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::core::state::State;
6use crate::{LibsyError, Result};
7use async_trait::async_trait;
8use switchyard_protocol::{Category, ModelId, Request, Response};
9
10/// One classifier's recommendation of a routing `target`, with a `[0.0, 1.0]` confidence.
11#[derive(Debug, Clone, PartialEq)]
12pub struct Score {
13    /// `[0.0, 1.0]` confidence in `target`.
14    pub confidence: f64,
15    /// The target (model / tier) being recommended.
16    pub target: ModelId,
17    /// The category `target` was drawn from, when the classifier picked one. The rest of
18    /// that category is what the turn falls through on failure. A decision made without
19    /// a category leaves this `None`.
20    pub category: Option<Category>,
21}
22
23/// A classifier's verdict for a request: a set of target [`Score`]s, flagged by how
24/// confident the classifier is that they are decisive.
25pub enum Classification {
26    /// Definite recommendations; [`argmax`](Self::argmax) always yields the top target.
27    Scores(Vec<Score>),
28    /// Recommendations the classifier considers ambiguous; [`argmax`](Self::argmax) yields
29    /// nothing unless the caller opts to ignore ambiguity.
30    Ambiguous(Vec<Score>),
31}
32
33impl Classification {
34    /// The top-scoring [`Score`], or `None` when the classifier abstained (an empty set).
35    ///
36    /// An [`Ambiguous`](Self::Ambiguous) classification also yields `None` unless
37    /// `ignore_ambiguous` is set, in which case it falls back to the plain argmax.
38    /// Errors if any confidence is `NaN` (an unorderable score the caller should surface).
39    pub fn argmax(&self, ignore_ambiguous: bool) -> Result<Option<Score>> {
40        match self {
41            Classification::Scores(scores) => argmax(scores),
42            Classification::Ambiguous(scores) => {
43                if ignore_ambiguous {
44                    argmax(scores)
45                } else {
46                    Ok(None)
47                }
48            }
49        }
50    }
51}
52
53/// The highest-confidence score, or `None` when the set is empty (the classifier abstained).
54/// Ties keep the first. Errors on a `NaN` confidence,
55/// which has no defined ordering.
56fn argmax(scores: &[Score]) -> Result<Option<Score>> {
57    let mut best: Option<&Score> = None;
58    for score in scores.iter() {
59        if score.confidence.is_nan() {
60            return Err(LibsyError::AlgorithmError {
61                message: format!(
62                    "classifier returned NaN confidence for target {:?}",
63                    score.target
64                ),
65            });
66        }
67        match best {
68            Some(cur_best) if score.confidence > cur_best.confidence => best = Some(score),
69            None => best = Some(score),
70            _ => {}
71        }
72    }
73    Ok(best.cloned())
74}
75
76/// Scores targets from the current request and the composition's state.
77#[async_trait]
78pub trait Classifier: Send + Sync {
79    /// Score the classifier's targets given the current state and request.
80    ///
81    /// `driver` lets a classifier inspect runtime models and offload model calls.
82    ///
83    /// `request` is borrowed mutably so a classifier may rewrite it in place — inject a
84    /// system prompt, drop tools, compact history. The edit is not scoped to this call:
85    /// later classifiers in the cascade score the rewritten request, and it is the
86    /// rewritten request that is finally sent to the selected model. Most classifiers
87    /// only read it.
88    async fn score(
89        &self,
90        state: &mut State,
91        request: &mut Request,
92        driver: &Driver,
93    ) -> Result<(Classification, Option<Response>)>;
94}
95
96#[cfg(test)]
97mod tests {
98    use super::*;
99    use crate::core::testing::empty_driver;
100    use switchyard_protocol::text_request;
101
102    /// Terse `Score` builder for the assertions below.
103    fn score(target: &str, confidence: f64) -> Score {
104        Score {
105            target: ModelId::from(target),
106            confidence,
107            category: None,
108        }
109    }
110
111    #[test]
112    fn argmax_picks_the_highest_confidence_score() -> Result<()> {
113        let scores = vec![score("weak", 0.2), score("strong", 0.9), score("mid", 0.5)];
114        let best = Classification::Scores(scores).argmax(false)?;
115        assert_eq!(best, Some(score("strong", 0.9)));
116        Ok(())
117    }
118
119    #[test]
120    fn argmax_breaks_ties_by_cascade_order() -> Result<()> {
121        // Equal confidence: the earlier target in cascade order wins the tie.
122        let scores = vec![score("first", 0.7), score("second", 0.7)];
123        let best = Classification::Scores(scores).argmax(false)?;
124        assert_eq!(best.map(|s| s.target), Some(ModelId::from("first")));
125        Ok(())
126    }
127
128    #[test]
129    fn argmax_on_an_empty_set_abstains() -> Result<()> {
130        // No scores means the classifier abstained — no choice to make.
131        assert_eq!(Classification::Scores(vec![]).argmax(false)?, None);
132        assert_eq!(Classification::Ambiguous(vec![]).argmax(true)?, None);
133        Ok(())
134    }
135
136    #[test]
137    fn argmax_errors_on_nan_confidence() {
138        // A NaN confidence has no defined ordering — surface it rather than guess.
139        let scores = vec![score("weak", 0.3), score("strong", f64::NAN)];
140        assert!(matches!(
141            Classification::Scores(scores).argmax(false),
142            Err(LibsyError::AlgorithmError { message })
143                if message == "classifier returned NaN confidence for target \"strong\""
144        ));
145        // A lone NaN errors too, even with nothing to compare it against.
146        assert!(matches!(
147            Classification::Scores(vec![score("only", f64::NAN)]).argmax(false),
148            Err(LibsyError::AlgorithmError { message })
149                if message == "classifier returned NaN confidence for target \"only\""
150        ));
151    }
152
153    #[test]
154    fn ambiguous_without_ignore_makes_no_choice() -> Result<()> {
155        // Ambiguous means "don't pick" unless the caller opts to ignore ambiguity.
156        let scores = vec![score("strong", 0.9)];
157        assert_eq!(Classification::Ambiguous(scores).argmax(false)?, None);
158        Ok(())
159    }
160
161    #[test]
162    fn ambiguous_with_ignore_falls_back_to_argmax() -> Result<()> {
163        let scores = vec![score("weak", 0.3), score("strong", 0.8)];
164        let best = Classification::Ambiguous(scores).argmax(true)?;
165        assert_eq!(best, Some(score("strong", 0.8)));
166        Ok(())
167    }
168
169    #[test]
170    fn scores_variant_ignores_the_ambiguous_flag() -> Result<()> {
171        // A definitive classification always yields its argmax, regardless of the flag.
172        let scores = vec![score("a", 0.4), score("b", 0.6)];
173        let with_ignore = Classification::Scores(scores.clone()).argmax(true)?;
174        let without_ignore = Classification::Scores(scores).argmax(false)?;
175        assert_eq!(with_ignore, without_ignore);
176        assert_eq!(with_ignore, Some(score("b", 0.6)));
177        Ok(())
178    }
179
180    /// Scores the request's requested model at full confidence and records that it ran.
181    struct RecordingClassifier;
182
183    #[async_trait]
184    impl Classifier for RecordingClassifier {
185        async fn score(
186            &self,
187            state: &mut State,
188            request: &mut Request,
189            _driver: &Driver,
190        ) -> Result<(Classification, Option<Response>)> {
191            state
192                .extra
193                .insert("ran".to_string(), crate::core::state::StateValue::Count(1));
194            let target = request.model_id().unwrap_or(ModelId::from("auto"));
195            Ok((
196                Classification::Scores(vec![Score {
197                    target,
198                    confidence: 1.0,
199                    category: None,
200                }]),
201                None,
202            ))
203        }
204    }
205
206    #[tokio::test]
207    async fn classifier_reads_request_and_mutates_state() -> Result<()> {
208        let mut state = State::default();
209        let mut request = Request {
210            llm_request: text_request(Some("strong".to_string()), "hi"),
211            raw_request: None,
212            metadata: None,
213        };
214        let (classification, _) = RecordingClassifier
215            .score(&mut state, &mut request, &empty_driver())
216            .await?;
217        assert_eq!(
218            classification.argmax(false)?.map(|s| s.target),
219            Some(ModelId::from("strong"))
220        );
221        assert_eq!(
222            state.extra.get("ran"),
223            Some(&crate::core::state::StateValue::Count(1))
224        );
225        Ok(())
226    }
227
228    /// Rewrites the request's model, then scores the rewritten value.
229    struct RewritingClassifier;
230
231    #[async_trait]
232    impl Classifier for RewritingClassifier {
233        async fn score(
234            &self,
235            _state: &mut State,
236            request: &mut Request,
237            _driver: &Driver,
238        ) -> Result<(Classification, Option<Response>)> {
239            request.llm_request.model = Some("rewritten".to_string());
240            Ok((
241                Classification::Scores(vec![Score {
242                    target: ModelId::from("rewritten"),
243                    confidence: 1.0,
244                    category: None,
245                }]),
246                None,
247            ))
248        }
249    }
250
251    #[tokio::test]
252    async fn classifier_rewrites_the_request_in_place() -> Result<()> {
253        let mut state = State::default();
254        let mut request = Request {
255            llm_request: text_request(Some("auto".to_string()), "hi"),
256            raw_request: None,
257            metadata: None,
258        };
259
260        RewritingClassifier
261            .score(&mut state, &mut request, &empty_driver())
262            .await?;
263
264        // The rewrite outlives the call: later classifiers in the cascade score this value,
265        // and it is what reaches the model.
266        assert_eq!(request.model_id().as_deref(), Some("rewritten"));
267        Ok(())
268    }
269}
switchyard_libsy/core/classifier.rs

switchyard_libsy/core/
classifier.rs
