Skip to main content

litellm_rust/
registry.rs

1use serde_json::Value;
2use std::collections::{HashMap, HashSet};
3use std::sync::OnceLock;
4
5use crate::error::{LiteLLMError, Result};
6
7#[derive(Debug, Clone)]
8pub struct ModelPricing {
9    pub input_cost_per_1k: Option<f64>,
10    pub output_cost_per_1k: Option<f64>,
11    pub max_input_tokens: Option<u32>,
12    pub max_output_tokens: Option<u32>,
13    pub mode: Option<String>,
14    pub provider: Option<String>,
15}
16
17#[derive(Debug, Clone)]
18pub struct Registry {
19    pub models: HashMap<String, ModelPricing>,
20}
21
22impl Registry {
23    pub fn load_embedded() -> Result<Self> {
24        Self::from_json_str(include_str!("../data/model_prices_and_context_window.json"))
25    }
26
27    pub fn from_json_str(raw: &str) -> Result<Self> {
28        let json: Value = serde_json::from_str(raw)
29            .map_err(|e| LiteLLMError::Parse(format!("model registry: {e}")))?;
30        let mut models = HashMap::new();
31        let map = json
32            .as_object()
33            .ok_or_else(|| LiteLLMError::Parse("model registry not an object".into()))?;
34        for (name, entry) in map {
35            if name == "sample_spec" {
36                continue;
37            }
38            if let Some(obj) = entry.as_object() {
39                let input = obj
40                    .get("input_cost_per_token")
41                    .and_then(|v| v.as_f64())
42                    .map(|v| v * 1000.0);
43                let output = obj
44                    .get("output_cost_per_token")
45                    .and_then(|v| v.as_f64())
46                    .map(|v| v * 1000.0);
47                let max_input = obj
48                    .get("max_input_tokens")
49                    .or_else(|| obj.get("max_tokens"))
50                    .and_then(|v| v.as_u64())
51                    .map(|v| v as u32);
52                let max_output = obj
53                    .get("max_output_tokens")
54                    .and_then(|v| v.as_u64())
55                    .map(|v| v as u32);
56                let mode = obj
57                    .get("mode")
58                    .and_then(|v| v.as_str())
59                    .map(|s| s.to_string());
60                let provider = obj
61                    .get("litellm_provider")
62                    .and_then(|v| v.as_str())
63                    .map(|s| s.to_string());
64                models.insert(
65                    name.to_string(),
66                    ModelPricing {
67                        input_cost_per_1k: input,
68                        output_cost_per_1k: output,
69                        max_input_tokens: max_input,
70                        max_output_tokens: max_output,
71                        mode,
72                        provider,
73                    },
74                );
75            }
76        }
77        Ok(Self { models })
78    }
79
80    pub fn get(&self, model: &str) -> Option<&ModelPricing> {
81        self.models.get(model)
82    }
83
84    pub fn estimate_cost(&self, model: &str, input_tokens: u32, output_tokens: u32) -> Option<f64> {
85        let pricing = self.models.get(model)?;
86        let input = pricing
87            .input_cost_per_1k
88            .map(|v| v * input_tokens as f64 / 1000.0)?;
89        let output = pricing
90            .output_cost_per_1k
91            .map(|v| v * output_tokens as f64 / 1000.0)?;
92        Some(input + output)
93    }
94}
95
96pub(crate) fn embedded_model_supports_output_config(model: &str) -> bool {
97    static SUPPORTED_MODELS: OnceLock<HashSet<String>> = OnceLock::new();
98
99    let supported = SUPPORTED_MODELS.get_or_init(|| {
100        let raw = include_str!("../data/model_prices_and_context_window.json");
101        let Ok(Value::Object(models)) = serde_json::from_str::<Value>(raw) else {
102            return HashSet::new();
103        };
104        models
105            .into_iter()
106            .filter_map(|(name, metadata)| {
107                metadata
108                    .get("supports_output_config")
109                    .and_then(Value::as_bool)
110                    .unwrap_or(false)
111                    .then(|| name.to_ascii_lowercase().replace('.', "-"))
112            })
113            .collect()
114    });
115
116    supported.contains(&model.to_ascii_lowercase().replace('.', "-"))
117}
118
119#[cfg(test)]
120mod tests {
121    use super::*;
122
123    #[test]
124    fn loads_registry() {
125        let registry = Registry::load_embedded().unwrap();
126        assert!(!registry.models.is_empty());
127    }
128
129    #[test]
130    fn parses_registry_from_json_string() {
131        let registry = Registry::from_json_str(
132            r#"{
133                "sample_spec": {},
134                "test/model": {
135                    "input_cost_per_token": 0.000001,
136                    "output_cost_per_token": 0.000002,
137                    "max_input_tokens": 4096,
138                    "max_output_tokens": 1024,
139                    "mode": "chat",
140                    "litellm_provider": "test"
141                }
142            }"#,
143        )
144        .unwrap();
145
146        let model = registry.get("test/model").unwrap();
147        assert_eq!(model.input_cost_per_1k, Some(0.001));
148        assert_eq!(model.output_cost_per_1k, Some(0.002));
149        assert_eq!(model.max_input_tokens, Some(4096));
150        assert_eq!(model.max_output_tokens, Some(1024));
151    }
152
153    #[test]
154    fn embedded_registry_retains_metadata_coverage() {
155        let registry = Registry::load_embedded().unwrap();
156        let priced_models = registry
157            .models
158            .values()
159            .filter(|model| model.output_cost_per_1k.is_some())
160            .count();
161        let models_with_context = registry
162            .models
163            .values()
164            .filter(|model| model.max_input_tokens.is_some())
165            .count();
166
167        assert!(
168            priced_models > 1000,
169            "expected broad completion-pricing coverage, found {priced_models} models"
170        );
171        assert!(
172            models_with_context > 1000,
173            "expected broad context-window coverage, found {models_with_context} models"
174        );
175    }
176
177    #[test]
178    fn embedded_capabilities_drive_anthropic_output_config_support() {
179        assert!(embedded_model_supports_output_config("claude-sonnet-5"));
180        assert!(embedded_model_supports_output_config("claude-opus-4.8"));
181        assert!(!embedded_model_supports_output_config(
182            "claude-3-7-sonnet-20250219"
183        ));
184        assert!(!embedded_model_supports_output_config(
185            "claude-3-5-sonnet-20241022"
186        ));
187    }
188}