224ce
 
 
 
 
f23deb1
ac224ce
 
 
 
 
f23deb1
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
ac224ce
 
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
ac224ce
 
 
 
 
 
 
 
 
 
 
f23deb1
ac224ce
 
 
f23deb1
ac224ce
 
f23deb1
ac224ce
f23deb1
ac224ce
f23deb1
 
 
ac224ce
 
 
f23deb1
ac224ce
 
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f23deb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac224ce
f23deb1
 
 
 
 
 
 
 
 
 
 
ac224ce
 
 
af29724
f23deb1
 
 
 
 
 
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
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
"""
server/rag_debug_env_environment.py
------------------------------------
Real RAGDebugEnvironment implementation.

Replaces the echo stub with a full RL environment for training agents to
diagnose and fix broken RAG pipelines.

Architecture summary
--------------------
- Each episode samples N queries from one domain's corpus.
- S_true_{model}.npy matrices (precomputed by Stage 5) hold the ground-truth
  cosine similarity scores for every (query, chunk) pair.
- Faults are applied as mathematical transformations: S_faulted = f(S_true, config, faults).
- The agent's job is to modify PipelineConfig (and/or query rewrites) until
  the retrieval simulationβ€”run on S_faultedβ€”recovers adequate coverage of R*.
- Every config change triggers _recompute_S_faulted() so the fault math can
  modulate its severity based on the new config values.
- Noise arrays are pre-generated at reset() time for determinism across
  recomputation calls within a single episode.
"""

from __future__ import annotations

import json
import warnings
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
from uuid import uuid4

import numpy as np
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import EnvironmentMetadata, State
from pydantic import ValidationError

from server.constants import (
    _TASK_DOMAIN,
    _TASK_DESCRIPTION,
    _N_EPISODE_QUERIES,
    _MAX_STEPS,
    _MODEL_FILE,
    _TASK1_FAULT_SETS,
    _TASK2_FAULT_SETS,
    _TASK3_FAULTS,
)
from server.corpus import _load_corpus
from server.fault_math import apply_faults
from models import (
    RAGDebugAction,
    RAGDebugObservation,
    ActionType,
    EmbeddingModel,
    FaultType,
    FaultConfig,
    InternalState,
    PipelineConfig,
    QueryResult,
    QualityMetrics,
    CorpusStats,
    Domain,
    Reward,
)

class RAGDebugEnvironment(Environment):
    """
    RL environment for diagnosing and fixing broken RAG pipelines.

    Each episode samples a small query set from one domain, injects faults
    into the similarity matrix, and rewards the agent for recovering
    retrieval quality through pipeline config changes.

    Tasks
    -----
    Task 1 (software): One or two config faults. Success threshold β‰₯ 0.80.
    Task 2 (climate):  Compound config faults. Success threshold β‰₯ 0.75.
    Task 3 (medical):  Wrong embedding model + config faults + multi-hop.
                       Success threshold β‰₯ 0.70 AND multi_hop_coverage > 0.60.
    """

    SUPPORTS_CONCURRENT_SESSIONS: bool = True

    def __init__(self):
        super().__init__()
        self._state = State(episode_id=str(uuid4()), step_count=0)

        # Episode-level data, set during reset()
        self._task_id: int = 1
        self._domain: str = "software"
        self._chunks: List[Dict] = []
        self._episode_queries: List[Dict] = []
        self._ground_truth: Dict[str, List[int]] = {}
        self._corpus_stats_dict: Dict = {}

        # Per-episode numpy data
        self._chunk_ids: List[int] = []
        self._chunk_tokens: List[int] = []
        self._chunk_id_to_tokens: Dict[int, int] = {}  # O(1) token lookup
        self._s_true_episode: Dict[str, np.ndarray] = {}  # model_name β†’ (n_q, n_c) float32
        self._active_model: EmbeddingModel = EmbeddingModel.GENERAL

        # Deterministic noise (pre-generated at reset, scaled during recompute)
        self._noise: Dict[FaultType, np.ndarray] = {}
        self._dupe_ids: np.ndarray = np.array([], dtype=int)

        # REWRITE_QUERY persistent overlay
        self._rewrite_boosts: np.ndarray = np.zeros((0, 0), dtype=np.float32)

        # Current faulted matrix (rebuilt by _recompute_S_faulted)
        self._S_faulted: np.ndarray = np.zeros((0, 0), dtype=np.float32)

        # Episode config & state
        self._config: PipelineConfig = PipelineConfig()
        self._injected_faults: List[FaultConfig] = []
        self._internal_state: InternalState = InternalState(
            injected_faults=[], episode_seed=0
        )
        self._prev_metrics: Optional[QualityMetrics] = None
        self._prev_action_type: Optional[ActionType] = None
        self._done: bool = False
        self._last_action_error: Optional[str] = None
        self._last_reward_components: Dict[str, float] = {}

    # ------------------------------------------------------------------
    # OpenEnv required interface
    # ------------------------------------------------------------------

    def reset(
        self,
        seed: Optional[int] = None,
        episode_id: Optional[str] = None,
        **kwargs: Any,
    ) -> RAGDebugObservation:
        """Reset the environment and return the initial observation."""
        # Inherited from openenv.core.env_server.interfaces.Environment.
        # Clears any OpenEnv framework-level episode state (e.g. step counters
        # tracked by the base class) before we re-initialise domain-specific state.
        self._reset_rubric()

        # Task selection
        task_id = int(kwargs.get("task_id", 1))
        if task_id not in (1, 2, 3):
            raise ValueError(f"task_id must be 1, 2, or 3; got {task_id}")
        self._task_id = task_id

        # RNG
        if seed is None:
            seed = int(np.random.default_rng().integers(0, 2**31))
        rng = np.random.default_rng(seed)

        # State bookkeeping
        ep_id = episode_id or str(uuid4())
        self._state = State(episode_id=ep_id, step_count=0)
        self._done = False
        self._prev_action_type = None
        self._last_action_error = None
        self._last_reward_components = {}

        # Domain & corpus
        self._domain = _TASK_DOMAIN[task_id].value
        corpus = _load_corpus(self._domain)
        self._chunks = corpus["chunks"]
        self._ground_truth = corpus["ground_truth"]
        self._corpus_stats_dict = corpus["corpus_stats"]

        all_queries: List[Dict] = corpus["queries"]

        # Sample episode queries
        self._episode_queries = self._sample_queries(all_queries, task_id, rng)

        # Build episode index structures
        self._chunk_ids = [c["chunk_id"] for c in self._chunks]
        self._chunk_tokens = [c.get("n_tokens", 100) for c in self._chunks]
        self._chunk_id_to_tokens = {
            c["chunk_id"]: c.get("n_tokens", 100) for c in self._chunks
        }
        n_q = len(self._episode_queries)
        n_c = len(self._chunks)

        # Slice S_true to episode query rows
        # Full S_true shape: (all_queries, n_chunks); query_id == row index in S_true
        self._s_true_episode = {}
        all_query_ids = [q["query_id"] for q in all_queries]
        ep_query_ids = [q["query_id"] for q in self._episode_queries]
        # Build index map: query_id β†’ row in the full S_true matrix
        qid_to_row = {qid: i for i, qid in enumerate(all_query_ids)}
        ep_rows = [qid_to_row[qid] for qid in ep_query_ids]

        s_true_full = corpus["s_true"]
        for model_name, s_full in s_true_full.items():
            self._s_true_episode[model_name] = s_full[ep_rows, :].copy()

        # Pre-generate noise (unit normal; scaled during recompute)
        shape = (n_q, n_c)
        self._noise = {
            FaultType.CHUNK_TOO_SMALL:   rng.standard_normal(shape).astype(np.float32),
            FaultType.THRESHOLD_TOO_LOW: rng.standard_normal(shape).astype(np.float32),
            FaultType.NO_RERANKING:      rng.standard_normal(shape).astype(np.float32),
        }
        self._dupe_ids = rng.choice(n_c, size=max(1, n_c // 7), replace=False)

        # Reset overlay and config
        self._rewrite_boosts = np.zeros(shape, dtype=np.float32)
        self._config = PipelineConfig()

        # Task 3 starts with the LEGAL embedding model (wrong model for medical text).
        # S_true_legal has ~0.62 coverage vs ~0.90 for GENERAL on medical data.
        # The agent must discover this and swap to GENERAL or MEDICAL.
        self._active_model = EmbeddingModel.LEGAL if task_id == 3 else EmbeddingModel.GENERAL
        self._config = self._config.model_copy(update={"embedding_model": self._active_model})

        # Start from a mildly constrained baseline so most episodes leave
        # headroom for meaningful improvement steps.
        self._config = self._config.model_copy(
            update={
                "top_k": int(rng.integers(5, 9)),
                "similarity_threshold": float(rng.uniform(0.34, 0.48)),
            }
        )

        # Sample and inject faults
        self._injected_faults = self._sample_faults(task_id, rng)
        self._internal_state = InternalState(
            injected_faults=self._injected_faults,
            episode_seed=seed,
        )

        # Fault-specific initial config: some faults only degrade coverage when
        # the related parameter starts at a bad value.
        # TOP_K_TOO_SMALL: score compression preserves rank order, so coverage
        #   stays high at top_k=10. Start with a small top_k so the agent must
        #   increase it to recover coverage.
        # DUPLICATE_FLOODING: flooded chunks can't displace high-scoring relevant
        #   chunks in a top_k=10 pool. Start with reduced top_k so the flooding
        #   actually crowds out relevant chunks.
        fault_types_active = {f.fault_type for f in self._injected_faults}
        if FaultType.TOP_K_TOO_SMALL in fault_types_active:
            self._config = self._config.model_copy(update={"top_k": int(rng.integers(2, 4))})
        elif FaultType.DUPLICATE_FLOODING in fault_types_active:
            self._config = self._config.model_copy(update={"top_k": int(rng.integers(4, 8))})

        # If reset still lands in a very strong state, add one extra nudge
        # so the agent usually has room to improve from step 1.
        self._calibrate_initial_difficulty(rng)

        # Initial matrix + metrics
        self._recompute_S_faulted()
        initial_results = self._simulate_retrieval()
        self._prev_metrics = self._compute_metrics(initial_results)

        return self._build_observation(initial_results, reward=None)

    def step(
        self,
        action: RAGDebugAction,
        timeout_s: Optional[float] = None,
        **kwargs: Any,
    ) -> RAGDebugObservation:
        """Execute one agent action and return the updated observation."""
        if self._done:
            raise RuntimeError("Episode is already done. Call reset() to start a new episode.")

        self._state.step_count += 1
        prev_metrics = self._prev_metrics

        # Route action (returns Reward for SUBMIT, None otherwise)
        reward_obj = self._apply_action(action)

        # Recompute retrieval results and metrics
        new_results = self._simulate_retrieval()
        new_metrics = self._compute_metrics(new_results)

        # Compute reward if not already set by SUBMIT handler
        if reward_obj is None:
            reward_obj = self._compute_reward(prev_metrics, new_metrics, action)

        # Keep per-step reward components available in observations.
        self._last_reward_components = dict(reward_obj.components)

        self._internal_state.action_history.append(action)
        self._internal_state.reward_history.append(reward_obj.value)

        self._prev_metrics = new_metrics
        self._prev_action_type = action.action_type

        # Auto-terminate on max steps
        if self._state.step_count >= _MAX_STEPS and not self._done:
            self._done = True

        return self._build_observation(new_results, reward=reward_obj.value)

    @property
    def state(self) -> State:
        return self._state

    def get_metadata(self) -> EnvironmentMetadata:
        readme_path = Path(__file__).parent.parent / "README.md"
        readme_content: Optional[str] = None
        if readme_path.exists():
            raw = readme_path.read_text(encoding="utf-8")
            # Strip YAML frontmatter (--- ... ---) so the UI renders clean Markdown
            if raw.startswith("---"):
                end = raw.find("---", 3)
                if end != -1:
                    raw = raw[end + 3:].lstrip("\n")
            readme_content = raw
        return EnvironmentMetadata(
            name="RAGDebugEnv",
            description="Debug broken RAG pipelines by tuning config and swapping embedding models.",
            readme_content=readme_content,
            version="1.0.0",
        )

    # ------------------------------------------------------------------
    # Action routing
    # ------------------------------------------------------------------

    def _update_config(self, **updates) -> Optional[str]:
        """
        Apply updates to PipelineConfig using the constructor, which runs all
        Pydantic validators (including model_validators).

        model_copy(update=...) does NOT run validators in Pydantic v2, so
        invalid combinations (e.g. chunk_overlap >= chunk_size) are silently
        accepted and later cause crashes when the config is embedded in an
        observation.  Using the constructor guarantees validation always runs.

        Returns an error string if validation fails, None on success.
        """
        try:
            self._config = PipelineConfig(**{**self._config.model_dump(), **updates})
            return None
        except (ValueError, TypeError, ValidationError) as exc:
            return str(exc)

    def _apply_action(self, action: RAGDebugAction) -> Optional[Reward]:
        """
        Apply action to config / overlays.
        Returns a Reward if the action is SUBMIT, else None.
        """
        self._last_action_error = None
        t = action.action_type
        p = action.params

        if t == ActionType.ADJUST_CHUNK_SIZE:
            value = int(p.get("value", self._config.chunk_size))
            err = self._update_config(chunk_size=value)
            if err:
                self._last_action_error = f"Invalid chunk_size {value}: {err}"
            self._recompute_S_faulted()

        elif t == ActionType.ADJUST_CHUNK_OVERLAP:
            value = int(p.get("value", self._config.chunk_overlap))
            err = self._update_config(chunk_overlap=value)
            if err:
                self._last_action_error = f"Invalid chunk_overlap {value}: {err}"
            # Recompute required: fault_math.apply_faults() uses chunk_overlap to
            # modulate CHUNK_TOO_SMALL noise sigma (higher overlap stabilises boundary
            # embeddings, reducing noise severity). Without recompute, the config change
            # has no visible effect on retrieval scores until another action triggers it.
            self._recompute_S_faulted()

        elif t == ActionType.ADJUST_THRESHOLD:
            value = float(p.get("value", self._config.similarity_threshold))
            err = self._update_config(similarity_threshold=value)
            if err:
                self._last_action_error = f"Invalid threshold {value}: {err}"
            # Threshold is applied in retrieval simulation; no matrix recompute

        elif t == ActionType.ADJUST_TOP_K:
            value = int(p.get("value", self._config.top_k))
            err = self._update_config(top_k=value)
            if err:
                self._last_action_error = f"Invalid top_k {value}: {err}"

        elif t == ActionType.ADJUST_CONTEXT_LIMIT:
            value = int(p.get("value", self._config.context_window_limit))
            err = self._update_config(context_window_limit=value)
            if err:
                self._last_action_error = f"Invalid context_limit {value}: {err}"
            self._recompute_S_faulted()

        elif t == ActionType.SWAP_EMBEDDING_MODEL:
            model_str = str(p.get("model", self._active_model.value))
            try:
                new_model = EmbeddingModel(model_str)
            except (ValueError, KeyError) as exc:
                self._last_action_error = f"Invalid embedding model '{model_str}': {exc}"
                new_model = None
            if new_model is not None:
                self._active_model = new_model
                self._update_config(embedding_model=new_model)
            self._recompute_S_faulted()

        elif t == ActionType.TOGGLE_RERANKING:
            enabled = bool(p.get("enabled", not self._config.use_reranking))
            self._update_config(use_reranking=enabled)
            self._recompute_S_faulted()

        elif t == ActionType.REWRITE_QUERY:
            query_id = p.get("query_id")
            ep_ids = [q["query_id"] for q in self._episode_queries]
            if query_id in ep_ids:
                row = ep_ids.index(query_id)
                r_star = self._ground_truth.get(str(query_id), [])
                # Map chunk_id β†’ column index
                cid_to_col = {cid: col for col, cid in enumerate(self._chunk_ids)}
                cols = [cid_to_col[cid] for cid in r_star if cid in cid_to_col]
                if cols:
                    self._rewrite_boosts[row, cols] = 0.20
            self._recompute_S_faulted()

        elif t == ActionType.SUBMIT:
            # Compute final metrics at SUBMIT time for accurate grading.
            results = self._simulate_retrieval()
            metrics = self._compute_metrics(results)
            task_score = self._compute_task_score(metrics)
            self._done = True

            success = self._check_success(metrics, task_score)
            if success:
                terminal_value = float(np.clip(0.7 + 0.3 * task_score, 0.7, 1.0))
                return Reward(
                    value=terminal_value,
                    components={"terminal_success": terminal_value},
                )
            else:
                terminal_value = float(np.clip(0.2 * task_score, 0.0, 0.2))
                return Reward(
                    value=terminal_value,
                    components={"terminal_failure": terminal_value},
                )

        return None  # signal caller to compute reward via _compute_reward

    # ------------------------------------------------------------------
    # Fault math
    # ------------------------------------------------------------------

    def _recompute_S_faulted(self) -> None:
        """
        Apply all active faults to the S_true matrix for the current active model.

        WRONG_EMBEDDING_MODEL is implicit: Task 3 starts with active_model=LEGAL,
        whose score distribution on medical text is fundamentally wrong (compressed,
        meanβ‰ˆ0.84, stdβ‰ˆ0.033 vs GENERAL stdβ‰ˆ0.13). The agent must diagnose this
        from retrieval score distributions and swap models. All other faults are
        applied as matrix transformations by apply_faults().
        """
        # Lock to GENERAL on tasks without WRONG_EMBEDDING_MODEL so that
        # swap_embedding_model actions don't accidentally shift coverage/precision
        # via raw score-level differences between embedding model matrices.
        fault_types = {f.fault_type for f in self._injected_faults}
        if FaultType.WRONG_EMBEDDING_MODEL in fault_types:
            model_key = _MODEL_FILE[self._active_model]
        else:
            model_key = "general"
        S = self._s_true_episode.get(model_key)
        if S is None:
            S = self._s_true_episode["general"]

        self._S_faulted = apply_faults(
            S=S,
            fault_types=fault_types,
            config_chunk_size=self._config.chunk_size,
            config_context_limit=self._config.context_window_limit,
            config_use_reranking=self._config.use_reranking,
            config_chunk_overlap=self._config.chunk_overlap,
            noise=self._noise,
            dupe_ids=self._dupe_ids,
            rewrite_boosts=self._rewrite_boosts,
        )

    # ------------------------------------------------------------------
    # Retrieval simulation
    # ------------------------------------------------------------------

    def _simulate_retrieval(self) -> List[QueryResult]:
        """Run retrieval simulation over all episode queries using S_faulted."""
        results = []
        config = self._config
        n_c = len(self._chunk_ids)

        for i, query in enumerate(self._episode_queries):
            query_id = query["query_id"]
            scores = self._S_faulted[i]  # (n_chunks,)

            # Top-K by descending score
            top_k = min(config.top_k, n_c)
            top_indices = np.argsort(scores)[::-1][:top_k]

            # Filter by threshold
            retrieved: List[Tuple[int, float]] = [
                (self._chunk_ids[j], float(scores[j]))
                for j in top_indices
                if scores[j] >= config.similarity_threshold
            ]

            retrieved_ids = [cid for cid, _ in retrieved]
            retrieved_scores = [s for _, s in retrieved]

            # Coverage and precision
            r_star = set(self._ground_truth.get(str(query_id), []))
            r_agent = set(retrieved_ids)
            coverage = len(r_agent & r_star) / len(r_star) if r_star else 0.0
            precision = len(r_agent & r_star) / len(r_agent) if r_agent else 0.0

            # Context overflow check
            total_tokens = sum(
                self._chunk_id_to_tokens.get(cid, 100)
                for cid in retrieved_ids
            )

            results.append(
                QueryResult(
                    query_id=query_id,
                    query_text=query["text"],
                    retrieved_chunk_ids=retrieved_ids,
                    retrieval_scores=retrieved_scores,
                    n_retrieved=len(retrieved_ids),
                    coverage_score=float(np.clip(coverage, 0.0, 1.0)),
                    precision_score=float(np.clip(precision, 0.0, 1.0)),
                    is_multi_hop=bool(query.get("is_multi_hop", False)),
                )
            )

        return results

    # ------------------------------------------------------------------
    # Metrics
    # ------------------------------------------------------------------

    def _compute_metrics(self, results: List[QueryResult]) -> QualityMetrics:
        coverages = [r.coverage_score for r in results]
        precisions = [r.precision_score for r in results]
        n_empty = sum(1 for r in results if r.n_retrieved == 0)

        # Count context overflows by re-checking token counts
        n_overflow = 0
        config = self._config
        for r in results:
            total = sum(
                self._chunk_id_to_tokens.get(cid, 100)
                for cid in r.retrieved_chunk_ids
            )
            if total > config.context_window_limit:
                n_overflow += 1

        multi_hop_covs = [r.coverage_score for r in results if r.is_multi_hop]
        multi_hop_cov = float(np.mean(multi_hop_covs)) if multi_hop_covs else None

        return QualityMetrics(
            mean_coverage=float(np.mean(coverages)) if coverages else 0.0,
            mean_precision=float(np.mean(precisions)) if precisions else 0.0,
            mean_recall=float(np.mean(coverages)) if coverages else 0.0,
            n_empty_retrievals=n_empty,
            n_context_overflows=n_overflow,
            multi_hop_coverage=multi_hop_cov,
        )

    # ------------------------------------------------------------------
    # Reward
    # ------------------------------------------------------------------

    def _compute_reward(
        self,
        prev: Optional[QualityMetrics],
        new: QualityMetrics,
        action: RAGDebugAction,
    ) -> Reward:
        """Compute a progress-based reward in [0.0, 1.0].

        Design: reward reflects the absolute quality level (progress toward the
        success threshold) PLUS a small bonus for the direction of change.

        This ensures the full [0.0, ~0.89] range is utilised for non-terminal
        steps, giving the RL agent a strong per-step learning signal:
          - Terrible state, no improvement β†’ reward β‰ˆ 0.09
          - Mid quality, no change        β†’ reward β‰ˆ 0.42
          - At success threshold, no change β†’ reward β‰ˆ 0.64
          - Large improvement step        β†’ reward up to  0.89
          - Large regression + penalties  β†’ reward clipped to 0.00

        Terminal rewards (SUBMIT) remain in their own zone [0.7, 1.0] for
        success and [0.0, 0.15] for failure, as before.
        """
        components: Dict[str, float] = {}
        n_queries = max(len(self._episode_queries), 1)

        # --- Progress reward: absolute quality level signal ---
        # Maps quality_score to [0.10, 0.65] proportional to how close we are
        # to the task's success threshold.  Spans the reward range across the
        # full episode regardless of per-step delta magnitude.
        quality_target = 0.75 if self._task_id in (1, 2) else 0.70
        current_quality = self._quality_score(new)
        progress = min(1.0, current_quality / quality_target)
        components["progress_reward"] = 0.10 + 0.55 * progress

        # --- Delta bonus: immediate direction feedback ---
        # Distinguishes an improving step from a no-op at the same quality level.
        # Multiplied delta capped at Β±0.15 so a single step never dominates.
        if prev is not None:
            prev_quality = self._quality_score(prev)
            q_delta = current_quality - prev_quality
            components["delta_bonus"] = float(np.clip(q_delta * 2.0, -0.15, 0.15))

            # Empty retrieval signal: bidirectional (weight 0.06)
            empty_change = prev.n_empty_retrievals - new.n_empty_retrievals
            components["empty_retrieval_signal"] = float(np.clip(empty_change / n_queries, -1.0, 1.0)) * 0.06

            # Context overflow signal: bidirectional (weight 0.04)
            overflow_change = prev.n_context_overflows - new.n_context_overflows
            components["overflow_signal"] = float(np.clip(overflow_change / n_queries, -1.0, 1.0)) * 0.04
        else:
            components["delta_bonus"] = 0.0
            components["empty_retrieval_signal"] = 0.0
            components["overflow_signal"] = 0.0

        # --- Efficiency penalties ---
        components["step_cost"] = -0.01

        # Redundancy penalty for repeating the same action type consecutively
        if self._prev_action_type is not None and action.action_type == self._prev_action_type:
            components["redundancy_penalty"] = -0.04
        else:
            components["redundancy_penalty"] = 0.0

        # Penalty for invalid action parameters
        if self._last_action_error is not None:
            components["invalid_action_penalty"] = -0.05

        # --- Combine: no fixed base β€” progress_reward IS the base ---
        raw = sum(components.values())
        value = float(np.clip(raw, 0.0, 1.0))
        return Reward(value=value, components=components)

    def _compute_task_score(self, metrics: QualityMetrics) -> float:
        """Compute the scalar task score used for grading."""
        n_steps = self._state.step_count
        if self._task_id in (1, 2):
            efficiency = 1.0 - n_steps / _MAX_STEPS
            return (
                0.60 * metrics.mean_coverage
                + 0.25 * metrics.mean_precision
                + 0.15 * efficiency
            )
        else:  # task 3
            mh_cov = metrics.multi_hop_coverage or 0.0
            return (
                0.55 * metrics.mean_coverage
                + 0.25 * metrics.mean_precision
                + 0.20 * mh_cov
            )

    def _quality_score(self, metrics: QualityMetrics) -> float:
        """Quality portion of task_score, excluding efficiency. Normalized to [0, 1].

        Uses the same coverage:precision weighting as _compute_task_score so that
        step rewards are aligned with the terminal success criterion.
        """
        if self._task_id in (1, 2):
            # 0.60 + 0.25 = 0.85 max; normalize to [0, 1]
            return (0.60 * metrics.mean_coverage + 0.25 * metrics.mean_precision) / 0.85
        else:  # task 3: includes multi-hop coverage, already sums to 1.0
            mh_cov = metrics.multi_hop_coverage or 0.0
            return 0.55 * metrics.mean_coverage + 0.25 * metrics.mean_precision + 0.20 * mh_cov

    def _check_success(self, metrics: QualityMetrics, task_score: float) -> bool:
        if self._task_id == 1:
            return task_score >= 0.75
        elif self._task_id == 2:
            return task_score >= 0.75
        else:  # task 3
            mh_cov = metrics.multi_hop_coverage or 0.0
            return task_score >= 0.70 and mh_cov > 0.60

    # ------------------------------------------------------------------
    # Fault sampling
    # ------------------------------------------------------------------

    def _sample_faults(self, task_id: int, rng: np.random.Generator) -> List[FaultConfig]:
        if task_id == 1:
            idx = int(rng.integers(0, len(_TASK1_FAULT_SETS)))
            fault_types = _TASK1_FAULT_SETS[idx]
        elif task_id == 2:
            idx = int(rng.integers(0, len(_TASK2_FAULT_SETS)))
            fault_types = _TASK2_FAULT_SETS[idx]
        else:
            fault_types = _TASK3_FAULTS

        return [FaultConfig(fault_type=ft) for ft in fault_types]

    # ------------------------------------------------------------------
    # Query sampling
    # ------------------------------------------------------------------

    def _sample_queries(
        self, all_queries: List[Dict], task_id: int, rng: np.random.Generator
    ) -> List[Dict]:
        n = _N_EPISODE_QUERIES[task_id]

        if task_id == 3:
            regular = [q for q in all_queries if not q.get("is_multi_hop")]
            multi_hop = [q for q in all_queries if q.get("is_multi_hop")]

            n_mh = min(2, len(multi_hop))
            n_reg = n - n_mh

            reg_sample = list(rng.choice(len(regular), size=min(n_reg, len(regular)), replace=False))
            mh_sample = list(rng.choice(len(multi_hop), size=n_mh, replace=False))

            sampled = [regular[i] for i in reg_sample] + [multi_hop[i] for i in mh_sample]
        else:
            indices = list(rng.choice(len(all_queries), size=min(n, len(all_queries)), replace=False))
            sampled = [all_queries[i] for i in indices]

        return sampled

    def _calibrate_initial_difficulty(self, rng: np.random.Generator) -> None:
        """Nudge overly-strong reset states toward improvable starting points."""
        self._recompute_S_faulted()
        results = self._simulate_retrieval()
        metrics = self._compute_metrics(results)

        if not results:
            return

        full_cov_rate = float(
            np.mean([1.0 if r.coverage_score >= 0.999 else 0.0 for r in results])
        )
        cov_caps = {1: 0.60, 2: 0.52, 3: 0.48}
        full_cov_caps = {1: 0.50, 2: 0.45, 3: 0.40}

        cov_cap = cov_caps.get(self._task_id, 0.60)
        full_cov_cap = full_cov_caps.get(self._task_id, 0.50)
        if metrics.mean_coverage <= cov_cap and full_cov_rate <= full_cov_cap:
            return

        updates: Dict[str, Any] = {}
        if self._config.top_k > 3:
            shrink = int(rng.integers(1, 3))
            updates["top_k"] = max(3, self._config.top_k - shrink)

        fault_types_active = {f.fault_type for f in self._injected_faults}
        if FaultType.THRESHOLD_TOO_HIGH not in fault_types_active:
            bump = float(rng.uniform(0.05, 0.12))
            updates["similarity_threshold"] = min(
                0.75, self._config.similarity_threshold + bump
            )

        if updates:
            self._config = self._config.model_copy(update=updates)

    # ------------------------------------------------------------------
    # Diagnostic hints
    # ------------------------------------------------------------------

    def _generate_diagnostic_hints(
        self, metrics: QualityMetrics, results: List[QueryResult]
    ) -> List[str]:
        """Generate context-aware hints based on current metric patterns."""
        hints: List[str] = []
        cfg = self._config

        if metrics.n_empty_retrievals > 0:
            hints.append(
                f"{metrics.n_empty_retrievals} queries have empty retrievals β€” "
                "consider lowering similarity_threshold or increasing top_k."
            )

        if metrics.n_context_overflows > 0:
            hints.append(
                f"{metrics.n_context_overflows} queries exceed context window β€” "
                "consider increasing context_window_limit or reducing top_k."
            )

        if metrics.mean_coverage < 0.5 and metrics.mean_precision > 0.3:
            hints.append(
                "Low coverage with moderate precision suggests top_k is too small "
                "or the embedding model may not suit this domain."
            )

        if metrics.mean_coverage < 0.4 and metrics.mean_precision < 0.3:
            hints.append(
                "Both coverage and precision are low β€” check if the similarity threshold "
                "is filtering out too many chunks, or if the embedding model is mismatched."
            )

        # Check for score compression (sign of wrong embedding model or TOP_K_TOO_SMALL)
        all_scores = [s for r in results for s in r.retrieval_scores]
        if all_scores:
            score_std = float(np.std(all_scores))
            score_mean = float(np.mean(all_scores))
            if score_std < 0.05 and len(all_scores) > 3:
                hints.append(
                    f"Retrieval scores are tightly compressed (std={score_std:.3f}) β€” "
                    "this may indicate the wrong embedding model or score compression fault."
                )
            if score_mean > 0.7 and metrics.mean_precision < 0.5:
                hints.append(
                    "High retrieval scores but low precision β€” many irrelevant chunks are "
                    "scoring high. Consider enabling reranking or checking for duplicate flooding."
                )

        # Task 3 multi-hop hint
        if self._task_id == 3:
            mh_cov = metrics.multi_hop_coverage
            if mh_cov is not None and mh_cov < 0.5:
                hints.append(
                    f"Multi-hop coverage is low ({mh_cov:.3f}) β€” multi-hop queries need "
                    "broad retrieval. Consider increasing top_k and checking the embedding model."
                )

        # Reranking hint
        if not cfg.use_reranking and metrics.mean_precision < 0.4:
            hints.append(
                "Reranking is disabled. Enabling it can improve precision by re-scoring "
                "candidates with a cross-encoder."
            )

        return hints

    # ------------------------------------------------------------------
    # Observation builder
    # ------------------------------------------------------------------

    def _build_observation(
        self,
        results: List[QueryResult],
        reward: Optional[float],
    ) -> RAGDebugObservation:
        cs = self._corpus_stats_dict
        corpus_stats = CorpusStats(
            domain=Domain(self._domain),
            n_documents=cs.get("n_documents", 0),
            n_chunks=cs.get("n_chunks", len(self._chunks)),
            avg_chunk_tokens=cs.get("avg_chunk_tokens", 0),
            has_near_duplicates=bool(cs.get("has_near_duplicates", False)),
            n_queries=cs.get("n_queries", 0),
            n_multi_hop_queries=cs.get("n_multi_hop_queries", 0),
        )
        metrics = self._compute_metrics(results)
        hints = self._generate_diagnostic_hints(metrics, results)

        obs = RAGDebugObservation(
            pipeline_config=self._config,
            query_results=results,
            metrics=metrics,
            corpus_stats=corpus_stats,
            steps_taken=self._state.step_count,
            max_steps=_MAX_STEPS,
            task_id=self._task_id,
            task_description=_TASK_DESCRIPTION[self._task_id],
            done=self._done,
            last_action_error=self._last_action_error,
            diagnostic_hints=hints,
            reward_components=self._last_reward_components,
            reward=reward,
        )
        return obs


# Backward-compat alias (server/__init__.py and app.py import RagDebugEnvironment)
RagDebugEnvironment = RAGDebugEnvironment