feat(server): add workbench API, DSH planner worker and RAG pipeline

2026-09-16 19:06:20 +08:00
parent 72fc5b47c7
commit d4b5b3f43a
122 changed files with 31906 additions and 0 deletions
@@ -0,0 +1,310 @@
+"""AI 辅助创建（小白友好）：大白话 → 结构化草稿 + 资源冲突检测。
+
+两个能力：
+1. draft_from_blab：用户用自己的话描述「我想要一个干什么的 Agent / 技能 / 模板」，
+   由当前默认模型解析成结构化草稿（名称/说明/规则/指令/提示模板字段），
+   用户在表单里确认修改后再正式创建——AI 只出草稿，不直接落库。
+2. conflict_scan：创建/修改前的冲突检测，回答「Agent、技能、模板之间会不会打架」：
+   - 重名/高相似名称
+   - 同名约束互相矛盾（规则 vs 技能 vs 模板约束）
+   - 关键词互斥（如「口语化」对「书面语」「输出 JSON」对「段落正文」）
+
+冲突不阻断创建，返回 warning 列表由前端展示，用户判断后可继续。
+"""
+from __future__ import annotations
+
+import re
+from difflib import SequenceMatcher
+from sqlmodel import Session, select
+
+from .llm import chat_json
+from .models import Agent, AgentSkill, AgentPromptTemplate
+
+# 语义互斥对：同一条消息里同时要求两边就会打架。
+MUTUAL_EXCLUSIVE_PAIRS: list[tuple[str, str]] = [
+    ("口语", "书面"), ("大白话", "书面语"), ("口语", "公文"),
+    ("简短", "详细"), ("精炼", "展开"), ("长文", "短句"),
+    ("严肃", "幽默"), ("专业", "接地气"), ("正式", "轻松"),
+    ("列表", "段落"), ("分点", "一段到底"), ("json", "口语"),
+    ("markdown", "纯文本"), ("提问", "陈述"), ("第一人称", "第三人称"),
+]
+
+DRAFT_SYSTEM_PROMPT = """你是私域Agent系统的配置助手。用户不懂 AI，只会用自己的话描述想要的东西。
+你的任务是把用户的描述整理成结构化草稿，输出 JSON。
+
+输出要求：
+- 只输出一个合法 JSON 对象，不要 markdown 代码块标记，不要解释。
+- 名称要具体（4-12 个字），说明是一句完整的话。
+- 用户没提到的字段用空字符串或空列表，不要编造细节。
+- 规则和指令用大白话写，像跟同事交代事情，不写 AI 术语。
+
+三种资源的 JSON 结构：
+1. agent（智能体）:
+   {"kind": "agent", "name": "...", "role": "production|qc|composer|service", "desc": "...",
+    "rules": ["规则1", "规则2"], "content_tags": ["标签"], "skill_suggestions": ["想要的能力描述"],
+    "advice": "给用户的一句提醒或建议"}
+2. skill（技能）:
+   {"kind": "skill", "name": "...", "desc": "...",
+    "instructions": ["步骤1", "步骤2"], "advice": "..."}
+3. template（提示模板）:
+   {"kind": "template", "name": "...", "desc": "...",
+    "role_prompt": "你是...（角色设定）", "task_template": "...",
+    "constraints": ["约束1"], "style": "文风描述", "advice": "..."}
+
+role 判断：写内容的选 production，检查内容质量的选 qc，把多段合成完整稿的选 composer，
+客服对话选 service。不确定时选 production。"""
+
+
+
+# ---------------- 知识库检索注入 ----------------
+
+def _kb_ids_for_kind(session: Session, kind: str) -> list[int]:
+    """按草稿类型挑选参与检索的知识库：
+    agent/skill/template 都查全部业务库——草稿可能涉及任何板块；
+    检索相关性排序会自动挑最相关的块。"""
+    from .models import KnowledgeBase
+    names = {"心法", "技法", "功法", "情欲内容", "产品信息", "销售逻辑"}
+    out = []
+    for kb in session.exec(select(KnowledgeBase)).all():
+        if kb.name in names:
+            out.append(kb.id)
+    return out
+
+
+def retrieve_kb_context(session: Session, description: str, kind: str, k: int = 8) -> tuple[list[dict], str]:
+    """按描述检索知识块，返回 (hits, 注入 prompt 的上下文文本)。检索失败返回空不阻塞。"""
+    try:
+        from .rag import pipeline as rag_pipeline
+        kb_ids = _kb_ids_for_kind(session, kind)
+        if not kb_ids:
+            return [], ""
+        hits = rag_pipeline.search_kbs_remote(
+            session, kb_ids, description, k=k,
+            user_id=0, organization_id=1)
+        hits = [h for h in hits if str(h.get("text") or "").strip()][:k]
+        if not hits:
+            return [], ""
+        parts = []
+        for i, h in enumerate(hits, 1):
+            kb_name = str(h.get("kb_name") or "")
+            src = str((h.get("meta") or {}).get("file") or "")
+            text = re.sub(r"\s+", " ", str(h.get("text") or ""))[:420]
+            parts.append(f"[知识{i}]（{kb_name}·{src}）{text}")
+        return hits, "\n\n".join(parts)
+    except Exception as exc:  # 检索失败不阻塞草稿生成
+        print(f"[ai-assist] 知识库检索失败（不影响生成）: {exc}", flush=True)
+        return [], ""
+
+
+def _pick_model(session: Session, organization_id: int | None = None) -> tuple[int, str] | None:
+    """挑一个可用的 Provider/模型：优先全局默认，其次任意启用的。"""
+    from .models import ModelProvider, ModelCatalog, AppSetting
+    setting = session.exec(select(AppSetting).where(AppSetting.key == "model_default")).first()
+    if setting:
+        try:
+            data = __import__("json").loads(setting.value)
+            provider = session.get(ModelProvider, int(data.get("provider_id") or 0))
+            if provider and provider.enabled and (organization_id is None or int(provider.organization_id or 1) in {0, int(organization_id)} ) and data.get("model_name"):
+                return provider.id, str(data["model_name"])
+        except (ValueError, TypeError, KeyError):
+            pass
+    # Do not fall back to the first enabled built-in provider: it may have no
+    # development credential while a user-configured provider is usable.
+    from .secret_store import resolve_secret
+    providers = session.exec(select(ModelProvider).where(ModelProvider.enabled == True)).all()  # noqa: E712
+    if organization_id is not None:
+        providers = [p for p in providers if int(p.organization_id or 1) in {0, int(organization_id)}]
+    for provider in providers:
+        if provider.default_model and resolve_secret(provider.api_key, f"provider-{provider.id}"):
+            return provider.id, provider.default_model
+    return None
+
+
+def draft_from_blab(session: Session, kind: str, description: str,
+                    user_id: int = 0, organization_id: int = 1,
+                    advice: str = "", previous: dict | None = None) -> dict:
+    """大白话 → 结构化草稿。无可用模型时抛 ValueError（前端显示可读提示）。
+    advice+previous：人工未确认时的"按建议重新生成"通道——意见最高优先级注入。"""
+    description = str(description or "").strip()
+    if not description:
+        raise ValueError("请先用一段话描述你想要的东西")
+    if len(description) < 5:
+        raise ValueError("描述太短了，多说几句你想要它干什么、风格是什么")
+    if kind not in {"agent", "skill", "template"}:
+        raise ValueError("不支持的资源类型")
+    model = _pick_model(session)
+    if not model:
+        raise ValueError("还没有可用的模型，请先到「模型管理」添加 Provider")
+    provider_id, model_name = model
+    hits, kb_context = retrieve_kb_context(session, description, kind)
+    user_msg = f"类型：{kind}\n用户说：{description[:2000]}"
+    if advice:
+        user_msg += f"\n\n【人工修改意见（最高优先级，必须逐条落实，输出修改后的完整草稿）】\n{advice[:1500]}"
+    if previous:
+        import json as _json
+        user_msg += "\n\n【上一版草稿（在其基础上按意见修改）】\n" + _json.dumps(previous, ensure_ascii=False)[:2500]
+    if kb_context:
+        user_msg += ("\n\n【系统知识库相关内容（供你理解业务语境，草稿的规则/指令应贴合这些真实知识）】\n"
+                     + kb_context)
+    result = chat_json(
+        session,
+        [{"role": "system", "content": DRAFT_SYSTEM_PROMPT},
+         {"role": "user", "content": user_msg}],
+        provider_id, model_name,
+        agent_name="AI 辅助创建", user_id=user_id, organization_id=organization_id,
+        max_retries=1, temperature=0.4,
+    )
+    if result.get("status") != "ok" or not isinstance(result.get("json"), dict):
+        raise ValueError(f"AI 没能理解这段描述：{str(result.get('error'))[:160]}")
+    draft = result["json"]
+    draft["kind"] = kind
+    # 模型信息回传，前端展示这次草稿花了多少
+    draft["_meta"] = {"provider": result.get("provider"), "model": result.get("model"),
+                      "tokens": result.get("tokens"), "duration_ms": result.get("duration_ms"),
+                      "kb_hits": [{"kb_id": h.get("kb_id"), "kb_name": h.get("kb_name"),
+                                   "file": str((h.get("meta") or {}).get("file") or ""),
+                                   "excerpt": re.sub(r"\s+", " ", str(h.get("text") or ""))[:80]}
+                                  for h in hits]}
+    return draft
+
+
+# ---------------- 冲突检测 ----------------
+
+def _norm(text: str) -> str:
+    return re.sub(r"[\s，。、！？：；「」\-—_]", "", str(text or "")).lower()
+
+
+def _similar(a: str, b: str) -> float:
+    return SequenceMatcher(None, _norm(a), _norm(b)).ratio()
+
+
+def _find_mutex_clauses(text: str, known_conflicts: list[str]) -> list[str]:
+    """在一段文本里找互斥表述。返回「A ↔ B」列表。"""
+    found: list[str] = []
+    norm = _norm(text)
+    for left, right in MUTUAL_EXCLUSIVE_PAIRS:
+        l, r = _norm(left), _norm(right)
+        if l in norm and r in norm:
+            pair = f"{left} ↔ {right}"
+            if pair not in known_conflicts:
+                found.append(pair)
+    return found
+
+
+def _single_side_hits(text: str) -> set[str]:
+    """文本里出现的互斥词（任一侧），用于跨资源单侧冲突比对。"""
+    norm = _norm(text)
+    hits: set[str] = set()
+    for left, right in MUTUAL_EXCLUSIVE_PAIRS:
+        if _norm(left) in norm:
+            hits.add(_norm(left))
+        if _norm(right) in norm:
+            hits.add(_norm(right))
+    return hits
+
+
+def conflict_scan(session: Session, kind: str, payload: dict,
+                  exclude_id: int = 0) -> dict:
+    """扫描与现有资源的冲突。返回 {conflicts: [...], warnings: [...]}。
+
+    conflicts 是应该处理的（重名），warnings 是提醒用户判断的（相似名/互斥表述）。
+    """
+    conflicts: list[str] = []
+    warnings: list[str] = []
+    name = str(payload.get("name") or "").strip()
+    norm_name = _norm(name)
+    if not norm_name:
+        return {"conflicts": ["名称不能为空"], "warnings": []}
+
+    # 1. 同类重名 → conflict
+    if kind == "agent":
+        rows = [(a.id, a.name, a) for a in session.exec(select(Agent)).all()]
+    elif kind == "skill":
+        rows = [(sk.id, sk.name, sk) for sk in session.exec(select(AgentSkill)).all()]
+    else:
+        rows = [(t.id, t.name, t) for t in session.exec(select(AgentPromptTemplate)).all()]
+    for row_id, row_name, _row in rows:
+        if row_id == exclude_id:
+            continue
+        if _norm(row_name) == norm_name:
+            conflicts.append(f"已存在同名的{'Agent' if kind == 'agent' else '技能' if kind == 'skill' else '提示模板'}「{row_name}」，换一个名字")
+            continue
+        if _similar(name, row_name) >= 0.82:
+            warnings.append(f"与现有「{row_name}」名字很像（相似度 {int(_similar(name, row_name) * 100)}%），确认不是重复创建？")
+
+    # 2. 互斥表述：收集本资源的所有文字
+    own_texts: list[str] = [name, str(payload.get("desc") or "")]
+    if kind == "agent":
+        own_texts += [str(item) for item in (payload.get("rules") or [])]
+    elif kind == "skill":
+        own_texts += [str(item) for item in (payload.get("instructions") or [])]
+    else:
+        own_texts += [str(payload.get("role_prompt") or ""), str(payload.get("task_template") or ""),
+                      str(payload.get("style") or "")]
+        own_texts += [str(item) for item in (payload.get("constraints") or [])]
+    own_conflicts: list[str] = []
+    for text in own_texts:
+        for pair in _find_mutex_clauses(text, own_conflicts):
+            own_conflicts.append(pair)
+    if own_conflicts:
+        warnings.append("这段描述里有互相打架的要求：" + "；".join(f"「{pair}」" for pair in own_conflicts)
+                        + "，AI 不知道该听哪个")
+
+    # 3. Agent 与挂接技能的冲突（创建/编辑 Agent 时）
+    if kind == "agent":
+        skill_ids = [int(item) for item in (payload.get("skill_ids") or []) if str(item).isdigit()]
+        for skill_id in skill_ids:
+            skill = session.get(AgentSkill, skill_id)
+            if not skill:
+                continue
+            skill_text = skill.instructions
+            own_pairs: set[tuple[str, str]] = set()
+            for text in own_texts:
+                own_pairs.update(_find_mutex_clauses(text, []))
+            skill_pairs = set(_find_mutex_clauses(skill_text, []))
+            # 双侧都在（两边各自出现完整互斥对）→ 直接冲突
+            for l, r in own_pairs:
+                for l2, r2 in skill_pairs:
+                    if _norm(l) in (_norm(l2), _norm(r2)) or _norm(r) in (_norm(l2), _norm(r2)):
+                        warnings.append(f"与挂接的技能「{skill.name}」要求相反：{l} ↔ {r}，生成时会互相拉扯")
+                        break
+            # 单侧词冲突：本资源只说「口语」，技能只说「书面」→ 两边拉扯
+            own_sides: set[str] = set()
+            for text in own_texts:
+                own_sides.update(_single_side_hits(text))
+            skill_sides = _single_side_hits(skill_text)
+            for l, r in MUTUAL_EXCLUSIVE_PAIRS:
+                if _norm(l) in own_sides and _norm(r) in skill_sides:
+                    warnings.append(f"与挂接的技能「{skill.name}」要求相反：{l} ↔ {r}，生成时会互相拉扯")
+                elif _norm(r) in own_sides and _norm(l) in skill_sides:
+                    warnings.append(f"与挂接的技能「{skill.name}」要求相反：{r} ↔ {l}，生成时会互相拉扯")
+
+    # 4. Agent 与其提示模板的约束冲突
+    if kind == "agent":
+        template_id = payload.get("prompt_template_id")
+        if template_id and str(template_id).isdigit():
+            template = session.get(AgentPromptTemplate, int(template_id))
+            if template:
+                template_texts = [template.role_prompt, template.task_template, template.style] + list(template.constraints or [])
+                template_sides: set[str] = set()
+                for text in template_texts:
+                    template_sides.update(_single_side_hits(text))
+                own_sides_t: set[str] = set()
+                for text in own_texts:
+                    own_sides_t.update(_single_side_hits(text))
+                for l, r in MUTUAL_EXCLUSIVE_PAIRS:
+                    if _norm(l) in own_sides_t and _norm(r) in template_sides:
+                        warnings.append(f"与提示模板「{template.name}」的约束相反：{l} ↔ {r}，模板会跟规则打架")
+                    elif _norm(r) in own_sides_t and _norm(l) in template_sides:
+                        warnings.append(f"与提示模板「{template.name}」的约束相反：{r} ↔ {l}，模板会跟规则打架")
+
+    return {"conflicts": conflicts, "warnings": warnings}
+
+
+def conflict_scan_text(session: Session, kind: str, payload: dict, exclude_id: int = 0) -> dict:
+    """conflict_scan 的文本摘要版，供 LLM 草稿后立即回检。"""
+    return conflict_scan(session, kind, payload, exclude_id)
+
+
+__all__ = ["draft_from_blab", "conflict_scan", "conflict_scan_text",
+           "MUTUAL_EXCLUSIVE_PAIRS", "DRAFT_SYSTEM_PROMPT"]