atomr-infer-runtime-anthropic 0.8.0

Anthropic Messages API provider for atomr-infer — implements ModelRunner against api.anthropic.com, with cost table, error classification, and SSE streaming.
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

use serde::Deserialize;

use atomr_infer_core::error::InferenceError;
use atomr_infer_core::runtime::ProviderKind;
use atomr_infer_remote_core::classify::{classify_http_status, parse_retry_after};

#[derive(Debug, Deserialize)]
struct ErrorEnvelope {
    #[serde(rename = "type", default)]
    kind: String,
    error: ErrorBody,
}

#[derive(Debug, Deserialize)]
struct ErrorBody {
    #[serde(rename = "type", default)]
    kind: String,
    #[serde(default)]
    message: String,
}

pub fn classify_anthropic_error(
    status: u16,
    retry_after_header: Option<&str>,
    body: Option<String>,
) -> InferenceError {
    let retry_after = parse_retry_after(retry_after_header);

    if let Some(body_str) = body.as_deref() {
        if let Ok(env) = serde_json::from_str::<ErrorEnvelope>(body_str) {
            // Anthropic uses `invalid_request_error` for length/filter
            // shapes too; differentiate by the inner type.
            match env.error.kind.as_str() {
                "invalid_request_error"
                    if env.error.message.to_lowercase().contains("context length")
                        || env.error.message.to_lowercase().contains("too long") =>
                {
                    return InferenceError::ContextLengthExceeded {
                        tokens: 0,
                        max_tokens: 0,
                    };
                }
                "permission_error" => {
                    return InferenceError::Forbidden {
                        message: env.error.message,
                    };
                }
                _ => {}
            }
            // Some refusals come back as `error.type = "overloaded_error"` with
            // 529; mapped via status. Content filter is per-message in stream
            // events, not in the envelope.
            let _ = env.kind;
        }
    }

    classify_http_status(ProviderKind::Anthropic, status, retry_after, body)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn maps_context_length() {
        let body = r#"{"type":"error","error":{"type":"invalid_request_error","message":"prompt too long"}}"#;
        let e = classify_anthropic_error(400, None, Some(body.into()));
        assert!(matches!(e, InferenceError::ContextLengthExceeded { .. }));
    }
}