use async_trait::async_trait;
use std::path::PathBuf;
use crate::vlm_bench::scoring::{self, LevelScore, Rating};
use crate::vlm_bench::{BenchScenario, Difficulty, ExpectedAnswer, VlmBenchLevel};
const PASS_THRESHOLD: f64 = 0.60;
pub struct L3Architecture {
fixtures_dir: PathBuf,
}
impl Default for L3Architecture {
fn default() -> Self {
Self::new()
}
}
impl L3Architecture {
pub fn new() -> Self {
Self {
fixtures_dir: PathBuf::from("vlm_fixtures/l3_architecture"),
}
}
}
#[async_trait]
impl VlmBenchLevel for L3Architecture {
fn name(&self) -> &str {
"L3 Architecture"
}
fn difficulty(&self) -> Difficulty {
Difficulty::Hard
}
fn description(&self) -> &str {
"Architecture diagram comprehension: identify components from diagrams, \
trace data flow between modules, and map diagram boxes to Rust module paths."
}
fn scenarios(&self) -> Vec<BenchScenario> {
vec![
BenchScenario {
id: "l3_evolution_engine".into(),
description: "Map the evolution engine architecture diagram to code modules".into(),
image_path: self.fixtures_dir.join("evolution_diagram.png"),
prompt: "This diagram shows the architecture of an evolution engine. \
Identify each component box and map it to a Rust module path. \
Respond with JSON: {\"components\": [{\"name\": \"<box label>\", \
\"module\": \"<rust module path>\", \"role\": \"<brief description>\"}], \
\"data_flow\": [{\"from\": \"<component>\", \"to\": \"<component>\", \
\"data\": \"<what flows>\"}]}"
.into(),
expected: ExpectedAnswer::Keywords(vec![
"daemon".into(),
"sandbox".into(),
"fitness".into(),
"mutation".into(),
"tournament".into(),
]),
},
BenchScenario {
id: "l3_agent_pipeline".into(),
description: "Trace the agent pipeline from diagram to implementation".into(),
image_path: self.fixtures_dir.join("agent_pipeline.png"),
prompt: "This diagram shows an AI agent's execution pipeline. \
Identify the stages and their corresponding Rust modules. \
Respond with JSON: {\"stages\": [{\"name\": \"<stage>\", \
\"module\": \"<path>\", \"inputs\": [\"<input>\"], \
\"outputs\": [\"<output>\"]}]}"
.into(),
expected: ExpectedAnswer::Keywords(vec![
"agent".into(),
"tool".into(),
"parser".into(),
"context".into(),
]),
},
BenchScenario {
id: "l3_safety_layers".into(),
description: "Identify the layered safety architecture from a diagram".into(),
image_path: self.fixtures_dir.join("safety_layers.png"),
prompt: "This diagram shows a multi-layered safety architecture. \
List each layer from outermost to innermost and describe \
what it validates. Respond with JSON: \
{\"layers\": [{\"name\": \"<layer>\", \"validates\": \"<what>\", \
\"module\": \"<path>\"}]}"
.into(),
expected: ExpectedAnswer::Keywords(vec![
"safety".into(),
"validation".into(),
"path".into(),
"command".into(),
]),
},
]
}
fn evaluate(&self, scenario: &BenchScenario, response: &str) -> LevelScore {
let (accuracy, details) = match &scenario.expected {
ExpectedAnswer::Keywords(keywords) => {
let acc = scoring::keyword_accuracy(response, keywords);
let details = keywords
.iter()
.map(|kw| {
let found = response.to_lowercase().contains(&kw.to_lowercase());
(kw.clone(), if found { 1.0 } else { 0.0 })
})
.collect();
(acc, details)
}
ExpectedAnswer::JsonFields(expected) => {
scoring::json_field_accuracy(response, expected)
}
ExpectedAnswer::KeyValuePairs(pairs) => {
let mut total = 0.0;
let mut details = Vec::new();
for (key, value) in pairs {
let score = if response.to_lowercase().contains(&key.to_lowercase()) {
scoring::keyword_overlap_score(response, value)
} else {
0.0
};
details.push((key.clone(), score));
total += score;
}
let acc = if pairs.is_empty() {
1.0
} else {
total / pairs.len() as f64
};
(acc, details)
}
_ => (0.0, vec![]),
};
let rating = Rating::from_accuracy(accuracy, PASS_THRESHOLD);
LevelScore {
accuracy,
detail_scores: details,
response_tokens: 0,
latency_ms: 0,
rating,
}
}
}
#[cfg(test)]
#[path = "../../../tests/unit/vlm_bench/levels/l3_architecture/l3_architecture_test.rs"]
mod tests;