rig-core 0.45.0

An opinionated library for building LLM powered applications.
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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187

use super::*;

/// The two live shapes, and the catch-all that keeps a third from failing
/// the transcription. Recorded turns of the first two are replayed in
/// `transcription_usage_matrix`; this pins the decode itself, including
/// the shapes no live model produces.
#[test]
fn usage_decodes_both_billing_shapes_and_keeps_unknown_ones() {
    fn usage(body: &str) -> Option<TranscriptionUsage> {
        serde_json::from_str::<TranscriptionResponse>(body)
            .expect("response should decode")
            .usage
    }

    assert_eq!(
        usage(r#"{"text":"hi","usage":{"type":"duration","seconds":6}}"#),
        Some(TranscriptionUsage::Duration {
            r#type: DurationTag::Duration,
            seconds: 6.0
        })
    );
    assert_eq!(
        usage(
            r#"{"text":"hi","usage":{"type":"tokens","input_tokens":54,
                   "input_token_details":{"audio_tokens":54,"text_tokens":0},
                   "output_tokens":16,"total_tokens":70}}"#
        ),
        Some(TranscriptionUsage::Tokens {
            r#type: TokensTag::Tokens,
            input_tokens: 54,
            input_token_details: Some(TranscriptionInputTokenDetails {
                audio_tokens: 54,
                text_tokens: 0,
            }),
            output_tokens: 16,
            total_tokens: 70,
        })
    );
    // The breakdown is optional: a provider that omits it still decodes as
    // a token-billed turn rather than falling to the catch-all.
    assert_eq!(
        usage(
            r#"{"text":"hi","usage":{"type":"tokens","input_tokens":54,
                   "output_tokens":16,"total_tokens":70}}"#
        ),
        Some(TranscriptionUsage::Tokens {
            r#type: TokensTag::Tokens,
            input_tokens: 54,
            input_token_details: None,
            output_tokens: 16,
            total_tokens: 70,
        })
    );
    // The tag decides, not which optional keys are present: a token-billed
    // payload that also reported `seconds` must not decode as a duration
    // and drop every token count.
    assert!(matches!(
        usage(
            r#"{"text":"hi","usage":{"type":"tokens","seconds":6,"input_tokens":54,
                   "output_tokens":16,"total_tokens":70}}"#
        ),
        Some(TranscriptionUsage::Tokens {
            total_tokens: 70,
            ..
        })
    ));
    assert!(matches!(
        usage(r#"{"text":"hi","usage":{"type":"credits","spent":3}}"#),
        Some(TranscriptionUsage::Other(_))
    ));
    // A token-shaped payload missing a required total degrades to the
    // catch-all rather than failing the transcription.
    assert!(matches!(
        usage(r#"{"text":"hi","usage":{"type":"tokens","input_tokens":54}}"#),
        Some(TranscriptionUsage::Other(_))
    ));
    assert_eq!(usage(r#"{"text":"hi"}"#), None);
    assert_eq!(usage(r#"{"text":"hi","usage":null}"#), None);
}

/// The duration-billed shape reports no token counts, so normalization must
/// leave the usage empty rather than inventing zeros; the token-billed shape
/// carries its counts across.
#[test]
fn normalize_carries_token_billing_and_leaves_duration_billing_empty() {
    fn normalized(body: &str) -> crate::completion::Usage {
        serde_json::from_str::<TranscriptionResponse>(body)
            .expect("response should decode")
            .normalize()
            .expect("normalization is infallible for this wire")
            .usage
    }

    let tokens = normalized(
        r#"{"text":"hi","usage":{"type":"tokens","input_tokens":54,
               "output_tokens":16,"total_tokens":70}}"#,
    );
    assert_eq!(tokens.input_tokens, Some(54));
    assert_eq!(tokens.output_tokens, Some(16));
    assert_eq!(tokens.total_tokens, Some(70));

    let duration = normalized(r#"{"text":"hi","usage":{"type":"duration","seconds":6}}"#);
    assert_eq!(duration.input_tokens, None);
    assert_eq!(duration.output_tokens, None);
    assert_eq!(duration.total_tokens, None);
}

/// Timings and speakers from each verbose shape normalize into the same
/// segments and words. The bodies are trimmed from live replies: OpenRouter
/// labels speakers by index, OpenAI by letter, and Mistral by name under
/// `speaker_id`, with untimed segments allowed.
#[test]
fn verbose_shapes_normalize_to_segments_words_and_speakers() {
    fn normalize(body: &str) -> transcription::TranscriptionResponse {
        serde_json::from_str::<TranscriptionResponse>(body)
            .expect("response should decode")
            .normalize()
            .expect("response should normalize")
    }
    fn segment(start: f64, end: f64, text: &str, speaker: Option<&str>) -> TranscriptSegment {
        TranscriptSegment {
            start,
            end,
            text: text.to_owned(),
            speaker: speaker.map(str::to_owned),
        }
    }

    let openrouter = normalize(
        r#"{"text":"Hello there! It was great.","usage":{"seconds":7,"cost":0.0002},
            "language":"en","duration":6.55,"confidence":0.93,
            "segments":[{"id":0,"start":0.12,"end":2,"text":"Hello there!","speaker":0},
                        {"id":1,"start":2.88,"end":5.96,"text":"It was great.","speaker":1}],
            "words":[{"word":"Hello","start":0.12,"end":0.32,"speaker":0,"confidence":0.9}]}"#,
    );
    assert_eq!(openrouter.language.as_deref(), Some("en"));
    assert_eq!(openrouter.duration, Some(6.55));
    assert_eq!(openrouter.confidence, Some(0.93));
    assert_eq!(
        openrouter.segments,
        vec![
            segment(0.12, 2.0, "Hello there!", Some("0")),
            segment(2.88, 5.96, "It was great.", Some("1")),
        ]
    );
    assert_eq!(
        openrouter.words,
        vec![TranscriptWord {
            word: "Hello".to_owned(),
            start: 0.12,
            end: 0.32,
            speaker: Some("0".to_owned()),
            confidence: Some(0.9),
        }]
    );

    let openai = normalize(
        r#"{"text":"Hello there!","task":"transcribe","duration":6.55,
            "segments":[{"type":"transcript.text.segment","text":" Hello there!",
                         "speaker":"A","start":0.0,"end":0.5,"id":"seg_0"}],
            "usage":{"type":"tokens","total_tokens":293,"input_tokens":65,"output_tokens":228}}"#,
    );
    assert_eq!(
        openai.segments,
        vec![segment(0.0, 0.5, " Hello there!", Some("A"))]
    );
    assert_eq!(openai.usage.total_tokens, Some(293));

    let mistral = normalize(
        r#"{"model":"voxtral-mini-latest","text":"Hello there. It was great.","language":null,
            "segments":[{"type":"transcription_segment","text":"Hello there.","start":0.0,"end":0.7,"speaker_id":"speaker_1"},
                        {"type":"transcription_segment","text":" It was great.","start":2.9,"end":4.2,"speaker_id":null},
                        {"type":"transcription_segment","text":" untimed","start":null,"end":null}],
            "usage":{"prompt_audio_seconds":6,"prompt_tokens":5,"total_tokens":438,"completion_tokens":58}}"#,
    );
    assert_eq!(mistral.language, None);
    assert_eq!(
        mistral.segments,
        vec![
            segment(0.0, 0.7, "Hello there.", Some("speaker_1")),
            segment(2.9, 4.2, " It was great.", None),
        ]
    );

    let plain = normalize(r#"{"text":"hi","segments":null,"words":null}"#);
    assert!(plain.segments.is_empty() && plain.words.is_empty());
}