latch_core/
score.rs

1use serde::{Deserialize, Serialize};
2use std::time::SystemTime;
3
4// ── Observation types ──
5
6#[derive(Clone, Debug)]
7pub struct RequestObservation {
8    pub endpoint_id: String,
9    pub pool_id: String,
10    pub started_at: SystemTime,
11    pub success: bool,
12    pub error: Option<ObservationError>,
13    pub was_retry: bool,
14    pub latency: LatencyBreakdown,
15    pub tokens: TokenStats,
16    pub stream: Option<StreamMetrics>,
17}
18
19#[derive(Clone, Debug)]
20pub struct LatencyBreakdown {
21    pub total_ms: u64,
22    pub ttft_ms: Option<u64>,
23}
24
25#[derive(Clone, Debug, Default)]
26pub struct TokenStats {
27    pub input: u64,
28    pub output: u64,
29}
30
31#[derive(Clone, Debug)]
32pub struct StreamMetrics {
33    pub ttft_ms: u64,
34    pub tokens_per_second: f64,
35    pub max_inter_chunk_ms: u64,
36    pub broken: bool,
37}
38
39#[derive(Clone, Debug)]
40pub enum ObservationError {
41    Timeout,
42    HttpStatus(u16),
43    ConnectionFailed,
44    RateLimited,
45    StreamBroken,
46    EmptyResponse,
47    Truncated,
48    MalformedResponse,
49    Other(String),
50}
51
52// ── Scoring config ──
53
54#[derive(Clone, Debug, Serialize, Deserialize)]
55pub struct ScoreConfig {
56    pub window_size: usize,
57    pub decay_period_secs: u64,
58    pub baseline_score: f32,
59    pub availability_weight: f32,
60    pub latency_weight: f32,
61    pub quality_weight: f32,
62    pub cost_weight: f32,
63    pub good_ttft_ms: u64,
64    pub acceptable_ttft_ms: u64,
65    pub good_tps: f64,
66    pub max_error_rate: f32,
67    pub max_truncation_rate: f32,
68    pub max_empty_response_rate: f32,
69}
70
71impl Default for ScoreConfig {
72    fn default() -> Self {
73        Self {
74            window_size: 100,
75            decay_period_secs: 300,
76            baseline_score: 60.0,
77            availability_weight: 0.35,
78            latency_weight: 0.25,
79            quality_weight: 0.25,
80            cost_weight: 0.15,
81            good_ttft_ms: 500,
82            acceptable_ttft_ms: 2000,
83            good_tps: 50.0,
84            max_error_rate: 0.20,
85            max_truncation_rate: 0.10,
86            max_empty_response_rate: 0.05,
87        }
88    }
89}
90
91// ── Scoring results ──
92
93#[derive(Clone, Debug, Serialize, Deserialize)]
94pub struct EndpointScore {
95    pub endpoint_id: String,
96    pub pool_id: String,
97    pub score: f32,
98    pub tier: ScoreTier,
99    pub observation_count: usize,
100    pub last_updated: SystemTime,
101    pub breakdown: ScoreBreakdown,
102}
103
104#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
105pub enum ScoreTier {
106    Poor,    // < 40
107    Bronze,  // 40-69
108    Silver,  // 70-89
109    Gold,    // >= 90
110}
111
112// Note: ScoreTier ordering: Poor < Bronze < Silver < Gold
113impl PartialOrd for ScoreTier {
114    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
115        Some(self.cmp(other))
116    }
117}
118
119impl Ord for ScoreTier {
120    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
121        let self_val = match self {
122            ScoreTier::Poor => 0,
123            ScoreTier::Bronze => 1,
124            ScoreTier::Silver => 2,
125            ScoreTier::Gold => 3,
126        };
127        let other_val = match other {
128            ScoreTier::Poor => 0,
129            ScoreTier::Bronze => 1,
130            ScoreTier::Silver => 2,
131            ScoreTier::Gold => 3,
132        };
133        self_val.cmp(&other_val)
134    }
135}
136
137#[derive(Clone, Debug, Serialize, Deserialize)]
138pub struct ScoreBreakdown {
139    pub availability: f32,
140    pub latency: f32,
141    pub quality: f32,
142    pub cost: f32,
143    pub penalty: f32,
144}
145
146#[derive(Clone, Debug, Serialize, Deserialize)]
147pub struct RankingResult {
148    pub ranked_endpoints: Vec<EndpointScore>,
149    pub recommended: Option<EndpointScore>,
150    pub recommended_fallback: Option<EndpointScore>,
151    pub excluded: Vec<EndpointScore>,
152}
latch_core/score.rs

latch_core/
score.rs
