argo-search 1.0.1 → 2.9.7

package/scripts/query_variants.py

@@ -0,0 +1,298 @@
+#!/usr/bin/env python3
+"""
+query_variants.py — 无 LLM 查询变体生成（深度研究多路召回）
+
+生成与原查询「真正不同」的多个检索变体，弥补单查询漏检。
+策略：
+  1. 原查询（恒保留）
+  2. 问句化（陈述 → 问题）
+  3. 概念扩展（近义/上位概念替换）
+  4. 反方观点（找反方/批评视角）
+  5. 范围调整（短查询加限定 / 长查询去限定词）
+  6. 缩写与全称互换
+
+纯规则实现，无外部依赖，中英双语。
+"""
+
+from __future__ import annotations
+
+import re
+from cli_io import dumps
+
+# 概念扩展映射（不只同义词——相关表述框架）
+CONCEPT_MAP = {
+    # 技术
+    "ai": ["artificial intelligence", "machine learning", "deep learning"],
+    "ml": ["machine learning", "statistical learning"],
+    "llm": ["large language model", "foundation model"],
+    "api": ["interface", "integration", "SDK", "endpoint"],
+    "saas": ["software as a service", "cloud software"],
+    "devops": ["deployment automation", "CI/CD", "infrastructure as code"],
+    "kubernetes": ["k8s", "container orchestration"],
+    "docker": ["containerization", "container runtime"],
+    "database": ["data store", "persistence layer"],
+    "microservices": ["service-oriented architecture", "distributed systems"],
+    # 商业
+    "roi": ["return on investment", "cost-benefit", "payback period"],
+    "kpi": ["key performance indicator", "metric", "benchmark"],
+    "b2b": ["business to business", "enterprise sales"],
+    "b2c": ["business to consumer", "consumer market"],
+    "revenue": ["income", "sales", "top line"],
+    "profit": ["margin", "bottom line", "earnings"],
+    "strategy": ["approach", "framework", "playbook"],
+    "market": ["industry", "sector", "vertical"],
+    # 制造
+    "manufacturing": ["production", "fabrication", "industrial"],
+    "supply chain": ["logistics", "procurement", "sourcing"],
+    # 金融
+    "stock": ["equity", "shares", "securities"],
+    "bond": ["fixed income", "debt instrument", "treasury"],
+    "crypto": ["cryptocurrency", "digital assets", "blockchain"],
+    "inflation": ["price increases", "purchasing power", "CPI"],
+    "interest rate": ["fed funds rate", "monetary policy", "yield"],
+    # 中文
+    "人工智能": ["AI", "机器学习", "深度学习", "大模型"],
+    "大模型": ["大语言模型", "LLM", "基础模型", "AI"],
+    "算法": ["模型", "方法", "框架"],
+    "机器人": ["自动化", "智能体", "Agent"],
+    "股票": ["股价", "行情", "证券", "equity"],
+    "基金": ["ETF", "公募", "理财"],
+    "区块链": ["加密货币", "分布式账本", "比特币"],
+    "云计算": ["云服务", "SaaS", "PaaS"],
+    "开源": ["开放源码", "open source"],
+    "软件": ["应用程序", "系统", "工具"],
+    "硬件": ["芯片", "处理器", "设备"],
+}
+
+QUESTION_PREFIXES = [
+    "what is", "how does", "why is", "what are the benefits of",
+    "what are the risks of", "how to", "what is the future of",
+]
+
+OPPOSITION_TRIGGERS = {
+    "best": "worst problems with",
+    "benefits": "risks drawbacks of",
+    "advantages": "disadvantages limitations of",
+    "why": "why not criticism of",
+    "success": "failure case study",
+    "growing": "declining stagnating",
+    "popular": "overrated criticism",
+    "recommended": "alternatives to avoid",
+    "safe": "risks dangers of",
+    "cheap": "hidden costs of",
+    "easy": "challenges difficulties of",
+    "fast": "slow problems with",
+    "good": "problems criticism of",
+    "pros": "cons drawbacks",
+    # 中文反方触发词
+    "推荐": "吐槽 避坑 负面评价",
+    "优势": "劣势 缺点 局限",
+    "好处": "风险 坏处 缺点",
+    "为什么": "为什么不 反对 批评",
+    "最佳": "最差 问题 坑",
+    "安全": "风险 隐患 危险",
+    "值得": "不值得 翻车 后悔",
+}
+
+# 长查询限定词（范围调整时移除）
+QUALIFIERS = {
+    "latest", "best", "top", "new", "recent", "current", "modern",
+    "2024", "2025", "2026", "today", "now", "ultimate", "complete",
+    "comprehensive", "definitive", "essential",
+    "最新", "最好", "最热", "当前", "现在", "近期", "最全",
+}
+
+# 缩写/全称互换
+ACRONYM_MAP = {
+    "ai": "artificial intelligence",
+    "ml": "machine learning",
+    "nlp": "natural language processing",
+    "llm": "large language model",
+    "api": "application programming interface",
+    "sdk": "software development kit",
+    "ui": "user interface",
+    "ux": "user experience",
+    "db": "database",
+    "oss": "open source software",
+    "saas": "software as a service",
+    "crm": "customer relationship management",
+    "erp": "enterprise resource planning",
+    "kpi": "key performance indicator",
+    "ipo": "initial public offering",
+    "cto": "chief technology officer",
+    "ceo": "chief executive officer",
+}
+
+
+def generate_query_variations(original_query: str) -> list[str]:
+    """生成 3-5 个真正不同的查询变体（含原查询）。"""
+    if not original_query or not original_query.strip():
+        return []
+    original = original_query.strip()
+    variations = [original]
+    query_lower = original.lower()
+    words = query_lower.split()
+
+    # 策略 2：问句化
+    question = _to_question(query_lower, words)
+    if question and question.lower() != query_lower:
+        variations.append(question)
+
+    # 策略 3：概念扩展
+    expanded = _expand_concepts(original, words)
+    if expanded and expanded.lower() != query_lower:
+        variations.append(expanded)
+
+    # 策略 4：反方观点
+    opposing = _opposing_viewpoint(original, query_lower, words)
+    if opposing and opposing.lower() != query_lower:
+        variations.append(opposing)
+
+    # 策略 5：范围调整
+    scoped = _adjust_scope(original, query_lower, words)
+    if scoped and scoped.lower() != query_lower:
+        variations.append(scoped)
+
+    # 策略 6：缩写/全称互换
+    expanded_acronym = _expand_acronym(original, query_lower)
+    if expanded_acronym and expanded_acronym.lower() != query_lower:
+        variations.append(expanded_acronym)
+
+    # 去重（大小写不敏感）
+    seen = set()
+    unique = []
+    for v in variations:
+        key = v.strip().lower()
+        if key not in seen:
+            seen.add(key)
+            unique.append(v)
+
+    return unique[:6]
+
+
+def _to_question(query: str, words: list[str]) -> str | None:
+    """陈述 → 问题形式。"""
+    if query.endswith("?") or words[0] in (
+        "how", "what", "why", "when", "where", "who", "which", "is", "are",
+        "can", "do", "does",
+    ):
+        return None
+
+    # 英文问句壳只对纯拉丁查询生成：套在中文实体上（「what is 贵州茅台
+    # 2024年 年报 and how does it work」）对英文引擎是噪声、对中文引擎是
+    # 无效壳，实测占掉 2/4 子查询槽位。中文语义（是什么/怎么用）由
+    # research_expand 的领域模板与工作包承担。
+    if any("\u4e00" <= c <= "\u9fff" for c in query):
+        return None
+
+    # 动作类 → "how to"
+    action_words = {"install", "setup", "configure", "build", "create", "deploy",
+                    "fix", "solve", "debug", "optimize", "improve", "migrate"}
+    if words[0] in action_words or (len(words) > 1 and words[1] in action_words):
+        return f"how to {query}"
+
+    # 名词短语 → "what is ... and how does it work"
+    if len(words) <= 4:
+        return f"what is {query} and how does it work"
+
+    # 长查询 → "why"
+    return f"why {query}"
+
+
+def _expand_concepts(original: str, words: list[str]) -> str | None:
+    """概念扩展：替换为上位/替代概念。"""
+    result = original
+    expanded = False
+
+    for word in words:
+        if word in CONCEPT_MAP:
+            alternatives = CONCEPT_MAP[word]
+            replacement = alternatives[0]
+            result = re.sub(r"\b" + re.escape(word) + r"\b", replacement,
+                            result, count=1, flags=re.IGNORECASE)
+            expanded = True
+            break
+
+    if not expanded:
+        for phrase, alternatives in CONCEPT_MAP.items():
+            if " " in phrase and phrase.lower() in original.lower():
+                result = original.lower().replace(phrase, alternatives[0], 1)
+                expanded = True
+                break
+
+    return result if expanded else None
+
+
+def _opposing_viewpoint(original: str, query_lower: str,
+                        words: list[str]) -> str | None:
+    """反方观点生成。"""
+    for trigger, opposition in OPPOSITION_TRIGGERS.items():
+        if trigger in words:
+            return query_lower.replace(trigger, opposition, 1)
+
+    # 通用 criticism 壳与 _to_question 同理：不套中文实体（英文模板对
+    # 中文查询是无效壳）。触发词命中（词替换）保留——产出与查询同语言。
+    if any("\u4e00" <= c <= "\u9fff" for c in original):
+        return None
+    if len(words) >= 2:
+        return f"criticism problems with {original}"
+    return None
+
+
+def _adjust_scope(original: str, query_lower: str, words: list[str]) -> str | None:
+    """范围调整：短查询加限定，长查询去限定词。"""
+    # 短查询 → 加时间限定；限定词与查询同语言（英文壳套中文实体是无效变体）
+    if len(words) <= 2:
+        if any("\u4e00" <= c <= "\u9fff" for c in original):
+            return f"{original} 最新进展"
+        return f"{original} in 2026 latest developments"
+
+    # 长查询 → 去掉限定词（最新/最热等）
+    narrowed_words = [w for w in words if w not in QUALIFIERS]
+    if len(narrowed_words) < len(words) and len(narrowed_words) >= 2:
+        return " ".join(narrowed_words)
+
+    # 深度补充
+    if not any(w in words for w in ("research", "study", "analysis", "paper",
+                                    "academic", "研究", "综述", "分析", "报告")):
+        return f"{original} research analysis"
+    return None
+
+
+def _expand_acronym(original: str, query_lower: str) -> str | None:
+    """缩写 → 全称。"""
+    for word in query_lower.split():
+        clean = word.strip(".,;:()[]")
+        if clean in ACRONYM_MAP:
+            return re.sub(r"\b" + re.escape(clean) + r"\b",
+                          ACRONYM_MAP[clean], original, count=1,
+                          flags=re.IGNORECASE)
+    return None
+
+
+# ── CLI ────────────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    import argparse
+    import json
+    import sys
+
+    p = argparse.ArgumentParser(description="Argo 查询变体生成（无 LLM 多路召回）")
+    p.add_argument("query", nargs="?", default=None, help="原始查询")
+    p.add_argument("--stdin", action="store_true", help="从 stdin 读取")
+    p.add_argument("--json", action="store_true", help="JSON 输出")
+    args = p.parse_args()
+
+    q = args.query
+    if args.stdin:
+        q = sys.stdin.read().strip()
+    if not q:
+        p.error("需要查询或 --stdin")
+
+    variants = generate_query_variations(q)
+    if args.json:
+        print(dumps({"original": q, "variants": variants}))
+    else:
+        print(f"原查询：{q}")
+        for i, v in enumerate(variants, 1):
+            print(f"  {i}. {v}")