//! Embedding router tests (EM-01 ~ EM-06)
//!
//! Validates SemanticSkillRouter with embedding, TF-IDF fallback,
//! dimension mismatch handling, empty queries, CJK queries, and LLM fallback.

use std::collections::HashMap;
use std::sync::Arc;
use async_trait::async_trait;
use zclaw_skills::semantic_router::{
    Embedder, NoOpEmbedder, SemanticSkillRouter, RuntimeLlmIntent,
    RoutingResult, ScoredCandidate, cosine_similarity,
};
use zclaw_skills::{SkillRegistry, PromptOnlySkill, SkillManifest, SkillMode};
use zclaw_types::id::SkillId;

fn make_manifest(id: &str, name: &str, triggers: Vec<&str>) -> SkillManifest {
    SkillManifest {
        id: SkillId::new(id),
        name: name.to_string(),
        description: format!("{} description", name),
        version: "1.0.0".to_string(),
        mode: SkillMode::PromptOnly,
        triggers: triggers.into_iter().map(String::from).collect(),
        enabled: true,
        author: None,
        capabilities: Vec::new(),
        input_schema: None,
        output_schema: None,
        tags: Vec::new(),
        category: None,
        tools: Vec::new(),
        body: None,
    }
}

/// Mock embedder that returns fixed 768-dim vectors with variation by text hash.
struct MockEmbedder {
    dim: usize,
    should_fail: bool,
}

impl MockEmbedder {
    fn new(dim: usize) -> Self {
        Self { dim, should_fail: false }
    }
    fn failing() -> Self {
        Self { dim: 768, should_fail: true }
    }
}

#[async_trait]
impl Embedder for MockEmbedder {
    async fn embed(&self, text: &str) -> Option<Vec<f32>> {
        if self.should_fail {
            return None;
        }
        // Deterministic vector based on text content
        let mut vec = vec![0.0f32; self.dim];
        for (i, b) in text.as_bytes().iter().enumerate() {
            vec[i % self.dim] += (*b as f32) / 255.0;
        }
        // Normalize
        let norm: f32 = vec.iter().map(|v| v * v).sum::<f32>().sqrt().max(1e-8);
        for v in vec.iter_mut() {
            *v /= norm;
        }
        Some(vec)
    }
}

/// Helper: register skills and build router with embedding.
async fn build_router_with_skills(
    embedder: Arc<dyn Embedder>,
    skills: Vec<(&str, &str, Vec<&str>)>,
) -> SemanticSkillRouter {
    let registry = Arc::new(SkillRegistry::new());
    for (id, name, triggers) in skills {
        let manifest = make_manifest(id, name, triggers);
        registry
            .register(
                Arc::new(zclaw_skills::PromptOnlySkill::new(
                    manifest.clone(),
                    format!("Execute {}", name),
                )),
                manifest,
            )
            .await;
    }
    let mut router = SemanticSkillRouter::new(registry, embedder);
    router.rebuild_index().await;
    router
}

/// EM-01: Embedding API normal routing with 70/30 hybrid scoring.
#[tokio::test]
async fn em01_embedding_normal_routing() {
    let router = build_router_with_skills(
        Arc::new(MockEmbedder::new(768)),
        vec![
            ("finance", "财务追踪", vec!["财务", "花销", "支出", "账单"]),
            ("scheduling", "排班管理", vec!["排班", "班表", "值班"]),
            ("news", "新闻搜索", vec!["新闻", "资讯", "头条"]),
        ],
    )
    .await;

    let result = router.route("帮我查一下上个月的花销").await;
    assert!(result.is_some(), "should match a skill");
    let r = result.unwrap();
    assert_eq!(r.skill_id, "finance", "should match finance skill");
    assert!(
        r.confidence > 0.1,
        "confidence should be positive: {}",
        r.confidence
    );
}

/// EM-02: Embedding API failure degrades to TF-IDF.
#[tokio::test]
async fn em02_embedding_failure_fallback_to_tfidf() {
    let router = build_router_with_skills(
        Arc::new(MockEmbedder::failing()),
        vec![
            ("finance", "财务追踪", vec!["财务", "花销"]),
            ("scheduling", "排班管理", vec!["排班", "班表"]),
        ],
    )
    .await;

    // Should still return results via TF-IDF fallback
    let result = router.route("帮我查花销").await;
    assert!(
        result.is_some(),
        "TF-IDF fallback should still produce results"
    );
}

/// EM-03: Embedding dimension mismatch — no panic.
#[tokio::test]
async fn em03_embedding_dimension_mismatch() {
    // Use a mismatched embedder that returns different dimensions
    struct MismatchedEmbedder;
    #[async_trait]
    impl Embedder for MismatchedEmbedder {
        async fn embed(&self, _text: &str) -> Option<Vec<f32>> {
            // Return a small vector — won't match index embeddings
            Some(vec![0.5; 64])
        }
    }

    let router = build_router_with_skills(
        Arc::new(MismatchedEmbedder),
        vec![("finance", "财务追踪", vec!["财务", "花销"])],
    )
    .await;

    // Should not panic
    let result = router.route("查花销").await;
    // May return None or a result via TF-IDF — key assertion: no panic
    let _ = result;
}

/// EM-04: Empty query handling.
#[tokio::test]
async fn em04_empty_query_handling() {
    let router = build_router_with_skills(
        Arc::new(MockEmbedder::new(768)),
        vec![("finance", "财务追踪", vec!["财务"])],
    )
    .await;

    let result = router.route("").await;
    // Empty query may return None or a low-confidence result
    // Key: no panic
    let _ = result;
}

/// EM-05: Pure Chinese CJK query with bigram matching.
#[tokio::test]
async fn em05_cjk_query_matching() {
    let router = build_router_with_skills(
        Arc::new(NoOpEmbedder), // TF-IDF only
        vec![
            ("scheduling", "排班管理", vec!["排班", "班表", "值班"]),
            ("news", "新闻搜索", vec!["新闻"]),
        ],
    )
    .await;

    let result = router.route("我这个月值班表怎么排").await;
    assert!(result.is_some(), "CJK query should match");
    assert_eq!(
        result.unwrap().skill_id,
        "scheduling",
        "should match scheduling skill"
    );
}

/// EM-06: LLM fallback triggered for ambiguous queries.
#[tokio::test]
async fn em06_llm_fallback_triggered() {
    struct MockLlmFallback {
        target: String,
    }

    #[async_trait]
    impl RuntimeLlmIntent for MockLlmFallback {
        async fn resolve_skill(
            &self,
            _query: &str,
            candidates: &[ScoredCandidate],
        ) -> Option<RoutingResult> {
            let c = candidates
                .iter()
                .find(|c| c.manifest.id.as_str() == self.target)?;
            Some(RoutingResult {
                skill_id: c.manifest.id.to_string(),
                confidence: 0.75,
                parameters: serde_json::json!({}),
                reasoning: "LLM selected".to_string(),
            })
        }
    }

    let registry = Arc::new(SkillRegistry::new());
    let manifest = make_manifest("helper", "通用助手", vec!["帮助", "处理"]);
    registry
        .register(
            Arc::new(zclaw_skills::PromptOnlySkill::new(
                manifest.clone(),
                "Help".to_string(),
            )),
            manifest,
        )
        .await;

    let mut router = SemanticSkillRouter::new_tf_idf_only(registry)
        .with_confidence_threshold(100.0) // Force all to be below threshold
        .with_llm_fallback(Arc::new(MockLlmFallback {
            target: "helper".to_string(),
        }));
    router.rebuild_index().await;

    let result = router.route("帮我处理一下那个东西").await;
    assert!(result.is_some(), "LLM fallback should resolve");
    assert_eq!(result.unwrap().skill_id, "helper");
}

/// Bonus: cosine_similarity utility correctness.
#[test]
fn cosine_similarity_identical_vectors() {
    let v = vec![1.0, 0.0, 1.0, 0.0];
    let sim = cosine_similarity(&v, &v);
    assert!((sim - 1.0).abs() < 1e-6, "identical vectors => cosine=1.0");
}

#[test]
fn cosine_similarity_orthogonal_vectors() {
    let a = vec![1.0, 0.0];
    let b = vec![0.0, 1.0];
    let sim = cosine_similarity(&a, &b);
    assert!(sim.abs() < 1e-6, "orthogonal => cosine≈0");
}

#[test]
fn cosine_similarity_mismatched_dimensions() {
    let a = vec![1.0, 0.0, 1.0];
    let b = vec![1.0, 0.0];
    let sim = cosine_similarity(&a, &b);
    assert_eq!(sim, 0.0, "mismatched dimensions => 0.0");
}