File size: 9,614 Bytes
94193b5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 | import { ProviderId } from '@/lib/llm/providers/types';
import { getProvider } from '@/lib/llm/providers/registry';
import type { UsageInfo } from '@/lib/llm/types';
/** Normalizes a judge provider's response usage into UsageInfo (undefined if absent). */
export function extractJudgeUsage(provider: string, model: string, data: unknown): UsageInfo | undefined {
const d = data as Record<string, unknown>;
if (provider === 'gemini') {
const u = d?.usageMetadata as Record<string, number> | undefined;
if (!u) return undefined;
return {
promptTokens: u.promptTokenCount || 0,
completionTokens: u.candidatesTokenCount || 0,
totalTokens: u.totalTokenCount || 0,
model, provider,
};
}
const u = d?.usage as Record<string, number> | undefined;
if (!u) return undefined;
if (provider === 'anthropic') {
const pt = u.input_tokens || 0;
const ct = u.output_tokens || 0;
return { promptTokens: pt, completionTokens: ct, totalTokens: pt + ct, model, provider };
}
const pt = u.prompt_tokens || 0;
const ct = u.completion_tokens || 0;
return { promptTokens: pt, completionTokens: ct, totalTokens: u.total_tokens || pt + ct, model, provider };
}
export interface JudgeConfig {
provider: ProviderId;
apiKey: string;
model: string;
}
export interface JudgeContext {
prompt: string;
files: Record<string, string>;
summary: string;
}
export interface JudgeResult {
passed: boolean;
reasoning: string;
}
const SYSTEM_PROMPT = `You are a benchmark judge evaluating whether an AI coding assistant completed a task correctly.
You will be given:
1. The original task prompt
2. The final state of project files
3. A summary from the AI assistant
Evaluate whether the task was completed correctly based on the criteria provided.
Respond in EXACTLY this format:
VERDICT: PASS
REASONING: <one paragraph explaining your judgment>
Or:
VERDICT: FAIL
REASONING: <one paragraph explaining what was missing or incorrect>`;
function buildUserMessage(criteria: string, context: JudgeContext): string {
const fileSummary = Object.entries(context.files)
.map(([path, content]) => `--- ${path} ---\n${content.substring(0, 2000)}`)
.join('\n\n');
return `## Task Prompt
${context.prompt}
## Evaluation Criteria
${criteria}
## Assistant Summary
${context.summary}
## Project Files
${fileSummary}
Evaluate whether the task was completed correctly based on the criteria above.`;
}
function parseVerdict(response: string): JudgeResult {
const verdictMatch = /VERDICT:\s*(PASS|FAIL)/i.exec(response);
const reasoningMatch = /REASONING:\s*([\s\S]+)/i.exec(response);
return {
passed: verdictMatch ? verdictMatch[1].toUpperCase() === 'PASS' : false,
reasoning: reasoningMatch ? reasoningMatch[1].trim() : response.substring(0, 200),
};
}
async function callOpenAICompatible(
baseUrl: string,
apiKey: string,
model: string,
provider: ProviderId,
systemPrompt: string,
userMessage: string
): Promise<{ text: string; usage?: UsageInfo }> {
const headers: Record<string, string> = { 'Content-Type': 'application/json' };
if (apiKey) headers['Authorization'] = `Bearer ${apiKey}`;
if (provider === 'openrouter') {
headers['HTTP-Referer'] = typeof window !== 'undefined' ? window.location.origin : 'http://localhost:3000';
headers['X-Title'] = 'OSW-Studio';
}
const response = await fetch(`${baseUrl}/chat/completions`, {
method: 'POST',
headers,
body: JSON.stringify({
model,
messages: [
{ role: 'system', content: systemPrompt },
{ role: 'user', content: userMessage },
],
temperature: 0.2,
max_tokens: 512,
stream: false,
}),
});
if (!response.ok) {
const error = await response.text();
throw new Error(`Judge API error (${provider}): ${error}`);
}
const data = await response.json();
return { text: data.choices?.[0]?.message?.content || '', usage: extractJudgeUsage(provider, model, data) };
}
async function callAnthropic(
apiKey: string,
model: string,
systemPrompt: string,
userMessage: string
): Promise<{ text: string; usage?: UsageInfo }> {
const response = await fetch('https://api.anthropic.com/v1/messages', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
'x-api-key': apiKey,
'anthropic-version': '2023-06-01',
},
body: JSON.stringify({
model,
system: systemPrompt,
messages: [{ role: 'user', content: userMessage }],
temperature: 0.2,
max_tokens: 512,
}),
});
if (!response.ok) {
const error = await response.text();
throw new Error(`Judge API error (anthropic): ${error}`);
}
const data = await response.json();
return { text: data.content?.[0]?.text || '', usage: extractJudgeUsage('anthropic', model, data) };
}
async function callGemini(
apiKey: string,
model: string,
systemPrompt: string,
userMessage: string
): Promise<{ text: string; usage?: UsageInfo }> {
const response = await fetch(
`https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent?key=${apiKey}`,
{
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
contents: [
{ role: 'user', parts: [{ text: systemPrompt }] },
{ role: 'user', parts: [{ text: userMessage }] },
],
generationConfig: { temperature: 0.2, maxOutputTokens: 512 },
}),
}
);
if (!response.ok) {
const error = await response.text();
throw new Error(`Judge API error (gemini): ${error}`);
}
const data = await response.json();
return { text: data.candidates?.[0]?.content?.parts?.[0]?.text || '', usage: extractJudgeUsage('gemini', model, data) };
}
export async function runJudgeEvaluation(
criteria: string,
context: JudgeContext,
config: JudgeConfig
): Promise<JudgeResult> {
const providerConfig = getProvider(config.provider);
const userMessage = buildUserMessage(criteria, context);
let responseText: string;
if (config.provider === 'anthropic') {
responseText = (await callAnthropic(config.apiKey, config.model, SYSTEM_PROMPT, userMessage)).text;
} else if (config.provider === 'gemini') {
responseText = (await callGemini(config.apiKey, config.model, SYSTEM_PROMPT, userMessage)).text;
} else {
const baseUrl = providerConfig.baseUrl || 'https://openrouter.ai/api/v1';
responseText = (await callOpenAICompatible(
baseUrl, config.apiKey, config.model, config.provider, SYSTEM_PROMPT, userMessage
)).text;
}
return parseVerdict(responseText);
}
// ---- Structured (multi-criteria) judge: one call, one verdict per criterion ----
const STRUCTURED_SYSTEM_PROMPT = `You are a completion judge. You are given several numbered criteria and the current state of project files. For EACH criterion, decide whether the files actually satisfy it.
Judge ONLY against what is actually recorded in the files — do not assume or invent. Respond with one line per criterion, EXACTLY in this format:
<number>: PASS
or
<number>: FAIL - <short reason of what is missing>
Output nothing else.`;
function buildStructuredUserMessage(criteria: string[], context: JudgeContext): string {
const fileSummary = Object.entries(context.files)
.map(([path, content]) => `--- ${path} ---\n${content.substring(0, 2000)}`)
.join('\n\n') || '(no files)';
const criteriaList = criteria.map((c, i) => `${i + 1}. ${c}`).join('\n');
return `## Context
${context.prompt}
## Criteria to evaluate
${criteriaList}
## Project Files
${fileSummary}
Evaluate each numbered criterion against the files above.`;
}
/**
* Parses one PASS/FAIL verdict per criterion from the judge's response.
* Tolerates "1:", "1.", "1)", "ITEM 1:", and en/em-dash reason separators.
* Any criterion without a parseable verdict fails closed.
*/
export function parseStructuredVerdicts(response: string, count: number): JudgeResult[] {
const results: JudgeResult[] = [];
for (let i = 1; i <= count; i++) {
const re = new RegExp(`^\\s*(?:item\\s*|#)?${i}\\s*[:.)\\]]\\s*(PASS|FAIL)\\b\\s*[-–—:]?\\s*(.*)$`, 'im');
const m = re.exec(response);
if (m) {
results.push({ passed: m[1].toUpperCase() === 'PASS', reasoning: (m[2] || '').trim() });
} else {
results.push({ passed: false, reasoning: 'Could not verify this item.' });
}
}
return results;
}
/**
* Evaluates multiple criteria in a single judge call. Returns one verdict per
* criterion, in the same order as the input.
*/
export async function runStructuredJudge(
criteria: string[],
context: JudgeContext,
config: JudgeConfig
): Promise<{ verdicts: JudgeResult[]; usage?: UsageInfo }> {
if (criteria.length === 0) return { verdicts: [] };
const providerConfig = getProvider(config.provider);
const userMessage = buildStructuredUserMessage(criteria, context);
let result: { text: string; usage?: UsageInfo };
if (config.provider === 'anthropic') {
result = await callAnthropic(config.apiKey, config.model, STRUCTURED_SYSTEM_PROMPT, userMessage);
} else if (config.provider === 'gemini') {
result = await callGemini(config.apiKey, config.model, STRUCTURED_SYSTEM_PROMPT, userMessage);
} else {
const baseUrl = providerConfig.baseUrl || 'https://openrouter.ai/api/v1';
result = await callOpenAICompatible(
baseUrl, config.apiKey, config.model, config.provider, STRUCTURED_SYSTEM_PROMPT, userMessage
);
}
return { verdicts: parseStructuredVerdicts(result.text, criteria.length), usage: result.usage };
}
|