switchyard_libsy/core/
classifier.rs

1// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2// SPDX-License-Identifier: Apache-2.0
3
4use crate::core::algorithm::Driver;
5use crate::{LibsyError, Result};
6use async_trait::async_trait;
7use switchyard_protocol::{Request, Response};
8
9/// One classifier's recommendation of a routing `target`, with a `[0.0, 1.0]` confidence.
10#[derive(Debug, Clone, PartialEq)]
11pub struct Score {
12    /// `[0.0, 1.0]` confidence in `target`.
13    pub confidence: f64,
14    /// The target (model / tier) being recommended.
15    pub target: String,
16}
17
18/// A classifier's verdict for a request: a set of target [`Score`]s, flagged by how
19/// confident the classifier is that they are decisive.
20pub enum Classification {
21    /// Definite recommendations; [`argmax`](Self::argmax) always yields the top target.
22    Scores(Vec<Score>),
23    /// Recommendations the classifier considers ambiguous; [`argmax`](Self::argmax) yields
24    /// nothing unless the caller opts to ignore ambiguity.
25    Ambiguous(Vec<Score>),
26}
27
28impl Classification {
29    /// The top-scoring [`Score`], or `None` when the classifier abstained (an empty set).
30    ///
31    /// An [`Ambiguous`](Self::Ambiguous) classification also yields `None` unless
32    /// `ignore_ambiguous` is set, in which case it falls back to the plain argmax.
33    /// Errors if any confidence is `NaN` (an unorderable score the caller should surface).
34    pub fn argmax(&self, ignore_ambiguous: bool) -> Result<Option<Score>> {
35        match self {
36            Classification::Scores(scores) => argmax(scores),
37            Classification::Ambiguous(scores) => {
38                if ignore_ambiguous {
39                    argmax(scores)
40                } else {
41                    Ok(None)
42                }
43            }
44        }
45    }
46}
47
48/// The highest-confidence score, or `None` when the set is empty (the classifier abstained).
49/// Ties keep the first. Errors on a `NaN` confidence,
50/// which has no defined ordering.
51fn argmax(scores: &[Score]) -> Result<Option<Score>> {
52    let mut best: Option<&Score> = None;
53    for score in scores.iter() {
54        if score.confidence.is_nan() {
55            return Err(LibsyError::AlgorithmError {
56                message: format!(
57                    "classifier returned NaN confidence for target {:?}",
58                    score.target
59                ),
60            });
61        }
62        match best {
63            Some(cur_best) if score.confidence > cur_best.confidence => best = Some(score),
64            None => best = Some(score),
65            _ => {}
66        }
67    }
68    Ok(best.cloned())
69}
70
71/// Scores targets from the current request and the composition's state.
72#[async_trait]
73pub trait Classifier<S = ()>: Send + Sync {
74    /// Stable tier represented by `selected_model`, when this classifier defines one.
75    fn routing_tier(&self, _selected_model: &str) -> Option<&'static str> {
76        None
77    }
78
79    /// Drops retained routing state when `target` was unavailable for `request`.
80    ///
81    /// Stateless classifiers do not need to implement this hook.
82    fn target_unavailable(&self, _request: &Request, _target: &str) {}
83
84    /// Score the classifier's targets given the current state and request.
85    ///
86    /// When present, `driver` lets a classifier offload model calls. It is `None`
87    /// when the classifier is evaluated outside an algorithm run.
88    ///
89    /// `request` is borrowed mutably so a classifier may rewrite it in place — inject a
90    /// system prompt, drop tools, compact history. The edit is not scoped to this call:
91    /// later classifiers in the cascade score the rewritten request, and it is the
92    /// rewritten request that is finally sent to the selected model. Most classifiers
93    /// only read it.
94    async fn score(
95        &self,
96        state: &mut S,
97        request: &mut Request,
98        driver: Option<&Driver>,
99    ) -> Result<(Classification, Option<Response>)>;
100}
101
102#[cfg(test)]
103mod tests {
104    use super::*;
105    use switchyard_protocol::text_request;
106
107    /// Terse `Score` builder for the assertions below.
108    fn score(target: &str, confidence: f64) -> Score {
109        Score {
110            target: target.to_string(),
111            confidence,
112        }
113    }
114
115    #[test]
116    fn argmax_picks_the_highest_confidence_score() -> Result<()> {
117        let scores = vec![score("weak", 0.2), score("strong", 0.9), score("mid", 0.5)];
118        let best = Classification::Scores(scores).argmax(false)?;
119        assert_eq!(best, Some(score("strong", 0.9)));
120        Ok(())
121    }
122
123    #[test]
124    fn argmax_breaks_ties_by_cascade_order() -> Result<()> {
125        // Equal confidence: the earlier target in cascade order wins the tie.
126        let scores = vec![score("first", 0.7), score("second", 0.7)];
127        let best = Classification::Scores(scores).argmax(false)?;
128        assert_eq!(best.map(|s| s.target), Some("first".to_string()));
129        Ok(())
130    }
131
132    #[test]
133    fn argmax_on_an_empty_set_abstains() -> Result<()> {
134        // No scores means the classifier abstained — no choice to make.
135        assert_eq!(Classification::Scores(vec![]).argmax(false)?, None);
136        assert_eq!(Classification::Ambiguous(vec![]).argmax(true)?, None);
137        Ok(())
138    }
139
140    #[test]
141    fn argmax_errors_on_nan_confidence() {
142        // A NaN confidence has no defined ordering — surface it rather than guess.
143        let scores = vec![score("weak", 0.3), score("strong", f64::NAN)];
144        assert!(matches!(
145            Classification::Scores(scores).argmax(false),
146            Err(LibsyError::AlgorithmError { message })
147                if message == "classifier returned NaN confidence for target \"strong\""
148        ));
149        // A lone NaN errors too, even with nothing to compare it against.
150        assert!(matches!(
151            Classification::Scores(vec![score("only", f64::NAN)]).argmax(false),
152            Err(LibsyError::AlgorithmError { message })
153                if message == "classifier returned NaN confidence for target \"only\""
154        ));
155    }
156
157    #[test]
158    fn ambiguous_without_ignore_makes_no_choice() -> Result<()> {
159        // Ambiguous means "don't pick" unless the caller opts to ignore ambiguity.
160        let scores = vec![score("strong", 0.9)];
161        assert_eq!(Classification::Ambiguous(scores).argmax(false)?, None);
162        Ok(())
163    }
164
165    #[test]
166    fn ambiguous_with_ignore_falls_back_to_argmax() -> Result<()> {
167        let scores = vec![score("weak", 0.3), score("strong", 0.8)];
168        let best = Classification::Ambiguous(scores).argmax(true)?;
169        assert_eq!(best, Some(score("strong", 0.8)));
170        Ok(())
171    }
172
173    #[test]
174    fn scores_variant_ignores_the_ambiguous_flag() -> Result<()> {
175        // A definitive classification always yields its argmax, regardless of the flag.
176        let scores = vec![score("a", 0.4), score("b", 0.6)];
177        let with_ignore = Classification::Scores(scores.clone()).argmax(true)?;
178        let without_ignore = Classification::Scores(scores).argmax(false)?;
179        assert_eq!(with_ignore, without_ignore);
180        assert_eq!(with_ignore, Some(score("b", 0.6)));
181        Ok(())
182    }
183
184    /// Scores the request's requested model at full confidence and records that it ran.
185    struct RecordingClassifier;
186
187    #[async_trait]
188    impl Classifier<bool> for RecordingClassifier {
189        async fn score(
190            &self,
191            state: &mut bool,
192            request: &mut Request,
193            _driver: Option<&Driver>,
194        ) -> Result<(Classification, Option<Response>)> {
195            *state = true;
196            let target = request.requested_model().unwrap_or("auto").to_string();
197            Ok((
198                Classification::Scores(vec![Score {
199                    target,
200                    confidence: 1.0,
201                }]),
202                None,
203            ))
204        }
205    }
206
207    #[tokio::test]
208    async fn classifier_reads_request_and_mutates_state() -> Result<()> {
209        let mut state = false;
210        let mut request = Request {
211            llm_request: text_request(Some("strong".to_string()), "hi"),
212            raw_request: None,
213            metadata: None,
214        };
215        // A `None` driver is valid: the classifier scored without offloading a model call.
216        let (classification, _) = RecordingClassifier
217            .score(&mut state, &mut request, None)
218            .await?;
219        assert_eq!(
220            classification.argmax(false)?.map(|s| s.target),
221            Some("strong".to_string())
222        );
223        assert!(state);
224        Ok(())
225    }
226
227    /// Rewrites the request's model, then scores the rewritten value.
228    struct RewritingClassifier;
229
230    #[async_trait]
231    impl Classifier for RewritingClassifier {
232        async fn score(
233            &self,
234            _state: &mut (),
235            request: &mut Request,
236            _driver: Option<&Driver>,
237        ) -> Result<(Classification, Option<Response>)> {
238            request.llm_request.model = Some("rewritten".to_string());
239            Ok((
240                Classification::Scores(vec![Score {
241                    target: "rewritten".to_string(),
242                    confidence: 1.0,
243                }]),
244                None,
245            ))
246        }
247    }
248
249    #[tokio::test]
250    async fn classifier_rewrites_the_request_in_place() -> Result<()> {
251        let mut state = ();
252        let mut request = Request {
253            llm_request: text_request(Some("auto".to_string()), "hi"),
254            raw_request: None,
255            metadata: None,
256        };
257
258        RewritingClassifier
259            .score(&mut state, &mut request, None)
260            .await?;
261
262        // The rewrite outlives the call: later classifiers in the cascade score this value,
263        // and it is what reaches the model.
264        assert_eq!(request.requested_model(), Some("rewritten"));
265        Ok(())
266    }
267}
switchyard_libsy/core/classifier.rs

switchyard_libsy/core/
classifier.rs
