Blame for construction/utils.py

"""Shared parsing, JSONL and local-model helpers extracted from the original code."""
import json
import re
import os
import time
from pathlib import Path

PROJECT_ROOT = Path(__file__).resolve().parent
RESULTS_DIR = PROJECT_ROOT / "work" / "candidate_discovery"
STAGE1_OUT = str(PROJECT_ROOT / "work" / "query_curation" / "intent.jsonl")
MODEL_PATH = str(PROJECT_ROOT / "models" / "Qwen" / "Qwen3.6-27B")
MAX_IMAGE_PIXELS = 2 * 1024 * 1024

def extract_json_obj(text):
    """Extract a JSON object ({...}) from model output. Returns dict or fallback."""
    try:
        match = re.search(r"\{.*\}", text, re.DOTALL)
        if match:
            parsed = json.loads(match.group())
            if isinstance(parsed, dict):
                return parsed
    except (json.JSONDecodeError, ValueError):
        pass
    return {"content": "未知", "intent": "未知", "note": "未知", "_raw": text}

def extract_keywords(text):
    try:
        match = re.search(r"\[.*?\]", text, re.DOTALL)
        if match:
            parsed = json.loads(match.group())
            if isinstance(parsed, list) and all(isinstance(x, str) for x in parsed):
                return parsed
    except (json.JSONDecodeError, ValueError):
        pass

    lines = text.strip().split("\n")
    keywords = []
    for line in lines:
        line = line.strip()
        cleaned = re.sub(r"^[\d]+[.、)\]]\s*", "", line)
        cleaned = re.sub(r"^[-*·•]\s*", "", cleaned)
        cleaned = re.sub(r"\*\*(.+?)\*\*", r"\1", cleaned)
        cleaned = cleaned.split("：")[0].split(":")[0].strip()
        if cleaned and len(cleaned) < 30:
            keywords.append(cleaned)

    if not keywords:
        keywords = [k.strip() for k in text.split("，") if k.strip()]
    if not keywords:
        keywords = [k.strip() for k in text.split(",") if k.strip()]

    return keywords[:15]

def append_result(result, pipeline_name, out_dir=None):
    """增量追加 JSONL。"""
    base = Path(out_dir) if out_dir is not None else RESULTS_DIR
    base.mkdir(exist_ok=True, parents=True)
    out_path = base / f"{pipeline_name}.jsonl"
    with open(out_path, "a", encoding="utf-8") as f:
        f.write(json.dumps(result, ensure_ascii=False) + "\n")
        f.flush()

def load_done_idx(pipeline_name, out_dir=None):
    """读取已完成的 idx 集合，用于断点续存跳过。"""
    base = Path(out_dir) if out_dir is not None else RESULTS_DIR
    base.mkdir(exist_ok=True, parents=True)
    out_path = base / f"{pipeline_name}.jsonl"
    done = set()
    if out_path.exists():
        with open(out_path, encoding="utf-8") as f:
            for line in f:
                line = line.strip()
                if line:
                    try:
                        done.add(json.loads(line)["idx"])
                    except (json.JSONDecodeError, KeyError):
                        continue
    return done

def load_scenes(stage1_path=None):
    """读 Stage1 缓存 -> {idx: scene_dict}。这是下游模块消费 Stage1 输出的主接口。"""
    path = stage1_path or STAGE1_OUT
    scenes = {}
    if os.path.exists(path):
        with open(path, encoding="utf-8") as f:
            for line in f:
                line = line.strip()
                if not line:
                    continue
                try:
                    r = json.loads(line)
                    scenes[str(r["idx"])] = r
                except (json.JSONDecodeError, KeyError):
                    continue
    return scenes

def load_model_and_processor(model_path=None):
    import torch
    from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
    path = model_path or MODEL_PATH
    print(f"Loading processor from {path} ...")
    processor = AutoProcessor.from_pretrained(path, trust_remote_code=True)
    # 限制图片分辨率。
    processor.image_processor.size["longest_edge"] = MAX_IMAGE_PIXELS
    print(f"  max_pixels(longest_edge)={MAX_IMAGE_PIXELS}")

    print(f"Loading model (bf16, device_map=cuda) from {path} ...")
    t0 = time.time()
    model = Qwen3_5ForConditionalGeneration.from_pretrained(
        path,
        dtype=torch.bfloat16,
        device_map="cuda",
        trust_remote_code=True,
    )
    model.eval()
    print(f"Model loaded in {time.time() - t0:.1f}s")
    return model, processor

def strip_think_block(text):
    return re.sub(r"🤔.*?</think>\s*", "", text, flags=re.DOTALL).strip()

def run_text_only(model, processor, text_prompt, max_new_tokens=1024, enable_thinking=False,
                  do_sample=True, temperature=0.7, top_p=0.8, top_k=20, seed=None):
    import torch
    """纯文本推理（单条）。采样参数语义同 run_multimodal：默认采样兼容旧 pipeline，
    需可复现时传 do_sample=False + seed=0。"""
    tokenizer = processor.tokenizer
    messages = [{"role": "user", "content": text_prompt}]
    text = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True,
        enable_thinking=enable_thinking,
    )
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    if seed is not None:
        torch.manual_seed(seed)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(seed)
    gen_kwargs = dict(max_new_tokens=max_new_tokens, do_sample=do_sample)
    if do_sample:
        gen_kwargs.update(temperature=temperature, top_p=top_p, top_k=top_k)
    with torch.no_grad():
        output_ids = model.generate(**inputs, **gen_kwargs)
    generated = output_ids[0][inputs["input_ids"].shape[1]:]
    response = tokenizer.decode(generated, skip_special_tokens=True)
    if enable_thinking:
        return strip_think_block(response)
    return response.strip()

def run_text_only_batch(model, processor, prompts, batch_size=8, max_new_tokens=256,
                        enable_thinking=False, do_sample=True, temperature=0.7,
                        top_p=0.8, top_k=20, seed=None, sort_by_length=False,
                        show_progress=True):
    import torch
    """批量纯文本推理，保留输入顺序。"""
    tokenizer = processor.tokenizer
    n = len(prompts)
    if n == 0:
        return []

    # 1. 套 chat template（不 tokenize，留待批量 padding）
    chat_texts = []
    for p in prompts:
        messages = [{"role": "user", "content": p}]
        chat_texts.append(tokenizer.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True,
            enable_thinking=enable_thinking,
        ))

    # 2. 分桶顺序（按 prompt 长度升序，连续分批 → 同批长度接近、padding 最少）
    order = list(range(n))
    if sort_by_length:
        lens = [len(tokenizer(t, add_special_tokens=False)["input_ids"]) for t in chat_texts]
        order.sort(key=lambda i: lens[i])

    # 3. 设置 left-padding（decoder-only 生成必须）+ pad token
    prev_side = tokenizer.padding_side
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "left"

    results = [None] * n
    n_batches = (n + batch_size - 1) // batch_size
    try:
        gen_kwargs = dict(max_new_tokens=max_new_tokens, do_sample=do_sample)
        if do_sample:
            gen_kwargs.update(temperature=temperature, top_p=top_p, top_k=top_k)
        for bi, b_start in enumerate(range(0, n, batch_size)):
            batch_idx = order[b_start:b_start + batch_size]
            batch_texts = [chat_texts[i] for i in batch_idx]
            inputs = tokenizer(
                batch_texts, padding=True, return_tensors="pt",
            ).to(model.device)
            if seed is not None:
                torch.manual_seed(seed)
                if torch.cuda.is_available():
                    torch.cuda.manual_seed_all(seed)
            with torch.no_grad():
                output_ids = model.generate(**inputs, **gen_kwargs)
            # left-pad 下所有行 prompt 都被右对齐到同一长度，切掉前 input_len 即得新生成
            input_len = inputs["input_ids"].shape[1]
            gen_ids = output_ids[:, input_len:]
            for j, i in enumerate(batch_idx):
                text = tokenizer.decode(gen_ids[j], skip_special_tokens=True)
                if enable_thinking:
                    text = strip_think_block(text)
                results[i] = text.strip()
            if show_progress:
                done = b_start + len(batch_idx)
                print(f"  batch {bi + 1}/{n_batches} done ({done}/{n})", flush=True)
    finally:
        tokenizer.padding_side = prev_side

    return results



GRADE_ALIASES = {"L1": "L1", "L2": "L2", "N1": "N1", "N2": "N2", "N3": "N3",
                 "hard_neg_constraint": "N1", "hard_neg_lookalike": "N2", "irrelevant": "N3"}


def parse_grade(text):
    """Require an explicit valid JSON label; failed generation is not a vote."""
    obj = extract_json_obj(text)
    grade = obj.get("grade")
    if not isinstance(grade, str) or grade not in GRADE_ALIASES:
        raise ValueError("Model response does not contain a valid five-way grade")
    internal = {"N1": "hard_neg_constraint", "N2": "hard_neg_lookalike", "N3": "irrelevant"}
    return {"grade": internal.get(grade, grade), "reason": obj.get("reason")}


def read_jsonl(path):
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]


def write_jsonl(path, rows):
    path = Path(path)
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as f:
        for row in rows:
            f.write(json.dumps(row, ensure_ascii=False) + "\n")


def read_pairs(path):
    """Accept flat pairs or the historical per-query candidates format."""
    pairs = []
    for row in read_jsonl(path):
        if "item_id" in row:
            candidates = [row]
        elif isinstance(row.get("candidates"), list):
            candidates = row["candidates"]
        else:
            raise ValueError(f"Expected item_id or candidates in {path}")
        for candidate in candidates:
            pairs.append({**candidate, "idx": str(row["idx"]),
                          "item_id": str(candidate["item_id"])})
    return pairs


def pair_key(row):
    return str(row["idx"]), str(row["item_id"])


def index_pairs(rows):
    indexed = {}
    for row in rows:
        key = pair_key(row)
        if key in indexed:
            raise ValueError(f"Duplicate pair: {key}")
        indexed[key] = row
    return indexed


def leader_clusters(sub, gidxs, threshold):
    """按中心性降序分配；只吸收与 leader 直接相似的成员。sub 对角为 -1。"""
    m = len(gidxs)
    if m <= 1:
        return [list(gidxs)] if m else []
    cent = (sub.sum(axis=1) + 1.0) / (m - 1)
    order = sorted(range(m), key=lambda i: -float(cent[i]))
    assigned = [False] * m
    clusters = []
    for li in order:
        if assigned[li]:
            continue
        members = [li]
        assigned[li] = True
        for j in order:
            if assigned[j] or j == li:
                continue
            if float(sub[li][j]) >= threshold:
                members.append(j)
                assigned[j] = True
        clusters.append([gidxs[x] for x in members])
    return clusters


def select_representative(members, emb, kind):
    import numpy as np
    sub = emb[members]
    if kind == "query":
        centroid = sub.mean(axis=0)
        centroid /= np.linalg.norm(centroid) + 1e-9
        scores = sub @ centroid
    else:
        sim = sub @ sub.T
        np.fill_diagonal(sim, 0)
        scores = sim.sum(axis=1) / max(len(members) - 1, 1)
    return members[int(np.argmax(scores))]


def deduplicate(items, emb, kind, threshold=None):
    import numpy as np
    from collections import defaultdict
    threshold = threshold if threshold is not None else (0.70 if kind == "query" else 0.60)
    emb = np.asarray(emb, dtype=np.float32)
    if emb.ndim != 2 or len(emb) != len(items):
        raise ValueError("向量行数必须与输入记录一一对应")
    norms = np.linalg.norm(emb, axis=1, keepdims=True)
    if not np.isfinite(emb).all() or (norms == 0).any():
        raise ValueError("向量必须有限且非零")
    emb = emb / norms
    buckets = defaultdict(list)
    for i, row in enumerate(items):
        key = row["treecate_full_path"] if kind == "product" else "all"
        if not key:
            raise ValueError("商品必须有叶子类目的完整路径")
        buckets[key].append(i)
    kept, mapping = [], {}
    id_key = "idx" if kind == "query" else "item_id"
    for indices in buckets.values():
        sim = emb[indices] @ emb[indices].T
        np.fill_diagonal(sim, -1)
        for members in leader_clusters(sim, indices, threshold):
            rep = select_representative(members, emb, kind)
            kept.append(rep)
            for i in members:
                mapping[str(items[i][id_key])] = str(items[rep][id_key])
    return [items[i] for i in sorted(kept)], mapping


def load_embeddings(path, ids):
    import numpy as np
    with np.load(path, allow_pickle=False) as data:
        if list(map(str, data["ids"])) != list(map(str, ids)):
            raise ValueError("向量缓存的 ID 顺序与输入不一致")
        return data["emb"]