recall_echo/graph/
confidence.rs

1//! Bayesian confidence model for relationship edges.
2//!
3//! Uses Beta-Binomial conjugate prior with pseudocount 10.
4//! Confidence moves slowly per observation but accumulates with repeated evidence.
5
6use serde::{Deserialize, Serialize};
7
8/// Pseudocount total for the Beta-Binomial prior.
9/// ~10 observations to overwhelm the prior.
10const PSEUDOCOUNT: f64 = 10.0;
11
12/// How a relationship was established — determines initial confidence prior.
13#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize)]
14#[serde(rename_all = "snake_case")]
15pub enum ExtractionContext {
16    Explicit,      // 0.9
17    Inferred,      // 0.6
18    Speculative,   // 0.3
19    Authoritative, // 1.0
20}
21
22impl ExtractionContext {
23    /// Initial confidence prior for this extraction context.
24    pub fn prior(self) -> f64 {
25        match self {
26            Self::Authoritative => 1.0,
27            Self::Explicit => 0.9,
28            Self::Inferred => 0.6,
29            Self::Speculative => 0.3,
30        }
31    }
32}
33
34impl std::str::FromStr for ExtractionContext {
35    type Err = String;
36
37    fn from_str(s: &str) -> Result<Self, Self::Err> {
38        match s.to_lowercase().as_str() {
39            "explicit" => Ok(Self::Explicit),
40            "inferred" => Ok(Self::Inferred),
41            "speculative" => Ok(Self::Speculative),
42            "authoritative" => Ok(Self::Authoritative),
43            other => Err(format!("unknown extraction context: {}", other)),
44        }
45    }
46}
47
48/// Bayesian update using Beta-Binomial conjugate prior.
49///
50/// Given a current confidence (interpreted as alpha / (alpha + beta) with
51/// total pseudocount), updates the posterior by adding one observation.
52///
53/// - `corroborate = true`: alpha += 1 (evidence supports the relationship)
54/// - `corroborate = false`: beta += 1 (evidence contradicts the relationship)
55pub fn bayesian_update(current_confidence: f64, corroborate: bool) -> f64 {
56    let alpha = current_confidence * PSEUDOCOUNT;
57    let beta = PSEUDOCOUNT - alpha;
58
59    if corroborate {
60        (alpha + 1.0) / (alpha + beta + 1.0)
61    } else {
62        alpha / (alpha + beta + 1.0)
63    }
64}
65
66/// Compound confidence along a multi-hop path.
67///
68/// Returns the product of edge confidences. An empty path returns 1.0.
69pub fn path_confidence(edge_confidences: &[f64]) -> f64 {
70    edge_confidences.iter().product()
71}
72
73#[cfg(test)]
74mod tests {
75    use super::*;
76
77    fn approx_eq(a: f64, b: f64) -> bool {
78        (a - b).abs() < 0.001
79    }
80
81    #[test]
82    fn bayesian_update_corroborate_0_6() {
83        let result = bayesian_update(0.6, true);
84        // alpha=6, beta=4 -> (6+1)/(10+1) = 7/11 ≈ 0.636
85        assert!(approx_eq(result, 0.636), "got {}", result);
86    }
87
88    #[test]
89    fn bayesian_update_contradict_0_6() {
90        let result = bayesian_update(0.6, false);
91        // alpha=6, beta=4 -> 6/(10+1) = 6/11 ≈ 0.545
92        assert!(approx_eq(result, 0.545), "got {}", result);
93    }
94
95    #[test]
96    fn bayesian_update_corroborate_0_9() {
97        let result = bayesian_update(0.9, true);
98        // alpha=9, beta=1 -> (9+1)/(10+1) = 10/11 ≈ 0.909
99        assert!(approx_eq(result, 0.909), "got {}", result);
100    }
101
102    #[test]
103    fn bayesian_update_contradict_0_9() {
104        let result = bayesian_update(0.9, false);
105        // alpha=9, beta=1 -> 9/(10+1) = 9/11 ≈ 0.818
106        assert!(approx_eq(result, 0.818), "got {}", result);
107    }
108
109    #[test]
110    fn bayesian_update_corroborate_0_3() {
111        let result = bayesian_update(0.3, true);
112        // alpha=3, beta=7 -> (3+1)/(10+1) = 4/11 ≈ 0.364
113        assert!(approx_eq(result, 0.364), "got {}", result);
114    }
115
116    #[test]
117    fn path_confidence_two_edges() {
118        let result = path_confidence(&[0.8, 0.7]);
119        assert!(approx_eq(result, 0.56), "got {}", result);
120    }
121
122    #[test]
123    fn path_confidence_empty() {
124        assert_eq!(path_confidence(&[]), 1.0);
125    }
126
127    #[test]
128    fn extraction_context_priors() {
129        assert_eq!(ExtractionContext::Authoritative.prior(), 1.0);
130        assert_eq!(ExtractionContext::Explicit.prior(), 0.9);
131        assert_eq!(ExtractionContext::Inferred.prior(), 0.6);
132        assert_eq!(ExtractionContext::Speculative.prior(), 0.3);
133    }
134
135    #[test]
136    fn extraction_context_from_str() {
137        assert_eq!(
138            "explicit".parse::<ExtractionContext>().unwrap(),
139            ExtractionContext::Explicit
140        );
141        assert_eq!(
142            "inferred".parse::<ExtractionContext>().unwrap(),
143            ExtractionContext::Inferred
144        );
145        assert_eq!(
146            "speculative".parse::<ExtractionContext>().unwrap(),
147            ExtractionContext::Speculative
148        );
149        assert_eq!(
150            "authoritative".parse::<ExtractionContext>().unwrap(),
151            ExtractionContext::Authoritative
152        );
153        assert!("unknown".parse::<ExtractionContext>().is_err());
154    }
155}
recall_echo/graph/confidence.rs

recall_echo/graph/
confidence.rs
