File size: 25,280 Bytes
12323e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""RAG Generation for single / cross / multi document settings."""

from __future__ import annotations

import os
import re
import json
import argparse
from typing import Dict, List, Tuple, Any

import pandas as pd


# ======================== Config ========================

GEN_MODEL_PATH = "../model/Qwen3-4B-Instruct-2507-FP8"

BASE_XLSX_PATH = "../Expert-Annotated Relevant Sources Dataset/ClimRetrieve_base.xlsx"
CROSS_XLSX_PATH = "../Expert-Annotated Relevant Sources Dataset/ClimRetrieve_cross.xlsx"

RESULT_DIR_BASE = "./Embedding_Search_Results_Qwen"

RETRIEVAL_TOP_K = 5
MAX_NEW_TOKENS = 768
TEMPERATURE = 0.0


# ======================== Prompt Templates ========================

FEW_SHOT_EXAMPLES_SINGLE = """Example 1:
Question: Does the company have a strategy on waste management?
Context:
- "To meet our commitment to being zero waste by 2030, we are reducing our waste footprint through reuse, recycling and recovery."
- "We increased our reuse and recycle rates of all cloud hardware to 82 percent."
Answer: [YES]. The company has a clear strategy on waste management focused on reduction, recycling, and recovery to meet its 2030 zero waste commitment.

Example 2:
Question: Does the company report the climate change scenarios used to test the resilience of its business strategy?
Context:
- "Our ESG ratings improved from 2021 to 2022."
- "We continue to focus on sustainable sourcing across our supply chain."
Answer: [NO]. The provided context does not contain information about climate change scenarios used to test business strategy resilience.
"""


FEW_SHOT_EXAMPLES_CROSS = """Example 1:
Question: Among [Company A Report, Company B Report], did at least one company address "Does the company have a strategy on waste management?"
Context:
- (Company A) "We are committed to reducing waste by 50% by 2030 through our circular economy approach."
- (Company B) "Our revenue grew by 15% year-over-year driven by strong demand."
Answer: [YES]. Company A provides clear evidence of a waste management strategy with specific targets, while Company B's context does not address waste management.

Example 2:
Question: Among [Company A Report, Company B Report], did Company A provide stronger disclosure than Company B on "Does the company report its Scope 1 emissions?"
Context:
- (Company A) "We focus on employee diversity and inclusion programs."
- (Company B) "Our sustainability team oversees environmental compliance."
Answer: [NO]. Neither company's context provides evidence about Scope 1 emissions reporting, so the comparison claim is not supported.
"""


MULTI_ZERO_SHOT_BASE = """You are a regulator/auditor analyzing multi-document climate disclosure questions.
Use ONLY the provided context chunks.
Do not fabricate facts, numeric values, years, units, page numbers, or source ids.
Prioritize quantitative comparison: extract comparable metrics (value + unit + period) only if explicitly stated.
If evidence is insufficient, keep fields conservative: null/[]/"insufficient".
Set `conclusion` as a direct decision in this exact style: "[YES] <reason>" or "[NO] <reason>".
Return strict JSON only (no markdown, no prose outside JSON).
"""

MULTI_OUTPUT_SCHEMA = (
    '{'
    '"dimension":"string",'
    '"rows":[{"report":"string","year":"string","disclosure_status":"explicit | partial | missing",'
    '"key_points":["string"],"evidence_chunks":["E1","E4"]}],'
    '"ranking":[{"rank":1,"report":"string","rationale":"string"}],'
    '"conclusion":"string ([YES]/[NO] + reason)"'
    '}'
)


MULTI_SKILL_SPECS = {
    "Comparative Table Builder": {
        "prompt": (
            "Build a regulator-facing comparative table. "
            "For each report, output disclosure maturity and quantified metrics extracted from context. "
            "Use `quant_metrics` with objects: {metric, value, unit, period, note}. "
            "If a metric is unavailable for a report, do not invent it; use null values or omit the metric."
        ),
        "schema": (
            '{"skill":"Comparative Table Builder","dimension":"string","reports":[{"report":"string","year":"number|null",'
            '"maturity_level":"high|moderate|low|insufficient","key_evidence":["string"],'
            '"quant_metrics":[{"metric":"string","value":"number|null","unit":"string|null","period":"string|null","note":"string|null"}]}],'
            '"comparison_metrics":["string"],"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Trend & Quant Comparator": {
        "prompt": (
            "Build a quantitative trend comparison across reports. "
            "For each report, return measurable indicators with fields: value, intensity, attainment_rate, "
            "change_magnitude, trend_direction. Keep values numeric when present in context, otherwise null. "
            "Include 1-3 concise key_evidence bullets per report."
        ),
        "schema": (
            '{"skill":"Trend & Quant Comparator","dimension":"string","reports":[{"report":"string","year":"number|null",'
            '"key_evidence":["string"],"strength_score":"number|null",'
            '"quant_metrics":[{"metric":"string","value":"number|null","unit":"string|null","period":"string|null",'
            '"intensity":"number|null","attainment_rate":"number|null","change_magnitude":"number|null",'
            '"trend_direction":"up|down|flat|unknown|insufficient","note":"string|null"}]}],'
            '"metric_highlights":["string"],"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Target Attainment & Delta Benchmark": {
        "prompt": (
            "Create a target-attainment benchmark. For each report, extract baseline/current/target values where available, "
            "compute attainment_rate and delta fields (absolute and percent), and infer trend_direction conservatively. "
            "If any field is unavailable in context, return null/insufficient rather than guessing."
        ),
        "schema": (
            '{"skill":"Target Attainment & Delta Benchmark","dimension":"string","reports":[{"report":"string","year":"number|null",'
            '"overall_strength":"high|moderate|low|insufficient","key_evidence":["string"],'
            '"benchmarks":[{"metric":"string","baseline_value":"number|null","baseline_period":"string|null",'
            '"current_value":"number|null","current_period":"string|null","target_value":"number|null","target_period":"string|null",'
            '"attainment_rate":"number|null","delta_abs":"number|null","delta_percent":"number|null","intensity":"number|null",'
            '"unit":"string|null","trend_direction":"up|down|flat|unknown|insufficient","note":"string|null"}]}],'
            '"leaderboard":[{"report":"string","score":"number|null","reason":"string"}],"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Compliance Checklist": {
        "prompt": (
            "Create a compliance checklist comparison. "
            "Return item-level pass/partial/fail for each report and include quantified indicators where available. "
            "Also compute summary counts per report and include 1-3 concise key_evidence bullets per report."
        ),
        "schema": (
            '{"skill":"Compliance Checklist","dimension":"string","required_checks":["string"],"reports":[{"report":"string",'
            '"key_evidence":["string"],'
            '"summary":{"pass":"number","partial":"number","fail":"number","completion_rate":"number|null"},'
            '"checks":[{"item":"string","status":"pass|partial|fail|insufficient","quant_value":"number|null","quant_unit":"string|null","note":"string"}]}],'
            '"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Dimension Extractor": {
        "prompt": (
            "Extract disclosure dimensions and quantify coverage by bucket. "
            "Return bucket-level counts and any explicit quantitative metric snippets, plus 1-3 key_evidence bullets per report."
        ),
        "schema": (
            '{"skill":"Dimension Extractor","dimension":"string","reports":[{"report":"string","bucket_counts":{"Process":"number","Input":"number","Output":"number","Outcome":"number","Governance":"number","Risk":"number"},'
            '"key_evidence":["string"],'
            '"quant_metrics":[{"metric":"string","value":"number|null","unit":"string|null","period":"string|null","note":"string|null"}],'
            '"coverage_level":"high|moderate|low|insufficient"}],"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Contradiction/Consistency Check": {
        "prompt": (
            "Evaluate consistency across reports. "
            "Return rule checks and a quantified consistency score per report pair/group. "
            "Include a concise key_evidence list that supports the consistency conclusion."
        ),
        "schema": (
            '{"skill":"Contradiction/Consistency Check","dimension":"string","key_evidence":["string"],'
            '"checks":[{"rule":"string","result":"consistent|inconsistent|insufficient","note":"string"}],'
            '"scores":{"consistent":"number","inconsistent":"number","insufficient":"number","consistency_rate":"number|null"},'
            '"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
    "Consensus/Count (Portfolio Statistics)": {
        "prompt": (
            "Produce portfolio-level statistics from retrieved context only. "
            "Count explicit/partial/missing disclosures and include percentage split. "
            "For each report, include 1-3 key_evidence bullets."
        ),
        "schema": (
            '{"skill":"Consensus/Count (Portfolio Statistics)","dimension":"string","counts":{"explicit":"number","partial":"number","missing":"number","total":"number"},'
            '"percentages":{"explicit":"number|null","partial":"number|null","missing":"number|null"},'
            '"per_report":[{"report":"string","label":"explicit|partial|missing|insufficient","quant_metrics":[{"metric":"string","value":"number|null","unit":"string|null","period":"string|null"}]}],'
            '"key_evidence_by_report":[{"report":"string","key_evidence":["string"]}],'
            '"consensus_items":["string"],"outliers":["string"],"conclusion":"string ([YES]/[NO] + reason)"}'
        ),
    },
}


METRIC_HINT_RULES = [
    ("trend", ["value", "intensity", "attainment_rate", "change_magnitude", "delta_percent"]),
    ("attainment", ["target_value", "current_value", "attainment_rate", "delta_percent"]),
    ("delta", ["delta_abs", "delta_percent", "baseline_value", "current_value"]),
    ("baseline", ["baseline_value", "current_value", "target_value", "delta_percent"]),
    ("kpi", ["value", "unit", "period", "change_magnitude"]),
    ("risk", ["number_of_risk_categories", "assessment_frequency_per_year", "scenario_count", "horizon_year"]),
    ("methodolog", ["framework_count", "method_steps_count", "indicator_count", "coverage_percent"]),
    ("scenario", ["scenario_count", "temperature_pathway_count", "horizon_year"]),
    ("resilience", ["scenario_count", "stress_test_count", "horizon_year"]),
    ("impact", ["scope1_emissions", "scope2_emissions", "scope3_emissions", "reduction_percent"]),
    ("waste", ["waste_reduction_percent", "recycling_rate_percent", "waste_diverted_tons", "target_year"]),
    ("industry peers", ["initiative_count", "partnership_count", "engagement_frequency_per_year"]),
    ("downstream", ["supplier_coverage_percent", "partner_count", "assessment_completion_percent"]),
    ("water", ["water_withdrawal_change_percent", "water_intensity_change_percent", "sites_count", "target_year"]),
]


def infer_metric_hints(question: str) -> List[str]:
    q = str(question or "").lower()
    hints = []
    for key, metrics in METRIC_HINT_RULES:
        if key in q:
            for m in metrics:
                if m not in hints:
                    hints.append(m)
    if not hints:
        hints = ["reported_metric_count", "target_year", "percent_change", "absolute_value"]
    return hints


def normalize_multi_skill_name(skill_name: str) -> str:
    s = str(skill_name or "").strip().lower()
    if "trend" in s and "quant" in s:
        return "Trend & Quant Comparator"
    if "attainment" in s or ("delta" in s and "benchmark" in s):
        return "Target Attainment & Delta Benchmark"
    if "compliance" in s and "check" in s:
        return "Compliance Checklist"
    if "dimension" in s and "extract" in s:
        return "Dimension Extractor"
    if "contradiction" in s or "consistency" in s:
        return "Contradiction/Consistency Check"
    if "consensus" in s or "portfolio" in s or "count" in s:
        return "Consensus/Count (Portfolio Statistics)"
    return "Comparative Table Builder"


def infer_multi_skill_name(question: str) -> str:
    q = str(question or "").lower()
    if ("trend" in q and "quant" in q) or "change magnitude" in q or "measurable progress" in q:
        return "Trend & Quant Comparator"
    if "attainment rate" in q or "target gap" in q or "baseline" in q or "delta percent" in q:
        return "Target Attainment & Delta Benchmark"
    if "checklist" in q or "compliance" in q:
        return "Compliance Checklist"
    if "extract" in q and "dimension" in q:
        return "Dimension Extractor"
    if "contradiction" in q or "consistent" in q or "inconsistent" in q:
        return "Contradiction/Consistency Check"
    if "consensus" in q or "outlier" in q or "portfolio" in q or "count" in q:
        return "Consensus/Count (Portfolio Statistics)"
    return "Comparative Table Builder"


def get_multi_skill_spec(skill_name: str) -> Dict[str, str]:
    name = normalize_multi_skill_name(skill_name)
    return {
        "skill_name": name,
        "skill_prompt": MULTI_SKILL_SPECS[name]["prompt"],
        "output_json_schema": MULTI_OUTPUT_SCHEMA,
    }


def build_yes_no_prompt(question: str, contexts: List[str], doc_mode: str) -> str:
    few_shot = FEW_SHOT_EXAMPLES_CROSS if doc_mode == "cross" else FEW_SHOT_EXAMPLES_SINGLE
    ctx_text = "\n".join(f'- "{c}"' for c in contexts)

    return f"""You are an expert analyst evaluating corporate sustainability reports.
Based ONLY on the provided context passages, answer the question with [YES] or [NO], followed by a brief reasoning.

Format your answer as:
[YES]. <reasoning> OR [NO]. <reasoning>

{few_shot.strip()}

Now answer the following:
Question: {question}
Context:
{ctx_text}
Answer:"""


def build_multi_zero_shot_prompt(
    question: str,
    contexts: List[str],
    skill_name: str = "",
    skill_prompt: str = "",
    output_json_schema: str = "",
    retrieval_query: str = "",
    metric_hints: List[str] | None = None,
) -> str:
    ctx_text = "\n".join(f'[E{i}] "{c}"' for i, c in enumerate(contexts, start=1))
    skill_name = str(skill_name or "").strip()
    skill_prompt = str(skill_prompt or "").strip()
    output_json_schema = str(output_json_schema or "").strip()
    retrieval_query = str(retrieval_query or "").strip()
    metric_hints = metric_hints or infer_metric_hints(question)
    metric_hint_text = ", ".join(metric_hints)

    return f"""{MULTI_ZERO_SHOT_BASE}

Matched skill: {skill_name}
Skill guidance: {skill_prompt}
Retrieval query dimension: {retrieval_query}
Suggested comparable metrics: {metric_hint_text}
Target JSON schema: {output_json_schema}

Question:
{question}

Context:
{ctx_text}

JSON Answer:"""


# ======================== Ground Truth ========================

def extract_yes_no(text: str) -> str | None:
    text_upper = str(text).strip().upper()
    m = re.match(r"\[*\s*(YES|NO)\s*\]*", text_upper)
    if m:
        return m.group(1)
    if text_upper.startswith("YES"):
        return "YES"
    if text_upper.startswith("NO"):
        return "NO"
    return None


def load_ground_truth(doc_mode: str) -> Dict[Any, str]:
    if doc_mode == "single":
        df = pd.read_excel(BASE_XLSX_PATH, index_col=0)
        gt = {}
        for _, row in df.iterrows():
            report = row["Document"]
            question = row["Question"]
            label = extract_yes_no(row["Answer"])
            if label is not None:
                gt[(report, question)] = label
        return gt

    if doc_mode == "cross":
        df = pd.read_excel(CROSS_XLSX_PATH)
        if "Unnamed: 0" in df.columns:
            df = df.drop(columns=["Unnamed: 0"])
        gt = {}
        for _, row in df.iterrows():
            question = row["Question"]
            label = extract_yes_no(row["Answer"])
            if label is not None and question not in gt:
                gt[question] = label
        return gt

    # multi has no closed-form YES/NO ground truth
    return {}


# ======================== Retrieval Result Loading ========================

def get_result_dir(chunk_mode: str, doc_mode: str) -> str:
    parts = [RESULT_DIR_BASE]
    if chunk_mode != "length":
        parts.append(chunk_mode)
    if doc_mode != "single":
        parts.append(doc_mode)
    return "_".join(parts)


def load_retrieval_contexts(chunk_mode: str, doc_mode: str, top_k: int) -> Dict[Any, Any]:
    result_dir = get_result_dir(chunk_mode, doc_mode)
    result_file = os.path.join(result_dir, f"Qwen3-Embedding-0.6B__{chunk_mode}__{doc_mode}__{top_k}.csv")
    if not os.path.exists(result_file):
        raise FileNotFoundError(f"Retrieval result file not found: {result_file}")

    df = pd.read_csv(result_file)
    retrieved = df[df["question_retrieved"] == 1].copy()
    retrieved = retrieved.sort_values("similarity_score", ascending=False)

    if doc_mode == "single":
        contexts = {}
        for (report, question), group in retrieved.groupby(["report", "question"], sort=False):
            contexts[(report, question)] = group["chunk_text"].tolist()
        return contexts

    contexts = {}
    for question, group in retrieved.groupby("question", sort=False):
        chunks = []
        for _, row in group.iterrows():
            chunks.append(f"({row['report'].replace('.pdf', '')}, chunk={row['chunk_idx']}) {row['chunk_text']}")

        payload = {"chunks": chunks}
        for col in ["skill_name", "skill_prompt", "output_json_schema", "retrieval_query"]:
            if col in group.columns:
                non_null = group[col].dropna()
                payload[col] = non_null.iloc[0] if len(non_null) > 0 else ""
            else:
                payload[col] = ""
        contexts[question] = payload
    return contexts


# ======================== Main ========================

def main():
    parser = argparse.ArgumentParser(description="RAG generation for ClimRetrieve")
    parser.add_argument("--chunk", type=str, default="length", choices=["length", "structure"])
    parser.add_argument("--doc", type=str, default="single", choices=["single", "cross", "multi"])
    parser.add_argument("--top_k", type=int, default=RETRIEVAL_TOP_K)
    parser.add_argument("--model", type=str, default=GEN_MODEL_PATH)
    parser.add_argument("--max_tokens", type=int, default=MAX_NEW_TOKENS)
    parser.add_argument("--temperature", type=float, default=TEMPERATURE)
    parser.add_argument("--limit", type=int, default=None, help="Only generate for first N questions")
    args = parser.parse_args()

    chunk_mode = args.chunk
    doc_mode = args.doc
    top_k = args.top_k
    limit = args.limit

    print("=" * 70)
    print("RAG Generation")
    print(f"  CHUNK={chunk_mode}, DOC={doc_mode}, TOP_K={top_k}, LIMIT={limit}")
    print(f"  Model: {args.model}")
    print("=" * 70)

    print("\n[Step 1] Load retrieval contexts")
    contexts = load_retrieval_contexts(chunk_mode, doc_mode, top_k)
    print(f"  Context groups: {len(contexts)}")

    if limit is not None and limit > 0:
        items = list(contexts.items())[:limit]
        contexts = dict(items)
        print(f"  Applied limit -> {len(contexts)} groups")

    gt = {}
    if doc_mode in ("single", "cross"):
        print("\n[Step 2] Load ground truth")
        gt = load_ground_truth(doc_mode)
        print(f"  Ground truth count: {len(gt)}")
    else:
        print("\n[Step 2] Multi mode: skip YES/NO ground truth")

    print("\n[Step 3] Build prompts")
    prompts = []
    prompt_keys = []

    if doc_mode == "single":
        for (report, question), ctx_list in contexts.items():
            prompts.append(build_yes_no_prompt(question, ctx_list[:top_k], doc_mode))
            prompt_keys.append({"report": report, "question": question})
    elif doc_mode == "cross":
        for question, payload in contexts.items():
            ctx_list = payload["chunks"] if isinstance(payload, dict) else payload
            prompts.append(build_yes_no_prompt(question, ctx_list[:top_k], doc_mode))
            prompt_keys.append({"question": question})
    else:
        for question, payload in contexts.items():
            ctx_list = payload.get("chunks", [])
            inferred_skill = payload.get("skill_name", "") or infer_multi_skill_name(question)
            spec = get_multi_skill_spec(inferred_skill)
            skill_name = spec["skill_name"]
            skill_prompt = payload.get("skill_prompt", "").strip() or spec["skill_prompt"]
            output_json_schema = spec["output_json_schema"]
            retrieval_query = payload.get("retrieval_query", "").strip() or question
            prompts.append(
                build_multi_zero_shot_prompt(
                    question=question,
                    contexts=ctx_list[:top_k],
                    skill_name=skill_name,
                    skill_prompt=skill_prompt,
                    output_json_schema=output_json_schema,
                    retrieval_query=retrieval_query,
                    metric_hints=infer_metric_hints(question),
                )
            )
            prompt_keys.append(
                {
                    "question": question,
                    "skill_name": skill_name,
                    "retrieval_query": retrieval_query,
                }
            )

    print(f"  Prompt count: {len(prompts)}")
    if prompts:
        print(f"  Prompt preview: {prompts[0][:400]}...")

    print("\n[Step 4] Run generation")
    from vllm import LLM, SamplingParams
    llm = LLM(model=args.model, max_model_len=8192, dtype="auto", trust_remote_code=True)
    sampling_params = SamplingParams(
        temperature=args.temperature,
        max_tokens=args.max_tokens,
        top_p=1.0,
    )
    outputs = llm.generate(prompts, sampling_params)

    print("\n[Step 5] Collect results")
    results = []
    correct = 0
    total_eval = 0

    for i, output in enumerate(outputs):
        generated_text = output.outputs[0].text.strip()
        key = prompt_keys[i]

        if doc_mode == "multi":
            results.append(
                {
                    **key,
                    "generated_answer": generated_text,
                }
            )
            continue

        predicted_label = extract_yes_no(generated_text)
        if doc_mode == "single":
            gt_label = gt.get((key["report"], key["question"]))
        else:
            gt_label = gt.get(key["question"])

        is_correct = None
        if predicted_label is not None and gt_label is not None:
            is_correct = 1 if predicted_label == gt_label else 0
            correct += is_correct
            total_eval += 1

        results.append(
            {
                **key,
                "generated_answer": generated_text[:800],
                "predicted_label": predicted_label,
                "gt_label": gt_label,
                "is_correct": is_correct,
            }
        )

    print("\n[Step 6] Save outputs")
    output_dir = get_result_dir(chunk_mode, doc_mode)
    os.makedirs(output_dir, exist_ok=True)

    limit_tag = f"__limit{limit}" if (limit is not None and limit > 0) else ""
    output_file = os.path.join(output_dir, f"generation__{chunk_mode}__{doc_mode}__top{top_k}{limit_tag}.csv")
    pd.DataFrame(results).to_csv(output_file, index=False, encoding="utf-8-sig")
    print(f"  Result file: {output_file}")

    if doc_mode == "multi":
        metrics = {
            "chunk_mode": chunk_mode,
            "doc_mode": doc_mode,
            "top_k": top_k,
            "limit": limit,
            "generated_count": len(results),
        }
    else:
        accuracy = correct / total_eval if total_eval > 0 else 0.0
        metrics = {
            "chunk_mode": chunk_mode,
            "doc_mode": doc_mode,
            "top_k": top_k,
            "limit": limit,
            "total_eval": total_eval,
            "correct": correct,
            "accuracy": accuracy,
            "unparsed": sum(1 for r in results if r.get("predicted_label") is None),
            "no_gt": sum(1 for r in results if r.get("gt_label") is None),
        }
        print(f"  Accuracy: {accuracy:.4f} ({correct}/{total_eval})")

    metrics_file = os.path.join(output_dir, f"generation_metrics__{chunk_mode}__{doc_mode}__top{top_k}{limit_tag}.json")
    with open(metrics_file, "w", encoding="utf-8") as f:
        json.dump(metrics, f, indent=2)
    print(f"  Metrics file: {metrics_file}")

    print("\n" + "=" * 70)
    print("Generation completed.")
    print("=" * 70)


if __name__ == "__main__":
    main()