Upload a batch of run outcomes for a single (lang, mode) combination. Normalizes model names and sanitizes volatile fields before upload. If `analysis` is provided, it is stored in the `llm_benchmark_analysis` table.
(&self, mode: &str, outcomes: &[RunOutcome], analysis: Option<&str>)
| 45 | /// Normalizes model names and sanitizes volatile fields before upload. |
| 46 | /// If `analysis` is provided, it is stored in the `llm_benchmark_analysis` table. |
| 47 | pub fn upload_batch(&self, mode: &str, outcomes: &[RunOutcome], analysis: Option<&str>) -> Result<usize> { |
| 48 | if outcomes.is_empty() { |
| 49 | return Ok(0); |
| 50 | } |
| 51 | |
| 52 | let mode = canonical_mode(mode); |
| 53 | |
| 54 | // Build in-memory Results so we can normalize model names |
| 55 | let mut results = Results::default(); |
| 56 | { |
| 57 | use crate::bench::normalize::{canonical_model_name, ensure_lang, ensure_mode, ensure_model}; |
| 58 | |
| 59 | for r in outcomes { |
| 60 | let lang_v = ensure_lang(&mut results, &r.lang); |
| 61 | let mode_v = ensure_mode(lang_v, mode, Some(r.hash.clone())); |
| 62 | let canonical_name = canonical_model_name(&r.model_name); |
| 63 | let model_v = ensure_model(mode_v, &canonical_name); |
| 64 | model_v.route_api_model = r.route_api_model.clone(); |
| 65 | |
| 66 | let mut sanitized = r.clone(); |
| 67 | sanitized.sanitize_for_commit(); |
| 68 | model_v.tasks.insert(r.task.clone(), sanitized); |
| 69 | } |
| 70 | } |
| 71 | normalize_model_names(&mut results); |
| 72 | |
| 73 | let url = format!("{}/api/llm-benchmark-upload", self.base_url); |
| 74 | let mut total_uploaded = 0usize; |
| 75 | |
| 76 | for lang_entry in &results.languages { |
| 77 | for mode_entry in &lang_entry.modes { |
| 78 | // Serialize models and inject analysis into each model object if provided |
| 79 | let mut models_json = serde_json::to_value(&mode_entry.models)?; |
| 80 | if let Some(text) = analysis |
| 81 | && let Some(arr) = models_json.as_array_mut() |
| 82 | { |
| 83 | for model in arr { |
| 84 | model["analysis"] = json!(text); |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | let payload = json!({ |
| 89 | "lang": lang_entry.lang, |
| 90 | "mode": mode_entry.mode, |
| 91 | "hash": mode_entry.hash, |
| 92 | "models": models_json, |
| 93 | }); |
| 94 | |
| 95 | let resp = self |
| 96 | .client |
| 97 | .post(&url) |
| 98 | .header("Authorization", format!("Bearer {}", self.api_key)) |
| 99 | .header("Content-Type", "application/json") |
| 100 | .json(&payload) |
| 101 | .send() |
| 102 | .with_context(|| format!("upload failed for {}/{}", lang_entry.lang, mode_entry.mode))?; |
| 103 | |
| 104 | if resp.status().is_success() { |
no test coverage detected