1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
use super::schema::TraceEntryV1;
use crate::judge::JudgeService;
use crate::model::EvalConfig;
use crate::providers::embedder::Embedder;
use anyhow::Context;
use std::fs::File;
use std::io::{BufRead, BufReader, Write};
use std::path::Path;
use std::sync::Arc;
pub async fn precompute_embeddings(
input: &Path,
output: &Path,
embedder: Arc<dyn Embedder>,
model: &str,
_config: &EvalConfig, // potentially unused if we trust trace
) -> anyhow::Result<()> {
let file = File::open(input).context("failed to open input trace file")?;
let reader = BufReader::new(file);
let mut out_file = File::create(output).context("failed to create output trace file")?;
for (i, line) in reader.lines().enumerate() {
let line = line?;
if line.trim().is_empty() {
continue; // Skip empty lines
}
let mut entry: TraceEntryV1 = serde_json::from_str(&line)
.context(format!("failed to parse trace entry at line {}", i + 1))?;
// Check if already embedded
let has_embedding = entry.meta.pointer("/assay/embeddings/response").is_some()
&& entry.meta.pointer("/assay/embeddings/reference").is_some();
if !has_embedding {
eprintln!("Embedding entry {}...", entry.request_id);
// Heuristic: We need a "reference" text to embed against?
// But traces don't inherently have "expected" values unless we join with Config?
// Ah, for "semantic similarity", the expected value is in the Config (TestCase), not the trace.
// BUT, the runner logic expects `meta.assay.embeddings.reference`.
// If we precompute, we must know the TestCase expected value.
// Implication: We MUST verify the trace against the config and find the matching TestCase.
// For MVP, if we can't find the test case, we skip precompute for that entry?
// Or we just embed the RESPONSE (which is always there).
// Let's see `runner.rs:enrich_semantic`: it embeds `resp.text` AND `expected.semantic_similarity_to`.
// This implies `precompute_embeddings` needs to look up the `TestCase` by `entry.request_id == tc.id`.
// If we find it, and it expects semantic similarity, we embed BOTH.
// Find test case
let matching_tc = _config.tests.iter().find(|tc| tc.id == entry.request_id);
if let Some(tc) = matching_tc {
use crate::model::Expected;
if let Expected::SemanticSimilarityTo {
semantic_similarity_to,
..
} = &tc.expected
{
let resp_vec = embedder.embed(&entry.response).await?;
let ref_vec = embedder.embed(semantic_similarity_to).await?;
// Patch meta
if !entry.meta.is_object() {
entry.meta = serde_json::json!({});
}
if !entry.meta.get("assay").is_some_and(|v| v.is_object()) {
entry.meta["assay"] = serde_json::json!({});
}
entry.meta["assay"]["embeddings"] = serde_json::json!({
"model": model,
"response": resp_vec,
"reference": ref_vec,
"source_response": "precomputed",
"source_reference": "precomputed"
});
}
// If not semantic similarity, do we need embeddings? Maybe for classifiers? Not yet.
}
}
let out_line = serde_json::to_string(&entry)?;
writeln!(out_file, "{}", out_line)?;
}
Ok(())
}
pub async fn precompute_judge(
input: &Path,
output: &Path,
judge: &JudgeService,
config: &EvalConfig,
) -> anyhow::Result<()> {
let file = File::open(input).context("failed to open input trace file")?;
let reader = BufReader::new(file);
let mut out_file = File::create(output).context("failed to create output trace file")?;
for (i, line) in reader.lines().enumerate() {
let line = line?;
if line.trim().is_empty() {
continue;
}
let mut entry: TraceEntryV1 = serde_json::from_str(&line)
.context(format!("failed to parse trace entry at line {}", i + 1))?;
// Check matching TestCase
let matching_tc = config.tests.iter().find(|tc| tc.id == entry.request_id);
if let Some(tc) = matching_tc {
use crate::model::Expected;
let (rubric_id, rubric_version) = match &tc.expected {
Expected::Faithfulness { rubric_version, .. } => {
("faithfulness", rubric_version.as_deref())
}
Expected::Relevance { rubric_version, .. } => {
("relevance", rubric_version.as_deref())
}
Expected::JudgeCriteria { .. } => ("custom", None),
_ => ("none", None),
};
if rubric_id != "none" {
// Check if already judged
let existing = entry
.meta
.pointer(&format!("/assay/judge/{}", rubric_id))
.is_some();
if !existing {
eprintln!("Judging entry {} ({}) ...", entry.request_id, rubric_id);
use crate::model::TestInput;
let input = TestInput {
prompt: entry.prompt.clone(),
context: None, // Traces usually contain full context in prompt or meta? For V1, flat prompt.
};
// We modify entry.meta in place
// But JudgeService::evaluate expects `&mut serde_json::Value` (meta)
// and typically writes to `assay.judge.{rubric_id}`.
judge
.evaluate(
&entry.request_id,
rubric_id,
&input,
&entry.response,
rubric_version,
&mut entry.meta,
)
.await?;
}
}
}
let out_line = serde_json::to_string(&entry)?;
writeln!(out_file, "{}", out_line)?;
}
Ok(())
}