绝尘
356c8f27f8
- 新建 df-ideas/adversarial_helpers.rs(366行): 6类型(EvaluatedBy/AdversarialEval/Argument/AnalystAnalysis/AssessmentLevel/Recommendation) + SYSTEM_PROMPT + 8纯fn(build_adversarial_prompt/parse_llm_eval/extract_json/parse_assessment_level/parse_recommendation/priority_label/assessment_desc/action_hint) + LlmEvalRaw中间结构 - adversarial.rs 930→596: AdversarialEngine impl保留 + pub use helpers(外部idea.rs路径零变更) - lib.rs: mod adversarial_helpers(私有) - ARC-260618-01-e(evaluate_with_llm一致性)逻辑保留(待决策, doc标注) 主代兜底(独立 -p df-ideas 避 df-ai): cargo 0 + test 20 strategy: 核心库, 纯函数/类型抽离, evaluate主入口+ARC-e保留 git add指定(df-ideas/*)
597 lines
28 KiB
Rust
597 lines
28 KiB
Rust
//! 对抗式评估系统 — 正反方辩论 + AI 分析师
|
||
//!
|
||
//! 双轨实现：
|
||
//! - **启发式**（默认/降级）：基于评分与内容信号生成正反方论点，稳定有区分度。
|
||
//! - **LLM**（注入 provider 后）：调一次 `complete()` 让论点由 LLM 生成，失败自动降级启发式。
|
||
//!
|
||
//! 评估来源由 [`EvaluatedBy`] 三态标记：`Llm`（LLM 深度评估）/ `Heuristic`（主动选启发式，
|
||
//! 无 provider）/ `HeuristicFallback`（LLM 调用失败降级）。前端可据此显示评估深度标签。
|
||
//!
|
||
//! LLM prompt 构造与 JSON 解析在 F-260614-03 接入：[`AdversarialEngine::evaluate_with_llm`]
|
||
//! 构造三角色辩论 prompt（正方/反方/分析师），调一次 `complete()` 要求返回对齐结构的 JSON，
|
||
//! 解析失败/字段缺失/枚举非法 → `bail` 触发降级（[`AdversarialEngine::evaluate`] 已兜底）。
|
||
//!
|
||
//! 重构（strategy·自底向上）：类型定义 / 常量 / prompt 构造 / JSON 解析等无副作用逻辑抽至
|
||
//! [`adversarial_helpers`]，本文件只保留 [`AdversarialEngine`]（有状态引擎 + evaluate 主入口，
|
||
//! 含 ARC-260618-01-e evaluate_with_llm 一致性待决策逻辑，原样保留）。外部已用路径（
|
||
//! `df_ideas::adversarial::{AdversarialEval, Recommendation, …}`）经 `pub use` 不变。
|
||
|
||
use std::sync::Arc;
|
||
|
||
use anyhow::Result;
|
||
|
||
use df_ai_core::model::ModelConfig;
|
||
use df_ai_core::provider::LlmProvider;
|
||
use df_types::types::Priority;
|
||
use crate::capture::Idea;
|
||
use crate::scoring::IdeaScores;
|
||
|
||
// 类型 / 常量 / 纯函数自 adversarial_helpers 引回，并 re-export：外部调用方经
|
||
// `df_ideas::adversarial::{...}` 路径访问的类型零变更（pub use 同时将符号带入当前
|
||
// 命名空间供 impl/tests 使用，无需额外 use）。AdversarialEngine 本身定义在本文件。
|
||
pub use crate::adversarial_helpers::{
|
||
AdversarialEval, AnalystAnalysis, Argument, AssessmentLevel, EvaluatedBy, Recommendation,
|
||
};
|
||
use crate::adversarial_helpers::{
|
||
SYSTEM_PROMPT, action_hint, assessment_desc, build_adversarial_prompt, parse_llm_eval,
|
||
priority_label,
|
||
};
|
||
|
||
/// 对抗评估引擎
|
||
pub struct AdversarialEngine {
|
||
/// 可选 LLM provider。Some → 优先 LLM 评估（失败降级启发式）；None → 纯启发式。
|
||
/// 构造注入（与 IdeaPromoter::new(policy) 同一模式），批量评估复用同一 provider。
|
||
provider: Option<Arc<dyn LlmProvider>>,
|
||
/// F-01 阶段5: 候选模型池。非空时 evaluate_with_llm 经 select_model_id 路由选模型;
|
||
/// 空(None provider 或未注入池)→ model 留空由 provider impl 回填自身 default_model
|
||
/// (与接入前行为一致,平稳过渡)。
|
||
model_pool: Vec<ModelConfig>,
|
||
}
|
||
|
||
impl AdversarialEngine {
|
||
/// 注入 LLM provider 构造（provider Some 时走 LLM，调用失败自动降级启发式）
|
||
pub fn new(provider: Arc<dyn LlmProvider>) -> Self {
|
||
Self { provider: Some(provider), model_pool: Vec::new() }
|
||
}
|
||
|
||
/// F-01 阶段5: 注入 provider + 候选模型池构造。池非空时 evaluate_with_llm 走路由。
|
||
pub fn with_pool(provider: Arc<dyn LlmProvider>, model_pool: Vec<ModelConfig>) -> Self {
|
||
Self { provider: Some(provider), model_pool }
|
||
}
|
||
|
||
/// 纯启发式构造（无 LLM 配置时的默认模式）
|
||
pub fn heuristic() -> Self {
|
||
Self { provider: None, model_pool: Vec::new() }
|
||
}
|
||
|
||
/// 执行完整的对抗评估（内部按 provider 有无调度 LLM / 启发式，失败降级）
|
||
pub async fn evaluate(&self, idea: &Idea) -> Result<AdversarialEval> {
|
||
match &self.provider {
|
||
Some(p) => match self.evaluate_with_llm(idea, p).await {
|
||
Ok(mut eval) => {
|
||
eval.evaluated_by = EvaluatedBy::Llm;
|
||
Ok(eval)
|
||
}
|
||
Err(e) => {
|
||
// LLM 调用失败/超时/格式异常 → 自动降级启发式，保证前端结构完整返回
|
||
tracing::warn!("LLM 对抗评估失败, 降级到启发式: {e}");
|
||
let mut eval = self.evaluate_heuristic(idea)?;
|
||
eval.evaluated_by = EvaluatedBy::HeuristicFallback;
|
||
Ok(eval)
|
||
}
|
||
},
|
||
None => {
|
||
let mut eval = self.evaluate_heuristic(idea)?;
|
||
eval.evaluated_by = EvaluatedBy::Heuristic;
|
||
Ok(eval)
|
||
}
|
||
}
|
||
}
|
||
|
||
/// LLM 对抗评估（注入 provider 后走此路）。
|
||
///
|
||
/// 构造三角色对抗式辩论 prompt（正方/反方/分析师），调一次 `complete()`，要求 LLM
|
||
/// 返回对齐 [`AdversarialEval`] 结构的 JSON。任一环节失败（HTTP / 非 JSON / 字段
|
||
/// 缺失 / 枚举非法 / 数值越界无法修正）→ `bail` 由 [`Self::evaluate`] 捕获降级启发式。
|
||
///
|
||
/// temperature 取 0.4：低于 0.3 偏机械重复启发式信号，高于 0.5 易发散到无关风险，
|
||
/// 0.4 在「稳定可复现」与「论点多样性」间取得平衡。
|
||
///
|
||
/// ARC-260618-01-e: evaluate_with_llm 返回值一致性未校验（final_score 与
|
||
/// analyst.final_assessment 自洽性等），待产品决策，当前逻辑原样保留不调整。
|
||
async fn evaluate_with_llm(&self, idea: &Idea, provider: &Arc<dyn LlmProvider>) -> Result<AdversarialEval> {
|
||
let prompt = build_adversarial_prompt(idea);
|
||
// F-01 阶段5: 智能路由 — 对抗评估 TaskRequirements(Standard,无工具)。
|
||
// 池非空 → select_model_id 选最优 model_id;池空/无匹配 → 留空由 provider impl
|
||
// 回填自身 default_model(与接入前行为一致,平稳过渡)。
|
||
let eval_req = df_ai::router::TaskRequirements {
|
||
modalities: vec![df_ai_core::model::Modality::Text],
|
||
needs_tool_use: false,
|
||
estimated_context: 0,
|
||
};
|
||
let model = df_ai::router::select_model_id(&eval_req, &self.model_pool).unwrap_or_default();
|
||
let request = df_ai_core::provider::CompletionRequest {
|
||
// 路由命中 → 用 model_id;否则留空让 provider impl 回填自身 default_model。
|
||
// (OpenAICompatProvider::convert_request 在 req.model.is_empty() 时回退 default_model)
|
||
model,
|
||
messages: vec![
|
||
df_ai_core::provider::ChatMessage::system(SYSTEM_PROMPT),
|
||
df_ai_core::provider::ChatMessage::user(prompt),
|
||
],
|
||
temperature: Some(0.4),
|
||
max_tokens: Some(2048),
|
||
stream: false,
|
||
tools: None,
|
||
tool_choice: None,
|
||
reasoning_content: None,
|
||
};
|
||
|
||
let resp = provider
|
||
.complete(request)
|
||
.await
|
||
.map_err(|e| anyhow::anyhow!("LLM complete 调用失败: {e}"))?;
|
||
|
||
// 数值 clamp 由 parse_llm_eval 单点收口（final_score∈[0,10]、confidence∈[0,1]，
|
||
// 均在 parse 内对所有 Ok 路径完成），此处不再重复 clamp（CR-40-1 去冗余）。
|
||
parse_llm_eval(&resp.text, &idea.id)
|
||
}
|
||
|
||
/// 启发式评估（基于评分与内容信号，稳定有区分度）
|
||
fn evaluate_heuristic(&self, idea: &Idea) -> Result<AdversarialEval> {
|
||
// 先做多维评分，作为正反方论点与置信度的依据
|
||
let scores = crate::scoring::ScoringEngine::compute_default(idea);
|
||
|
||
let positive = self.generate_positive_argument(idea, &scores)?;
|
||
let negative = self.generate_negative_argument(idea, &scores)?;
|
||
let analyst = self.analyst_analysis(idea, &scores)?;
|
||
let recommendation = self.recommendation_for(&analyst.final_assessment);
|
||
|
||
Ok(AdversarialEval {
|
||
idea_id: idea.id.clone(),
|
||
positive,
|
||
negative,
|
||
analyst,
|
||
final_score: scores.overall,
|
||
recommendation,
|
||
// 由 evaluate() 调用方按调度路径覆盖（Heuristic / HeuristicFallback）
|
||
evaluated_by: EvaluatedBy::Heuristic,
|
||
})
|
||
}
|
||
|
||
/// 生成正方观点（支持执行）— confidence 由可行性 + 影响力驱动
|
||
/// 注:返回 Result 为后续 LLM 注入失败预留,启发式阶段恒 Ok
|
||
fn generate_positive_argument(&self, idea: &Idea, scores: &IdeaScores) -> Result<Argument> {
|
||
let desc = idea.description.trim();
|
||
let mut evidence = Vec::new();
|
||
evidence.push(format!("优先级：{}", priority_label(&idea.priority)));
|
||
if desc.is_empty() {
|
||
evidence.push("需求待补充（建议补全描述）".to_string());
|
||
} else {
|
||
let head: String = desc.chars().take(60).collect();
|
||
evidence.push(format!("明确需求：{}", head));
|
||
}
|
||
if idea.tags.is_empty() {
|
||
evidence.push("关联领域待界定".to_string());
|
||
} else {
|
||
evidence.push(format!("关联领域：{}", idea.tags.join("、")));
|
||
}
|
||
if scores.impact >= 7.0 {
|
||
evidence.push("业务价值显著，影响面较广".to_string());
|
||
}
|
||
|
||
// 正方置信度:可行性+影响力等权折算到 [0.1, 0.95],满分≈0.95 留质疑余地
|
||
let confidence =
|
||
((scores.feasibility * 0.5 + scores.impact * 0.5) / 10.0).clamp(0.1, 0.95);
|
||
|
||
let reasoning = vec![
|
||
format!("可行性评分 {:.1}/10，路径相对清晰", scores.feasibility),
|
||
format!("影响力评分 {:.1}/10，预期回报可观", scores.impact),
|
||
"整体风险可控，适合推进".to_string(),
|
||
];
|
||
|
||
Ok(Argument {
|
||
thesis: format!("「{}」具备明确价值与可行性，建议优先推进", idea.title),
|
||
evidence,
|
||
reasoning,
|
||
confidence,
|
||
})
|
||
}
|
||
|
||
/// 生成反方观点（反对或谨慎）— 论点基于想法实际缺陷，confidence 随风险上升
|
||
fn generate_negative_argument(&self, idea: &Idea, scores: &IdeaScores) -> Result<Argument> {
|
||
let desc = idea.description.trim();
|
||
let mut evidence = Vec::new();
|
||
if desc.is_empty() {
|
||
evidence.push("描述过于简略，需求边界不清".to_string());
|
||
} else if desc.chars().count() < 50 {
|
||
evidence.push("描述偏短，实现细节尚未论证".to_string());
|
||
}
|
||
if idea.tags.is_empty() {
|
||
evidence.push("缺少标签，影响范围未界定".to_string());
|
||
}
|
||
if scores.feasibility < 6.0 {
|
||
evidence.push(format!("可行性 {:.1}/10 偏低，实现路径存疑", scores.feasibility));
|
||
}
|
||
if matches!(idea.priority, Priority::Low) {
|
||
evidence.push("优先级偏低，可能非当前关键路径".to_string());
|
||
}
|
||
if evidence.is_empty() {
|
||
evidence.push("机会成本需权衡，可能存在更优替代方案".to_string());
|
||
}
|
||
|
||
// 反方强度:feasibility 每降 1 分 +0.04,impact 每降 1 分 +0.03,基线 0.25(满分也保留最低质疑),clamp [0.1, 0.9]
|
||
let confidence = ((10.0 - scores.feasibility) * 0.04 + (10.0 - scores.impact) * 0.03 + 0.25)
|
||
.clamp(0.1, 0.9);
|
||
|
||
let reasoning = vec![
|
||
format!("资源投入与当前综合评分 {:.1} 需匹配", scores.overall),
|
||
"ROI 需进一步验证".to_string(),
|
||
"需评估是否存在更优解".to_string(),
|
||
];
|
||
|
||
Ok(Argument {
|
||
thesis: format!("「{}」需谨慎评估，存在风险与机会成本", idea.title),
|
||
evidence,
|
||
reasoning,
|
||
confidence,
|
||
})
|
||
}
|
||
|
||
/// AI 分析师综合分析 — 评估等级由综合评分决定，优势/劣势按维度动态生成
|
||
fn analyst_analysis(&self, idea: &Idea, scores: &IdeaScores) -> Result<AnalystAnalysis> {
|
||
let final_assessment = match scores.overall {
|
||
x if x >= 7.5 => AssessmentLevel::StrongGo,
|
||
x if x >= 6.0 => AssessmentLevel::Recommended,
|
||
x if x >= 4.5 => AssessmentLevel::Conditional,
|
||
x if x >= 3.0 => AssessmentLevel::Revised,
|
||
_ => AssessmentLevel::Defer,
|
||
};
|
||
|
||
let mut strengths = Vec::new();
|
||
if scores.impact >= 6.0 {
|
||
strengths.push("业务价值明确".to_string());
|
||
}
|
||
if scores.feasibility >= 6.0 {
|
||
strengths.push("技术路径清晰".to_string());
|
||
}
|
||
if scores.urgency >= 7.0 {
|
||
strengths.push("时间窗口合适".to_string());
|
||
}
|
||
if strengths.is_empty() {
|
||
strengths.push("方向值得探索".to_string());
|
||
}
|
||
|
||
let mut weaknesses = Vec::new();
|
||
if scores.feasibility < 6.0 {
|
||
weaknesses.push("可行性论证不足".to_string());
|
||
}
|
||
if idea.description.trim().is_empty() {
|
||
weaknesses.push("需求描述缺失".to_string());
|
||
}
|
||
if scores.urgency < 4.0 {
|
||
weaknesses.push("紧急度偏低，易被搁置".to_string());
|
||
}
|
||
if weaknesses.is_empty() {
|
||
weaknesses.push("资源需求待评估".to_string());
|
||
}
|
||
|
||
// 启发式占位:固定风险模板,与具体想法无关,接 LLM 后改动态生成
|
||
let risks = vec![
|
||
"技术实现难度可能超出预期".to_string(),
|
||
"优先级与资源争夺".to_string(),
|
||
"需求范围蔓延".to_string(),
|
||
];
|
||
|
||
let opportunities = vec![
|
||
"可能形成可复用能力".to_string(),
|
||
"积累技术资产".to_string(),
|
||
];
|
||
|
||
let summary = format!(
|
||
"「{}」综合评分 {:.1}/10，{}。建议{}。",
|
||
idea.title,
|
||
scores.overall,
|
||
assessment_desc(&final_assessment),
|
||
action_hint(&final_assessment)
|
||
);
|
||
|
||
Ok(AnalystAnalysis {
|
||
summary,
|
||
strengths,
|
||
weaknesses,
|
||
risks,
|
||
opportunities,
|
||
final_assessment,
|
||
})
|
||
}
|
||
|
||
/// 评估等级 → 最终建议
|
||
fn recommendation_for(&self, level: &AssessmentLevel) -> Recommendation {
|
||
match level {
|
||
AssessmentLevel::StrongGo => Recommendation::ImmediateAction,
|
||
AssessmentLevel::Recommended => Recommendation::Soon,
|
||
AssessmentLevel::Conditional => Recommendation::WithResources,
|
||
AssessmentLevel::Revised => Recommendation::ResearchMore,
|
||
AssessmentLevel::Defer => Recommendation::Monitor,
|
||
}
|
||
}
|
||
}
|
||
|
||
#[cfg(test)]
|
||
mod tests {
|
||
use super::*;
|
||
use crate::capture::Idea;
|
||
use crate::scoring::ScoringEngine;
|
||
use df_types::types::{IdeaStatus, Priority};
|
||
|
||
fn make_idea(title: &str, desc: &str, priority: Priority, tags: Vec<&str>) -> Idea {
|
||
Idea {
|
||
id: "test-id".to_string(),
|
||
title: title.to_string(),
|
||
description: desc.to_string(),
|
||
status: IdeaStatus::Draft,
|
||
priority,
|
||
scores: None,
|
||
tags: tags.into_iter().map(String::from).collect(),
|
||
source: None,
|
||
related_ids: Vec::new(),
|
||
created_at: chrono::Utc::now(),
|
||
updated_at: chrono::Utc::now(),
|
||
}
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a1_high_score_immediate_action() {
|
||
let desc = "面向用户的核心功能，带来显著增长，大幅提升效率。集成成熟方案，复用已有组件。".repeat(3);
|
||
let idea = make_idea("AI增长引擎", &desc, Priority::Critical, vec!["增长", "核心"]);
|
||
let scores = ScoringEngine::compute_default(&idea);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a1] 高分想法 → 期望 ImmediateAction");
|
||
println!(" scores: feas={:.2} impact={:.2} urg={:.2} overall={:.2}", scores.feasibility, scores.impact, scores.urgency, scores.overall);
|
||
println!(" eval: final_score={:.2} recommendation={:?}", eval.final_score, eval.recommendation);
|
||
println!(" 正方 confidence={:.2} 反方 confidence={:.2}", eval.positive.confidence, eval.negative.confidence);
|
||
assert!(eval.final_score >= 7.5, "final_score 应≥7.5, 实际 {:.2}", eval.final_score);
|
||
assert_eq!(eval.recommendation, Recommendation::ImmediateAction);
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a2_mid_score_soon() {
|
||
let desc = "面向用户的功能，集成已有方案，提升体验".to_string();
|
||
let idea = make_idea("体验优化", &desc, Priority::Medium, vec!["体验"]);
|
||
let scores = ScoringEngine::compute_default(&idea);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a2] 中分想法 → 期望 Soon");
|
||
println!(" scores overall={:.2} eval final_score={:.2} recommendation={:?}", scores.overall, eval.final_score, eval.recommendation);
|
||
assert_eq!(eval.recommendation, Recommendation::Soon);
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a3_low_score_monitor() {
|
||
let desc = "重构迁移大规模分布式重写从零全新架构高并发底层".to_string();
|
||
let idea = make_idea("过度工程", &desc, Priority::Low, vec![]);
|
||
let scores = ScoringEngine::compute_default(&idea);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a3] 低分想法 → 期望 Monitor");
|
||
println!(" scores overall={:.2} eval final_score={:.2} recommendation={:?}", scores.overall, eval.final_score, eval.recommendation);
|
||
assert!(eval.final_score < 3.0, "final_score 应<3.0, 实际 {:.2}", eval.final_score);
|
||
assert_eq!(eval.recommendation, Recommendation::Monitor);
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a4_confidence_ranges() {
|
||
let idea = make_idea("普通想法", "一般描述", Priority::Medium, vec!["标签"]);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a4] confidence 区间校验");
|
||
println!(" 正方={:.2} (应∈[0.1, 0.95]) 反方={:.2} (应∈[0.1, 0.9])", eval.positive.confidence, eval.negative.confidence);
|
||
assert!(eval.positive.confidence >= 0.1 && eval.positive.confidence <= 0.95);
|
||
assert!(eval.negative.confidence >= 0.1 && eval.negative.confidence <= 0.9);
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a5_positive_thesis_contains_title() {
|
||
let idea = make_idea("独家创意", "描述内容", Priority::High, vec![]);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a5] 正方论点含标题");
|
||
println!(" thesis: {}", eval.positive.thesis);
|
||
assert!(eval.positive.thesis.contains("独家创意"), "正方 thesis 应含标题");
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a6_negative_evidence_nonempty() {
|
||
let idea = make_idea("待质疑想法", "短", Priority::Low, vec![]);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a6] 反方证据非空 ({} 条)", eval.negative.evidence.len());
|
||
for (i, e) in eval.negative.evidence.iter().enumerate() {
|
||
println!(" 证据{}: {}", i + 1, e);
|
||
}
|
||
assert!(!eval.negative.evidence.is_empty(), "反方 evidence 不应为空");
|
||
}
|
||
|
||
#[tokio::test]
|
||
async fn a7_final_score_consistency() {
|
||
let desc = "面向用户的核心功能".to_string();
|
||
let idea = make_idea("一致性测试", &desc, Priority::High, vec!["核心"]);
|
||
let scores = ScoringEngine::compute_default(&idea);
|
||
let eval = AdversarialEngine::heuristic().evaluate(&idea).await.unwrap();
|
||
println!("\n[a7] final_score == scores.overall 一致性");
|
||
println!(" scores.overall={:.2} eval.final_score={:.2}", scores.overall, eval.final_score);
|
||
println!(" analyst.summary: {}", eval.analyst.summary);
|
||
assert!((eval.final_score - scores.overall).abs() < 0.001, "final_score 应等于 overall");
|
||
assert!(eval.analyst.summary.contains("一致性测试"), "summary 应含标题");
|
||
}
|
||
|
||
// ────────────────────────────────────────────────────────────
|
||
// LLM 路径测试（F-260614-03）
|
||
// ────────────────────────────────────────────────────────────
|
||
|
||
/// mock LlmProvider：按构造时给定的响应文本回放，仅供 adversarial 单测。
|
||
///
|
||
/// `complete` 返回预设 `text`（或预设错误），`stream` 不被 adversarial 路径调用，
|
||
/// 返回 Err 即可。不引入 futures 依赖。
|
||
struct MockProvider {
|
||
text: String,
|
||
}
|
||
|
||
#[async_trait::async_trait]
|
||
impl LlmProvider for MockProvider {
|
||
async fn complete(
|
||
&self,
|
||
_request: df_ai_core::provider::CompletionRequest,
|
||
) -> anyhow::Result<df_ai_core::provider::CompletionResponse> {
|
||
Ok(df_ai_core::provider::CompletionResponse {
|
||
text: self.text.clone(),
|
||
model: "mock-model".to_string(),
|
||
usage: df_ai_core::provider::TokenUsage::default(),
|
||
tool_calls: None,
|
||
// 对抗评估路径不消费推理内容,留 None 即可(CompletionResponse 新增字段,
|
||
// 老构造点漏填 → 测试编译失败,回归补齐)。
|
||
reasoning_content: None,
|
||
})
|
||
}
|
||
|
||
async fn stream(
|
||
&self,
|
||
_request: df_ai_core::provider::CompletionRequest,
|
||
) -> anyhow::Result<df_ai_core::provider::StreamResult> {
|
||
anyhow::bail!("MockProvider 不支持 stream（adversarial 路径不调用）")
|
||
}
|
||
|
||
fn name(&self) -> &str {
|
||
"mock-model"
|
||
}
|
||
}
|
||
|
||
/// LLM mock 返回结构良好的 JSON → 解析正确，evaluated_by=Llm
|
||
#[tokio::test]
|
||
async fn a8_llm_mock_parse_success() {
|
||
let idea = make_idea(
|
||
"AI 增长引擎",
|
||
"面向用户的核心功能，带来显著增长。集成成熟方案。",
|
||
Priority::Critical,
|
||
vec!["增长"],
|
||
);
|
||
// 一份对齐 schema 的合法 LLM 返回（含 ```json 围栏，验证容错）
|
||
let llm_text = r#"```json
|
||
{
|
||
"positive": {
|
||
"thesis": "该想法价值高，路径清晰，应立即推进",
|
||
"evidence": ["用户增长是当前关键路径", "已有成熟方案可复用"],
|
||
"reasoning": ["可行性高", "ROI 明显"],
|
||
"confidence": 0.9
|
||
},
|
||
"negative": {
|
||
"thesis": "资源投入与替代方案需权衡",
|
||
"evidence": ["机会成本存在"],
|
||
"reasoning": ["需验证更优解"],
|
||
"confidence": 0.35
|
||
},
|
||
"analyst": {
|
||
"summary": "综合双方，价值明确且路径清晰，建议立即推进",
|
||
"strengths": ["业务价值明确", "技术路径清晰"],
|
||
"weaknesses": ["资源需求待评估"],
|
||
"risks": ["需求范围蔓延"],
|
||
"opportunities": ["可形成可复用能力"],
|
||
"final_assessment": "StrongGo"
|
||
},
|
||
"final_score": 8.5,
|
||
"recommendation": "ImmediateAction"
|
||
}
|
||
```"#;
|
||
let provider = Arc::new(MockProvider { text: llm_text.to_string() });
|
||
let engine = AdversarialEngine::new(provider);
|
||
let eval = engine.evaluate(&idea).await.unwrap();
|
||
|
||
println!("\n[a8] LLM mock 成功路径");
|
||
println!(" evaluated_by = {:?}", eval.evaluated_by);
|
||
println!(" final_score = {:.2} recommendation = {:?}", eval.final_score, eval.recommendation);
|
||
println!(" positive.confidence = {:.2} negative.confidence = {:.2}",
|
||
eval.positive.confidence, eval.negative.confidence);
|
||
println!(" analyst.final_assessment = {:?}", eval.analyst.final_assessment);
|
||
|
||
assert_eq!(eval.evaluated_by, EvaluatedBy::Llm, "成功路径应为 Llm");
|
||
assert!((eval.final_score - 8.5).abs() < 1e-9);
|
||
assert_eq!(eval.recommendation, Recommendation::ImmediateAction);
|
||
assert_eq!(eval.analyst.final_assessment, AssessmentLevel::StrongGo);
|
||
assert!((eval.positive.confidence - 0.9).abs() < 1e-9);
|
||
assert!((eval.negative.confidence - 0.35).abs() < 1e-9);
|
||
assert_eq!(eval.positive.evidence.len(), 2);
|
||
assert!(eval.positive.thesis.contains("立即推进"));
|
||
}
|
||
|
||
/// LLM mock 返回非法 JSON → bail → 降级 HeuristicFallback（evaluated_by 标记）
|
||
#[tokio::test]
|
||
async fn a9_llm_mock_bad_json_fallback() {
|
||
let idea = make_idea("坏想法", "测试降级", Priority::Medium, vec![]);
|
||
let provider = Arc::new(MockProvider { text: "这不是 JSON".to_string() });
|
||
let engine = AdversarialEngine::new(provider);
|
||
let eval = engine.evaluate(&idea).await.unwrap();
|
||
|
||
println!("\n[a9] LLM mock 非 JSON → 降级");
|
||
println!(" evaluated_by = {:?} (期望 HeuristicFallback)", eval.evaluated_by);
|
||
assert_eq!(eval.evaluated_by, EvaluatedBy::HeuristicFallback, "非 JSON 应降级");
|
||
}
|
||
|
||
/// LLM mock 返回枚举非法值 → bail → 降级
|
||
#[tokio::test]
|
||
async fn a10_llm_mock_bad_enum_fallback() {
|
||
let idea = make_idea("枚举非法", "测试枚举映射", Priority::High, vec![]);
|
||
let llm_text = r#"{
|
||
"positive": {"thesis":"x","evidence":[],"reasoning":[],"confidence":0.5},
|
||
"negative": {"thesis":"y","evidence":[],"reasoning":[],"confidence":0.5},
|
||
"analyst": {"summary":"s","strengths":[],"weaknesses":[],"risks":[],"opportunities":[],"final_assessment":"GoNow"},
|
||
"final_score": 5.0,
|
||
"recommendation": "Soon"
|
||
}"#;
|
||
let provider = Arc::new(MockProvider { text: llm_text.to_string() });
|
||
let engine = AdversarialEngine::new(provider);
|
||
let eval = engine.evaluate(&idea).await.unwrap();
|
||
|
||
println!("\n[a10] LLM mock 枚举非法 → 降级");
|
||
println!(" evaluated_by = {:?} (期望 HeuristicFallback)", eval.evaluated_by);
|
||
assert_eq!(eval.evaluated_by, EvaluatedBy::HeuristicFallback);
|
||
}
|
||
|
||
/// 数值越界 → clamp（confidence>1 / final_score>10）应被截断，不触发降级
|
||
#[tokio::test]
|
||
async fn a11_llm_mock_value_clamp() {
|
||
let idea = make_idea("越界测试", "数值超出区间", Priority::Medium, vec![]);
|
||
let llm_text = r#"{
|
||
"positive": {"thesis":"p","evidence":[],"reasoning":[],"confidence":1.5},
|
||
"negative": {"thesis":"n","evidence":[],"reasoning":[],"confidence":-0.3},
|
||
"analyst": {"summary":"s","strengths":[],"weaknesses":[],"risks":[],"opportunities":[],"final_assessment":"Recommended"},
|
||
"final_score": 99.0,
|
||
"recommendation": "Soon"
|
||
}"#;
|
||
let provider = Arc::new(MockProvider { text: llm_text.to_string() });
|
||
let engine = AdversarialEngine::new(provider);
|
||
let eval = engine.evaluate(&idea).await.unwrap();
|
||
|
||
println!("\n[a11] LLM mock 数值越界 clamp");
|
||
println!(" positive.confidence={:.2} (原 1.5 → 期望 1.0)", eval.positive.confidence);
|
||
println!(" negative.confidence={:.2} (原 -0.3 → 期望 0.0)", eval.negative.confidence);
|
||
println!(" final_score={:.2} (原 99 → 期望 10.0)", eval.final_score);
|
||
|
||
assert_eq!(eval.evaluated_by, EvaluatedBy::Llm, "仅越界不降级");
|
||
assert!((eval.positive.confidence - 1.0).abs() < 1e-9, "confidence>1 应 clamp 到 1.0");
|
||
assert!((eval.negative.confidence - 0.0).abs() < 1e-9, "confidence<0 应 clamp 到 0.0");
|
||
assert!((eval.final_score - 10.0).abs() < 1e-9, "final_score>10 应 clamp 到 10.0");
|
||
}
|
||
|
||
/// LLM 返回「前缀文字 + ```json + {...} + ```」（围栏不在开头）→ extract_json
|
||
/// 正则兜底提取首个 JSON 对象 → 解析成功不降级（CR-40-2 兜底路径）。
|
||
#[tokio::test]
|
||
async fn a12_llm_mock_fenced_not_at_start() {
|
||
let idea = make_idea("围栏不在开头", "测试正则兜底", Priority::High, vec![]);
|
||
let llm_text = "好的，以下是我的评估：\n```json\n{\n \"positive\": {\"thesis\":\"p\",\"evidence\":[],\"reasoning\":[],\"confidence\":0.7},\n \"negative\": {\"thesis\":\"n\",\"evidence\":[],\"reasoning\":[],\"confidence\":0.4},\n \"analyst\": {\"summary\":\"s\",\"strengths\":[],\"weaknesses\":[],\"risks\":[],\"opportunities\":[],\"final_assessment\":\"Recommended\"},\n \"final_score\": 7.0,\n \"recommendation\": \"Soon\"\n}\n```\n希望对你有帮助。";
|
||
let provider = Arc::new(MockProvider { text: llm_text.to_string() });
|
||
let engine = AdversarialEngine::new(provider);
|
||
let eval = engine.evaluate(&idea).await.unwrap();
|
||
|
||
println!("\n[a12] 围栏不在开头 → 正则兜底提取");
|
||
println!(" evaluated_by = {:?} (期望 Llm，不应降级)", eval.evaluated_by);
|
||
assert_eq!(eval.evaluated_by, EvaluatedBy::Llm, "围栏不在开头应正则兜底解析成功");
|
||
assert!((eval.final_score - 7.0).abs() < 1e-9);
|
||
assert_eq!(eval.recommendation, Recommendation::Soon);
|
||
}
|
||
}
|