scripts/salvage_engine.py

#!/usr/bin/env python3
"""
Zero-Waste Disfluency Salvage Engine.
1. Deterministic cue-swapper for retractions / repetitive cues (0 GPU compute).
2. Qwen 3.5 9B salvage worker (Ollama port 11434 with thought bypass) for false starts and clean rejects.
"""

import json
import os
import re
import sys
import time
import urllib.request
from collections import Counter
from typing import Optional, Dict, Any, List

sys.path.insert(0, "/opt/vox/sandbox/scripts")
from corrupt_from_clean import normalize_span_stem, get_span_prefix2

CORPUS_PATH = "/opt/vox/sandbox/corpus/disfluency_15k.jsonl"
BAK_FS_PATH = "/opt/vox/temp/backups/disfluency_15k.jsonl.bak_repetitive_fs"
REJECTS_PATH = "/opt/vox/temp/backups/disfluency_15k_rejects.jsonl"

RETRACTION_CUES = [
    "or rather", "scratch that", "make that", "correction", "actually I meant",
    "let me rephrase", "sorry no", "no wait", "that is to say", "I mean",
    "whoops", "rather", "check that", "strike that", "wait scratch that",
    "excuse me", "pardon me", "no I mean", "or I should say", "actually"
]

def load_current_corpus_state():
    stem_counts = Counter()
    prefix2_counts = Counter()
    committed_clean = set()
    total_committed = 0

    if os.path.exists(CORPUS_PATH):
        with open(CORPUS_PATH) as f:
            for line in f:
                d = json.loads(line)
                committed_clean.add(d["clean_text"])
                total_committed += 1
                for s in d.get("spans", []):
                    text = s.get("text", "")
                    stem = normalize_span_stem(text)
                    stem_counts[stem] += 1
                    pfx = get_span_prefix2(text)
                    if pfx and len(pfx.split()) >= 2:
                        prefix2_counts[pfx] += 1

    return committed_clean, stem_counts, prefix2_counts, total_committed

def deterministic_cue_swap(example: Dict[str, Any], stem_counts: Counter, prefix2_counts: Counter, max_span: int = 10, max_pfx: int = 15) -> Optional[Dict[str, Any]]:
    """Deterministically swap saturated cue words with an unsaturated cue word."""
    spans = example.get("spans", [])
    if len(spans) != 1:
        return None

    span_info = spans[0]
    span_text = span_info.get("text", "")
    span_range = span_info.get("span", [])
    raw_text = example.get("raw_text", "")
    clean_text = example.get("clean_text", "")

    # Look for known cue phrases inside span
    matched_cue = None
    for cue in ["wait no", "sorry", "I mean", "actually", "no wait", "well anyway"]:
        if cue.lower() in span_text.lower():
            matched_cue = cue
            break

    if not matched_cue:
        return None

    # Try replacement cues in order of least frequency
    for rep in RETRACTION_CUES:
        # Case preservation
        pat = re.compile(re.escape(matched_cue), re.IGNORECASE)
        new_span_text = pat.sub(rep, span_text, count=1)
        if new_span_text == span_text:
            continue

        new_stem = normalize_span_stem(new_span_text)
        new_pfx = get_span_prefix2(new_span_text)

        if stem_counts[new_stem] >= max_span:
            continue
        if new_pfx and len(new_pfx.split()) >= 2 and prefix2_counts[new_pfx] >= max_pfx:
            continue

        # Build new raw text
        start, end = span_range
        new_raw_text = raw_text[:start] + new_span_text + raw_text[end:]
        new_end = start + len(new_span_text)

        # Strict roundtrip validation
        reconstructed_clean = new_raw_text[:start] + new_raw_text[new_end:]
        if reconstructed_clean != clean_text:
            continue

        # Accepted
        stem_counts[new_stem] += 1
        if new_pfx and len(new_pfx.split()) >= 2:
            prefix2_counts[new_pfx] += 1

        new_ex = dict(example)
        new_ex["raw_text"] = new_raw_text
        new_ex["spans"] = [{
            "label": "speech disfluency",
            "span": [start, new_end],
            "text": new_span_text,
            "confidence": 1.0,
            "origin": "salvage:deterministic_swap"
        }]
        return new_ex

    return None

def query_qwen(prompt: str, temp: float = 0.4, timeout: float = 8.0) -> Optional[str]:
    """Query Qwen 3.5 9B with thought bypass for fast completion."""
    payload = json.dumps({
        "model": "qwen3.5:9b",
        "prompt": prompt,
        "stream": False,
        "raw": True,
        "options": {"temperature": temp, "num_predict": 150}
    }).encode("utf-8")

    req = urllib.request.Request(
        "http://127.0.0.1:11434/api/generate",
        data=payload,
        headers={"Content-Type": "application/json"}
    )
    try:
        with urllib.request.urlopen(req, timeout=timeout) as resp:
            data = json.loads(resp.read().decode("utf-8"))
            return data.get("response", "")
    except Exception as e:
        return None

def qwen_salvage(clean_text: str, stem_counts: Counter, prefix2_counts: Counter, max_span: int = 10, max_pfx: int = 15) -> Optional[Dict[str, Any]]:
    """Ask Qwen 3.5 9B to generate a diverse disfluency with negative constraints."""
    # Find top saturated prefixes to ban
    banned = [p for p, c in prefix2_counts.most_common(12) if c >= (max_pfx - 3)]
    banned_str = ", ".join([f'"{b}"' for b in banned[:8]]) if banned else '"because i", "if we"'

    prompt = f"""<|im_start|>system
You are an expert speech disfluency annotator.
Given a clean sentence, produce a natural spoken version with exactly ONE realistic speech disfluency (repetition, retraction, hesitation, or restart).
Rules:
1. Do NOT start the disfluent span with: {banned_str}.
2. Output STRICT JSON only.
Format:
{{"raw_text": "...", "span_text": "..."}}
Example:
{{"raw_text": "Regarding the, yes I am also looking for information.", "span_text": "Regarding the, "}}
<|im_end|>
<|im_start|>user
Clean sentence: "{clean_text}"
Generate disfluent version:
<|im_end|>
<|im_start|>assistant
<|thought|>
Generating diverse disfluency avoiding banned prefixes.
<|endofthought|>
{{"raw_text":\""""

    raw_resp = query_qwen(prompt, temp=0.5)
    if not raw_resp:
        return None

    full_json_str = '{"raw_text":' + raw_resp.strip()
    # Find JSON boundaries
    m = re.search(r'\{.*\}', full_json_str, re.DOTALL)
    if not m:
        return None

    try:
        data = json.loads(m.group(0))
        raw_text = data.get("raw_text", "").strip()
        span_text = data.get("span_text", "")
        if not raw_text or not span_text:
            return None

        # Verify span in raw_text
        idx = raw_text.find(span_text)
        if idx == -1:
            return None

        start = idx
        end = idx + len(span_text)

        # Verify roundtrip
        reconstructed = raw_text[:start] + raw_text[end:]
        if reconstructed != clean_text:
            return None

        stem = normalize_span_stem(span_text)
        pfx = get_span_prefix2(span_text)

        if stem_counts[stem] >= max_span:
            return None
        if pfx and len(pfx.split()) >= 2 and prefix2_counts[pfx] >= max_pfx:
            return None

        # Accepted
        stem_counts[stem] += 1
        if pfx and len(pfx.split()) >= 2:
            prefix2_counts[pfx] += 1

        return {
            "category": "DISFLUENCY",
            "raw_text": raw_text,
            "clean_text": clean_text,
            "spans": [{
                "label": "speech disfluency",
                "span": [start, end],
                "text": span_text,
                "confidence": 1.0,
                "origin": "salvage:qwen3.5_worker"
            }]
        }
    except Exception:
        return None

def main():
    print("=== Starting Disfluency Salvage Engine ===")
    committed_clean, stem_counts, prefix2_counts, total_committed = load_current_corpus_state()
    print(f"Current committed records: {total_committed}")
    print(f"Tracking {len(stem_counts)} unique stems, {len(prefix2_counts)} 2-word prefixes")

    salvaged_deterministic = 0
    salvaged_qwen = 0

    # Collect clean sentences to salvage
    salvage_pool: List[Dict[str, Any]] = []

    # 1. From repetitive false start backup
    if os.path.exists(BAK_FS_PATH):
        with open(BAK_FS_PATH) as f:
            for line in f:
                d = json.loads(line)
                c = d.get("clean_text")
                if c and c not in committed_clean:
                    salvage_pool.append(d)

    # 2. From judge rejects
    if os.path.exists(REJECTS_PATH):
        with open(REJECTS_PATH) as f:
            for line in f:
                d = json.loads(line)
                c = d.get("clean_text")
                if c and c not in committed_clean:
                    salvage_pool.append({"clean_text": c})

    print(f"Candidates available for salvage: {len(salvage_pool)}")

    OUT_PATH = "/opt/vox/temp/salvaged_qwen.jsonl"
    with open(OUT_PATH, "a", encoding="utf-8") as out:
        for idx, item in enumerate(salvage_pool):
            clean = item.get("clean_text")
            if clean in committed_clean:
                continue

            # Step 1: Try deterministic cue swap if item has raw_text and spans
            res = None
            if "raw_text" in item and "spans" in item:
                res = deterministic_cue_swap(item, stem_counts, prefix2_counts)
                if res:
                    res["id"] = f"vxb_syn_{total_committed + 1:05d}"
                    out.write(json.dumps(res, ensure_ascii=False) + "\n")
                    out.flush()
                    committed_clean.add(clean)
                    total_committed += 1
                    salvaged_deterministic += 1
                    print(f"[{total_committed:5d}] Salvaged (Deterministic): {res['spans'][0]['text']!r}")
                    continue

            # Step 2: Try Qwen salvage worker
            res = qwen_salvage(clean, stem_counts, prefix2_counts)
            if res:
                res["id"] = f"vxb_syn_{total_committed + 1:05d}"
                out.write(json.dumps(res, ensure_ascii=False) + "\n")
                out.flush()
                committed_clean.add(clean)
                total_committed += 1
                salvaged_qwen += 1
                print(f"[{total_committed:5d}] Salvaged (Qwen Worker): {res['spans'][0]['text']!r}")
            else:
                pass

            if (idx + 1) % 50 == 0:
                print(f"Progress: {idx+1}/{len(salvage_pool)} | Det: {salvaged_deterministic} | Qwen: {salvaged_qwen} | Total: {total_committed}")

    print(f"\n=== Salvage Complete ===")
    print(f"Deterministic Salvaged: {salvaged_deterministic}")
    print(f"Qwen Salvaged: {salvaged_qwen}")
    print(f"Total Committed in Corpus: {total_committed}")

if __name__ == "__main__":
    main()