litellm-rs 0.6.0

A high-performance AI Gateway written in Rust, providing OpenAI-compatible APIs with intelligent routing, load balancing, and enterprise features
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133

use super::*;
use crate::core::budget::{ProviderLimitConfig, ResetPeriod};
use crate::core::cost::calculator::estimate_cost;
use crate::core::models::openai::requests::ChatCompletionRequest;
use crate::core::models::openai::{ChatMessage, ContentPart, MessageContent, MessageRole};

fn reserve_with_provider_limit(
    provider: &str,
    model: &str,
    max_output_tokens: u32,
) -> UnifiedBudgetReservation {
    let budget = UnifiedBudgetLimits::new();
    let messages = vec![ChatMessage {
        role: MessageRole::User,
        content: Some(MessageContent::Text("hello".to_string())),
        name: None,
        function_call: None,
        tool_calls: None,
        tool_call_id: None,
        audio: None,
    }];
    let prompt_tokens = estimate_chat_prompt_tokens(model, &messages, None, None, None, None);
    let estimate = estimate_cost(model, provider, prompt_tokens, Some(max_output_tokens)).unwrap();
    budget.providers.set_provider_limit(
        provider,
        ProviderLimitConfig::new(estimate.max_cost * 2.0, ResetPeriod::Monthly),
    );

    let mut request = ChatCompletionRequest {
        model: model.to_string(),
        messages,
        ..Default::default()
    };
    request.max_completion_tokens = Some(max_output_tokens);
    if provider == "bedrock" {
        request.max_tokens = Some(max_output_tokens);
    }

    let reservation = reserve_chat_completion_budget(&budget, provider, model, &request)
        .unwrap()
        .unwrap();
    assert!((reservation.reserved_amount() - estimate.max_cost).abs() < f64::EPSILON);
    reservation
}

#[test]
fn bedrock_chat_reservation_uses_bedrock_cost_pricing() {
    let reservation = reserve_with_provider_limit("bedrock", "amazon.titan-text-express-v1", 100);
    reservation.cancel();
}

#[test]
fn amazon_nova_chat_reservation_uses_provider_pricing() {
    let reservation = reserve_with_provider_limit("amazon_nova", "amazon.nova-2-lite-v1:0", 10);
    reservation.cancel();
}

#[test]
fn openai_like_prefixed_chat_reservation_uses_provider_pricing() {
    let reservation =
        reserve_with_provider_limit("openai_like", "groq/llama-3.3-70b-versatile", 100);
    reservation.cancel();
}

#[test]
fn openai_chat_reservation_uses_exact_tiktoken_prompt_count() {
    let budget = UnifiedBudgetLimits::new();
    let messages = vec![ChatMessage {
        role: MessageRole::User,
        content: Some(MessageContent::Text("Hello, how are you?".to_string())),
        name: None,
        function_call: None,
        tool_calls: None,
        tool_call_id: None,
        audio: None,
    }];
    let prompt_tokens =
        estimate_chat_prompt_tokens("gpt-3.5-turbo", &messages, None, None, None, None);
    assert_eq!(prompt_tokens, 13);

    let expected = estimate_cost("gpt-3.5-turbo", "openai", prompt_tokens, Some(10))
        .unwrap()
        .max_cost;
    budget.providers.set_provider_limit(
        "openai",
        ProviderLimitConfig::new(expected * 2.0, ResetPeriod::Monthly),
    );
    let mut request = ChatCompletionRequest {
        model: "gpt-3.5-turbo".to_string(),
        messages,
        ..Default::default()
    };
    request.max_tokens = Some(10);

    let reservation = reserve_chat_completion_budget(&budget, "openai", "gpt-3.5-turbo", &request)
        .unwrap()
        .unwrap();

    assert!((reservation.reserved_amount() - expected).abs() < f64::EPSILON);
    reservation.cancel();
}

#[test]
fn chat_prompt_estimate_accounts_for_serialized_tool_parts() {
    let payload = "x".repeat(4_000);
    let messages = vec![ChatMessage {
        role: MessageRole::User,
        content: Some(MessageContent::Parts(vec![
            ContentPart::ToolResult {
                tool_use_id: "call_1".to_string(),
                content: serde_json::json!({ "payload": payload }),
                is_error: None,
            },
            ContentPart::ToolUse {
                id: "call_2".to_string(),
                name: "lookup".to_string(),
                input: serde_json::json!({ "payload": "y".repeat(4_000) }),
            },
        ])),
        name: None,
        function_call: None,
        tool_calls: None,
        tool_call_id: None,
        audio: None,
    }];

    let prompt_tokens = estimate_chat_prompt_tokens("gpt-4o", &messages, None, None, None, None);

    assert!(
        prompt_tokens > 1_800,
        "large tool payloads should add a serialized prompt floor"
    );
}