fix_query_grammar.py

"""
Fix Query Grammar with LLM
============================
After mechanically adding "In YYYY," to TempLAMA queries,
use Gemini to fix grammar while preserving:
  - The year
  - The entity
  - The relation
  - The ___ blank (if present)
  - The factual meaning

Usage:
    cd ~/svd_kg/knowledge_drift
    python fix_query_grammar.py
"""

import json, os, re, time, sys
from collections import Counter
from concurrent.futures import ThreadPoolExecutor, as_completed
from threading import Lock

import google.generativeai as genai

# ============================================================
# CONFIG
# ============================================================

GEMINI_MODEL = "gemini-2.5-flash"
GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY", "")
if not GOOGLE_API_KEY:
    print("Set GOOGLE_API_KEY environment variable or edit this script")
    sys.exit(1)

RATE_LIMIT_DELAY = 0.15
MAX_WORKERS = 100
BATCH_SIZE = 20  # queries per LLM call

TIER1_PATH = "data/knowledge_drift_unified_tier1.json"
OUTPUT_PATH = "data/knowledge_drift_unified_tier1_clean.json"
BACKUP_PATH = "data/knowledge_drift_unified_tier1_prefixed_backup.json"

# ============================================================
# SETUP
# ============================================================

genai.configure(api_key=GOOGLE_API_KEY)
model = genai.GenerativeModel(GEMINI_MODEL)

print("=" * 70)
print("  FIX QUERY GRAMMAR WITH LLM")
print("=" * 70)

# ============================================================
# STEP 1: LOAD AND ADD YEAR PREFIXES MECHANICALLY
# ============================================================

print("\nStep 1: Loading and adding year prefixes...")

with open(TIER1_PATH) as f:
    tier1 = json.load(f)
samples = tier1["samples"]
print(f"  Loaded: {len(samples)} samples")

# Remove noise first
REMOVE_RELATIONS = {"", "رئيس", "رئيس الوزراء", "معايير الصكوك", "ولي العهد",
                     "member countries expansion", "political situation",
                     "board and leadership changes"}

samples = [s for s in samples if s.get("relation", "") not in REMOVE_RELATIONS]
print(f"  After noise removal: {len(samples)}")

def has_year_prefix(query):
    return bool(re.match(r'^In \d{4},?\s', query))

# Track which queries need fixing
queries_to_fix = []  # list of (index, original_query, year)
for i, s in enumerate(samples):
    query = s.get("query", "")
    year = s.get("year", None)
    
    if not has_year_prefix(query) and year:
        try:
            year_int = int(year)
            # Mechanical prefix first
            if query.startswith("_"):
                mechanical = f"In {year_int}, {query}"
            else:
                mechanical = f"In {year_int}, {query[0].lower()}{query[1:]}"
            queries_to_fix.append((i, query, mechanical, year_int))
            s["query_original"] = query
            s["query"] = mechanical  # temporary, will be overwritten by LLM
        except (ValueError, TypeError):
            pass

print(f"  Queries needing year prefix: {len(queries_to_fix)}")
print(f"  Queries already with prefix: {len(samples) - len(queries_to_fix)}")

# ============================================================
# STEP 2: FIX GRAMMAR WITH GEMINI IN BATCHES
# ============================================================

print(f"\nStep 2: Fixing grammar with {GEMINI_MODEL}...")
print(f"  Batches: {len(queries_to_fix) // BATCH_SIZE + 1} ({BATCH_SIZE} queries each)")

SYSTEM_PROMPT = """You are a grammar editor. You will receive a batch of factual queries that have been mechanically modified by prepending "In YYYY," to them. Some may read awkwardly.

Your job: Fix ONLY the grammar to make each query read naturally, while preserving:
1. The exact year mentioned
2. The entity names (do NOT change names)
3. The blank marker "___" (keep it exactly as ___)
4. The factual meaning (do NOT change what is being asked)

Rules:
- Keep it concise — don't add extra words unnecessarily
- Preserve the format: the query should be a factual statement or question
- If the query already reads fine, return it unchanged
- Do NOT add periods or punctuation that wasn't there

Return ONLY a JSON array of fixed queries, one per input, in the same order. No explanation."""

lock = Lock()
fixed_queries = {}  # index -> fixed query
errors = []
total_done = 0

def fix_batch(batch):
    """Send a batch of queries to Gemini, return fixed versions."""
    indices = [b[0] for b in batch]
    mechanicals = [b[2] for b in batch]
    
    prompt = f"""Fix the grammar of these {len(mechanicals)} queries. Return a JSON array of strings.

Queries:
{json.dumps(mechanicals, indent=2)}"""
    
    try:
        response = model.generate_content(
            [{"role": "user", "parts": [prompt]}],
            generation_config=genai.types.GenerationConfig(
                temperature=0.0,
                max_output_tokens=4096,
            ),
        )
        
        text = response.text.strip()
        # Extract JSON array from response
        # Handle markdown code blocks
        if "```json" in text:
            text = text.split("```json")[1].split("```")[0].strip()
        elif "```" in text:
            text = text.split("```")[1].split("```")[0].strip()
        
        fixed = json.loads(text)
        
        if len(fixed) != len(indices):
            # Length mismatch — fall back to mechanical
            return {idx: mech for idx, mech in zip(indices, mechanicals)}
        
        return {idx: f for idx, f in zip(indices, fixed)}
        
    except Exception as e:
        # Fall back to mechanical version
        return {idx: mech for idx, mech in zip(indices, mechanicals)}

# Create batches
batches = []
for i in range(0, len(queries_to_fix), BATCH_SIZE):
    batches.append(queries_to_fix[i:i + BATCH_SIZE])

print(f"  Processing {len(batches)} batches with {MAX_WORKERS} workers...")

with ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
    futures = {}
    for batch_idx, batch in enumerate(batches):
        future = executor.submit(fix_batch, batch)
        futures[future] = batch_idx
        time.sleep(RATE_LIMIT_DELAY)
    
    for future in as_completed(futures):
        batch_idx = futures[future]
        try:
            result = future.result()
            with lock:
                fixed_queries.update(result)
                total_done += len(result)
                if total_done % 200 == 0 or total_done == len(queries_to_fix):
                    print(f"    Progress: {total_done}/{len(queries_to_fix)} ({total_done/len(queries_to_fix)*100:.1f}%)")
        except Exception as e:
            errors.append((batch_idx, str(e)))
            # Use mechanical versions for failed batches
            batch = batches[batch_idx]
            with lock:
                for idx, orig, mech, yr in batch:
                    fixed_queries[idx] = mech
                total_done += len(batch)

print(f"  Done. Fixed: {len(fixed_queries)}, Errors: {len(errors)}")

# ============================================================
# STEP 3: APPLY FIXES
# ============================================================

print("\nStep 3: Applying fixes...")

applied = 0
for idx, orig, mech, yr in queries_to_fix:
    if idx in fixed_queries:
        fixed = fixed_queries[idx]
        # Validate: must still contain the year
        if str(yr) in fixed and len(fixed) > 10:
            samples[idx]["query"] = fixed
        else:
            # LLM broke it — use mechanical
            samples[idx]["query"] = mech
        applied += 1
    else:
        samples[idx]["query"] = mech
        applied += 1

print(f"  Applied: {applied} fixes")

# ============================================================
# STEP 4: SHOW EXAMPLES
# ============================================================

print("\nStep 4: Examples of fixes...")
print(f"\n  {'Original':50s} → {'Fixed':50s}")
print("  " + "-" * 105)

count = 0
for idx, orig, mech, yr in queries_to_fix[:15]:
    fixed = samples[idx]["query"]
    if fixed != mech:  # LLM actually changed something
        print(f"  {orig[:50]:50s} → {fixed[:50]:50s}")
        count += 1
    if count >= 10:
        break

if count == 0:
    print("  (LLM kept all mechanical versions unchanged — grammar was fine)")
    for idx, orig, mech, yr in queries_to_fix[:5]:
        print(f"  {orig[:50]:50s} → {samples[idx]['query'][:50]:50s}")

# ============================================================
# STEP 5: VERIFY
# ============================================================

print("\nStep 5: Verification...")

n_with = sum(1 for s in samples if has_year_prefix(s.get("query", "")))
n_without = len(samples) - n_with
print(f"  With year prefix: {n_with}/{len(samples)} ({n_with/len(samples)*100:.1f}%)")
print(f"  Without year prefix: {n_without}")

# Check drift rate is no longer confounded with format
for has_yr in [True, False]:
    subset = [s for s in samples if has_year_prefix(s.get("query", "")) == has_yr]
    if not subset:
        continue
    n_d = sum(1 for s in subset if s.get("is_drifted_qwen25", False))
    pct = n_d / len(subset) * 100
    label = "with year" if has_yr else "no year"
    print(f"  {label}: {n_d}/{len(subset)} drifted for Qwen ({pct:.1f}%)")

# ============================================================
# STEP 6: SAVE
# ============================================================

print("\nStep 6: Saving...")

# Re-number
for i, s in enumerate(samples):
    s["sample_id"] = f"tier1_v2_{str(i).zfill(6)}"

tier1["samples"] = samples
tier1["metadata"]["total_samples"] = len(samples)
tier1["metadata"]["version"] = "2.1_grammar_fixed"

# Save clean version
with open(OUTPUT_PATH, "w") as f:
    json.dump(tier1, f, indent=2, ensure_ascii=False)
print(f"  Saved: {OUTPUT_PATH} ({len(samples)} samples)")

# Also update main tier1
with open(TIER1_PATH, "w") as f:
    json.dump(tier1, f, indent=2, ensure_ascii=False)
print(f"  Updated: {TIER1_PATH}")

# ============================================================
# STEP 7: CREATE PER-MODEL DATASETS
# ============================================================

print("\nStep 7: Creating per-model datasets...")

MODELS = {
    "llama2":  "is_drifted_llama2",
    "mistral": "is_drifted_mistral",
    "llama31": "is_drifted_llama31",
    "qwen25":  "is_drifted_qwen25",
    "gemma2":  "is_drifted_gemma2",
}

for model_name, drift_key in MODELS.items():
    model_dataset = json.loads(json.dumps(tier1))
    for s in model_dataset["samples"]:
        s["is_drifted_query"] = s.get(drift_key, False)
        s["temporal_zone"] = "post_cutoff"
    
    path = f"data/tier1_{model_name}.json"
    with open(path, "w") as f:
        json.dump(model_dataset, f, indent=2, ensure_ascii=False)
    
    n_d = sum(1 for s in model_dataset["samples"] if s["is_drifted_query"])
    print(f"  {model_name:10s}: {path} ({n_d} drifted, {len(samples)-n_d} stable)")

# ============================================================
# DONE
# ============================================================

print(f"""
{'=' * 70}
  DONE — ALL DATASETS READY
{'=' * 70}

  Clean Tier 1:  {OUTPUT_PATH} ({len(samples)} samples)
  All queries now have uniform year-prefix format.
  Grammar verified by {GEMINI_MODEL}.
  Noise samples removed.
  Per-model datasets created.

  NEXT: Re-extract hidden states for Qwen on the fixed data:

    python disentanglement_v2.py \\
        --model Qwen/Qwen2.5-7B-Instruct \\
        --dataset data/tier1_qwen25.json \\
        --output_dir data/experiments/tier1_qwen25_v2

{'=' * 70}
""")