Skip to main content

harn_vm/llm/
trigger_predicate.rs

1use std::cell::RefCell;
2use std::collections::{BTreeMap, HashMap};
3use std::sync::{Mutex, OnceLock};
4use std::time::Duration;
5
6use serde::{Deserialize, Serialize};
7
8use super::api::{LlmRequestPayload, LlmResult};
9
10#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
11pub struct TriggerPredicateBudget {
12    pub max_cost_usd: Option<f64>,
13    pub tokens_max: Option<u64>,
14    pub timeout_ms: Option<u64>,
15}
16
17impl TriggerPredicateBudget {
18    pub fn timeout(&self) -> Option<Duration> {
19        self.timeout_ms.map(Duration::from_millis)
20    }
21}
22
23#[derive(Clone, Debug, Serialize, Deserialize)]
24pub(crate) struct PredicateCacheEntry {
25    pub request_hash: String,
26    pub(crate) result: LlmResult,
27}
28
29#[derive(Clone, Debug, Default)]
30pub struct PredicateEvaluationCapture {
31    pub entries: Vec<PredicateCacheEntry>,
32    pub total_tokens: u64,
33    pub total_cost_usd: f64,
34    pub cached: bool,
35    pub budget_exceeded: bool,
36}
37
38#[derive(Clone, Debug, Default)]
39struct PredicateEvaluationState {
40    budget: TriggerPredicateBudget,
41    replay_cache: HashMap<String, LlmResult>,
42    entries: BTreeMap<String, LlmResult>,
43    total_tokens: u64,
44    total_cost_usd: f64,
45    cached: bool,
46    budget_exceeded: bool,
47}
48
49thread_local! {
50    static ACTIVE_PREDICATE_EVALUATION: RefCell<Option<PredicateEvaluationState>> = const { RefCell::new(None) };
51}
52
53fn request_cache() -> &'static Mutex<HashMap<String, LlmResult>> {
54    static CACHE: OnceLock<Mutex<HashMap<String, LlmResult>>> = OnceLock::new();
55    CACHE.get_or_init(|| Mutex::new(HashMap::new()))
56}
57
58pub(crate) fn reset_trigger_predicate_state() {
59    ACTIVE_PREDICATE_EVALUATION.with(|slot| {
60        *slot.borrow_mut() = None;
61    });
62    if let Ok(mut cache) = request_cache().lock() {
63        cache.clear();
64    }
65}
66
67pub(crate) fn request_hash(request: &LlmRequestPayload) -> String {
68    use std::hash::{Hash, Hasher};
69
70    let canonical = serde_json::json!({
71        "provider": request.provider,
72        "model": request.model,
73        "messages": request.messages,
74        "system": request.system,
75        "max_tokens": request.max_tokens,
76        "temperature": request.temperature,
77        "top_p": request.top_p,
78        "top_k": request.top_k,
79        "stop": request.stop,
80        "seed": request.seed,
81        "frequency_penalty": request.frequency_penalty,
82        "presence_penalty": request.presence_penalty,
83        "output_format": request.output_format,
84        "response_format": request.response_format,
85        "json_schema": request.json_schema,
86        "thinking": request.thinking,
87        "anthropic_beta_features": request.anthropic_beta_features,
88        "native_tools": request.native_tools,
89        "tool_choice": request.tool_choice,
90        "cache": request.cache,
91        "timeout": request.timeout,
92        "stream": request.stream,
93        "provider_overrides": request.provider_overrides,
94        "prefill": request.prefill,
95        "mock_scope": request.mock_scope,
96    });
97    let mut hasher = std::collections::hash_map::DefaultHasher::new();
98    serde_json::to_string(&canonical)
99        .unwrap_or_default()
100        .hash(&mut hasher);
101    format!("{:016x}", hasher.finish())
102}
103
104pub(crate) struct PredicateEvaluationGuard;
105
106impl PredicateEvaluationGuard {
107    pub fn finish(self) -> PredicateEvaluationCapture {
108        finish_predicate_evaluation()
109    }
110}
111
112impl Drop for PredicateEvaluationGuard {
113    fn drop(&mut self) {
114        ACTIVE_PREDICATE_EVALUATION.with(|slot| {
115            *slot.borrow_mut() = None;
116        });
117    }
118}
119
120pub(crate) fn start_predicate_evaluation(
121    budget: TriggerPredicateBudget,
122    replay_entries: Vec<PredicateCacheEntry>,
123) -> PredicateEvaluationGuard {
124    ACTIVE_PREDICATE_EVALUATION.with(|slot| {
125        *slot.borrow_mut() = Some(PredicateEvaluationState {
126            budget,
127            replay_cache: replay_entries
128                .into_iter()
129                .map(|entry| (entry.request_hash, entry.result))
130                .collect(),
131            ..Default::default()
132        });
133    });
134    PredicateEvaluationGuard
135}
136
137fn finish_predicate_evaluation() -> PredicateEvaluationCapture {
138    ACTIVE_PREDICATE_EVALUATION.with(|slot| {
139        let Some(state) = slot.borrow_mut().take() else {
140            return PredicateEvaluationCapture::default();
141        };
142        PredicateEvaluationCapture {
143            entries: state
144                .entries
145                .into_iter()
146                .map(|(request_hash, result)| PredicateCacheEntry {
147                    request_hash,
148                    result,
149                })
150                .collect(),
151            total_tokens: state.total_tokens,
152            total_cost_usd: state.total_cost_usd,
153            cached: state.cached,
154            budget_exceeded: state.budget_exceeded,
155        }
156    })
157}
158
159pub(crate) fn lookup_cached_result(request: &LlmRequestPayload) -> Option<LlmResult> {
160    ACTIVE_PREDICATE_EVALUATION.with(|slot| {
161        let mut borrowed = slot.borrow_mut();
162        let state = borrowed.as_mut()?;
163        if state.budget_exceeded {
164            return None;
165        }
166        let hash = request_hash(request);
167        let cached = state.replay_cache.get(&hash).cloned().or_else(|| {
168            request_cache()
169                .lock()
170                .ok()
171                .and_then(|cache| cache.get(&hash).cloned())
172        });
173        if let Some(result) = cached {
174            state.cached = true;
175            state.entries.insert(hash, result.clone());
176            return Some(result);
177        }
178        None
179    })
180}
181
182pub(crate) fn note_result(request: &LlmRequestPayload, result: &LlmResult) {
183    ACTIVE_PREDICATE_EVALUATION.with(|slot| {
184        let mut borrowed = slot.borrow_mut();
185        let Some(state) = borrowed.as_mut() else {
186            return;
187        };
188        let hash = request_hash(request);
189        state.entries.insert(hash.clone(), result.clone());
190        if let Ok(mut cache) = request_cache().lock() {
191            cache.insert(hash, result.clone());
192        }
193        let call_tokens = result
194            .input_tokens
195            .saturating_add(result.output_tokens)
196            .max(0) as u64;
197        state.total_tokens = state.total_tokens.saturating_add(call_tokens);
198        state.total_cost_usd += result.priced_cost_usd().unwrap_or(0.0);
199        if state
200            .budget
201            .tokens_max
202            .is_some_and(|limit| state.total_tokens > limit)
203        {
204            state.budget_exceeded = true;
205        }
206        if state
207            .budget
208            .max_cost_usd
209            .is_some_and(|limit| state.total_cost_usd > limit)
210        {
211            state.budget_exceeded = true;
212        }
213    });
214}