litellm-rs 0.8.2

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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150

use crate::core::models::openai::{
    ChatCompletionRequest, ChatCompletionResponse, ChatMessage, ContentPart, MessageContent,
    MessageRole,
};
use crate::core::providers::ChatMessageContinuation;
use crate::utils::error::gateway_error::GatewayError;

pub(super) fn guardrail_response_with_continuation(
    response: &ChatCompletionResponse,
    choice_extensions: &[ChatMessageContinuation],
) -> Result<ChatCompletionResponse, GatewayError> {
    if response.choices.len() != choice_extensions.len() {
        return Err(GatewayError::internal(format!(
            "chat choice extensions length mismatch: expected {}, got {}",
            response.choices.len(),
            choice_extensions.len()
        )));
    }
    let mut projected = response.clone();
    for (choice, extension) in projected.choices.iter_mut().zip(choice_extensions) {
        append_guardrail_visible_thinking(&mut choice.message, extension);
    }
    Ok(projected)
}

pub(super) fn guardrail_request_with_continuation(
    request: &ChatCompletionRequest,
    message_extensions: &[ChatMessageContinuation],
) -> Result<ChatCompletionRequest, GatewayError> {
    if request.messages.len() != message_extensions.len() {
        return Err(GatewayError::validation(format!(
            "chat message extensions length mismatch: expected {}, got {}",
            request.messages.len(),
            message_extensions.len()
        )));
    }
    let mut projected = request.clone();
    for (message, extension) in projected.messages.iter_mut().zip(message_extensions) {
        extension.validate().map_err(GatewayError::validation)?;
        if !extension.is_empty() && message.role != MessageRole::Assistant {
            return Err(GatewayError::validation(
                "Anthropic continuation is only valid on assistant messages",
            ));
        }
        append_guardrail_visible_thinking(message, extension);
    }
    Ok(projected)
}

pub(super) fn continuation_after_input_projection(
    original: &ChatCompletionRequest,
    projected: &ChatCompletionRequest,
    extensions: Vec<ChatMessageContinuation>,
) -> Result<Vec<ChatMessageContinuation>, GatewayError> {
    if original.messages.len() != projected.messages.len()
        || original.messages.len() != extensions.len()
    {
        return Err(GatewayError::validation(
            "chat message continuation length changed during guardrail projection",
        ));
    }

    original
        .messages
        .iter()
        .zip(&projected.messages)
        .zip(extensions)
        .map(|((before, after), extension)| {
            sanitize_continuation_after_content_projection(
                &before.content,
                &after.content,
                extension,
            )
        })
        .collect()
}

pub(super) fn continuation_after_output_projection(
    original: &ChatCompletionResponse,
    projected: &ChatCompletionResponse,
    extensions: Vec<ChatMessageContinuation>,
) -> Result<Vec<ChatMessageContinuation>, GatewayError> {
    if original.choices.len() != projected.choices.len()
        || original.choices.len() != extensions.len()
    {
        return Err(GatewayError::internal(
            "chat choice continuation length changed during guardrail projection",
        ));
    }

    original
        .choices
        .iter()
        .zip(&projected.choices)
        .zip(extensions)
        .map(|((before, after), extension)| {
            sanitize_continuation_after_content_projection(
                &before.message.content,
                &after.message.content,
                extension,
            )
        })
        .collect()
}

fn sanitize_continuation_after_content_projection(
    before: &Option<MessageContent>,
    after: &Option<MessageContent>,
    extension: ChatMessageContinuation,
) -> Result<ChatMessageContinuation, GatewayError> {
    let before = serde_json::to_value(before).map_err(|cause| {
        GatewayError::internal(format!(
            "failed to compare pre-guardrail continuation content: {cause}"
        ))
    })?;
    let after = serde_json::to_value(after).map_err(|cause| {
        GatewayError::internal(format!(
            "failed to compare post-guardrail continuation content: {cause}"
        ))
    })?;
    Ok(if before == after {
        extension
    } else {
        extension.without_anthropic_block_order()
    })
}

fn append_guardrail_visible_thinking(
    message: &mut ChatMessage,
    extension: &ChatMessageContinuation,
) {
    let Some(visible) = extension
        .anthropic_thinking()
        .and_then(|thinking| thinking.as_text())
    else {
        return;
    };
    match &mut message.content {
        Some(MessageContent::Text(content)) => {
            if !content.is_empty() {
                content.push('\n');
            }
            content.push_str(&visible);
        }
        Some(MessageContent::Parts(parts)) => parts.push(ContentPart::Text {
            text: visible.into_owned(),
        }),
        None => message.content = Some(MessageContent::Text(visible.into_owned())),
    }
}