ollama-eval.mjs
Rama main · réplica de lectura del repositorio original.
#!/usr/bin/env node
/**
* ollama-eval.mjs — Ollama-powered Job Offer Evaluator for career-ops
*
* Local, free, private alternative to the Claude-based pipeline.
* Reads evaluation logic from modes/oferta.md + modes/_shared.md,
* reads the user's resume from cv.md, and evaluates a Job Description
* passed as a CLI argument or file.
*
* Usage:
* node ollama-eval.mjs "Paste full JD text here"
* node ollama-eval.mjs --file ./jds/my-job.txt
* node ollama-eval.mjs --model qwen2.5:72b --file ./jds/my-job.txt
*
* Requires:
* Ollama running locally — https://ollama.com
* A model pulled: ollama pull llama3.3
*
* Context window guidance:
* The prompt (cv + modes + JD) is ~10K-15K tokens.
* Recommended models (32K+ context): llama3.3, mistral-nemo, qwen2.5, gemma3
* Smaller models (llama3.2:3b, phi3) may produce incomplete evaluations.
*/
import { readFileSync, existsSync, writeFileSync, mkdirSync } from 'fs';
import { join, dirname } from 'path';
import { fileURLToPath } from 'url';
import { getCareerOpsRoot, getProfilePath } from './path-resolver.mjs';
import { localToday } from './lib/local-today.mjs';
import { withJdArchiveSection } from './lib/jd-archive-section.mjs';
import { TSV_ADDITION_HEADER } from './tracker-parse.mjs';
import { outputLanguageInstruction, parseOutputLanguage } from './profile-language.mjs';
import {
formatReportNumber, releaseReportNumbers, reserveReportNumbers,
} from './reserve-report-num.mjs';
import { TokenAccumulator, formatBreakdown, normalizeOpenAIUsage } from './utils/token-tracker.mjs';
import { buildBudgetedPrompt } from './lib/context-budget.mjs';
import { assertLlmRequestAllowed } from './lib/llm-outbound.mjs';
import { installPiiInterceptor } from './lib/pii-interceptor.mjs';
import {
isPostingUrl, normalizedTrackerScore, parseScoreSummary, slugifyCompany, tsvSafe,
} from './lib/tracker-addition.mjs';
// Ambient guard over global fetch (idempotent); the explicit gate below is the
// primary layer.
installPiiInterceptor();
const tracker = new TokenAccumulator();
tracker.recordZeroToken('scan');
tracker.recordZeroToken('pdf payload');
try {
const { config } = await import('dotenv');
// The data root's .env first: doctor.mjs, remember-cli.mjs, scan.mjs and the
// plugin engine all read it there. The working directory's .env, the only one
// read before, still supplies what the data root's leaves unset.
config({ path: [join(getCareerOpsRoot(), '.env'), join(process.cwd(), '.env')] });
} catch { /* dotenv optional */ }
const ROOT = dirname(fileURLToPath(import.meta.url));
const DATA_ROOT = getCareerOpsRoot();
// ---------------------------------------------------------------------------
// Paths
// ---------------------------------------------------------------------------
const PATHS = {
shared: join(ROOT, 'modes', '_shared.md'),
oferta: join(ROOT, 'modes', 'oferta.md'),
cv: join(DATA_ROOT, 'cv.md'),
profile: join(DATA_ROOT, 'modes', '_profile.md'),
custom: join(DATA_ROOT, 'modes', '_custom.md'),
profileYml: getProfilePath(),
reports: join(DATA_ROOT, 'reports'),
// CAREER_OPS_ADDITIONS mirrors merge-tracker.mjs:43. Writing under DATA_ROOT
// regardless would drop the addition somewhere the merge it instructs never
// looks, so the evaluation would sit there unread.
trackerAdditions: process.env.CAREER_OPS_ADDITIONS
? process.env.CAREER_OPS_ADDITIONS
: join(DATA_ROOT, 'batch', 'tracker-additions'),
};
// ---------------------------------------------------------------------------
// CLI argument parsing
// ---------------------------------------------------------------------------
const args = process.argv.slice(2);
if (args.length === 0 || args[0] === '--help' || args[0] === '-h') {
console.log(`
╔══════════════════════════════════════════════════════════════════╗
║ career-ops — Ollama Evaluator (local / free) ║
╚══════════════════════════════════════════════════════════════════╝
Evaluate a job offer using a local Ollama model instead of Claude.
USAGE
node ollama-eval.mjs "<JD text>"
node ollama-eval.mjs --file ./jds/my-job.txt
node ollama-eval.mjs --model qwen2.5:72b "<JD text>"
OPTIONS
--file <path> Read JD from a file instead of inline text
--model <name> Ollama model to use (default: llama3.3)
--url <url> Ollama base URL (default: http://localhost:11434)
--posting-url <url> Posting URL, recorded in the report header and
used as the tracker's dedup key
--no-save Do not save report to reports/ directory
--help Show this help
SETUP
1. Install Ollama: https://ollama.com
2. Pull a model: ollama pull llama3.3
3. Start server: ollama serve (or it auto-starts)
4. Run this script
EXAMPLES
node ollama-eval.mjs "We are looking for a Senior AI Engineer..."
node ollama-eval.mjs --file ./jds/openai-swe.txt
OLLAMA_MODEL=mistral-nemo node ollama-eval.mjs --file ./jds/job.txt
`);
process.exit(0);
}
// Parse flags
let jdText = '';
let postingUrl = '';
let modelName = process.env.OLLAMA_MODEL || 'llama3.3';
let baseUrl = (process.env.OLLAMA_BASE_URL || 'http://localhost:11434').replace(/\/$/, '');
// Context window for the request AND the prompt budget. Defaults to the previous hardcoded
// 32768, so behaviour is unchanged unless OLLAMA_NUM_CTX is set. Raise it for a model with a
// bigger window; `ollama show` reports each model's ceiling (qwen2.5 caps at 32768).
// Digits only: parseInt read "32k" as 32, a 32-token context, and the prompt was
// cut to fit it.
const numCtxRaw = (process.env.OLLAMA_NUM_CTX || '32768').trim();
const numCtx = /^\d+$/.test(numCtxRaw) ? Number(numCtxRaw) : NaN;
if (!Number.isSafeInteger(numCtx) || numCtx <= 0) {
console.error(`❌ Invalid OLLAMA_NUM_CTX: "${process.env.OLLAMA_NUM_CTX}" — must be a positive integer (tokens).`);
process.exit(1);
}
// Checked here, with the other settings, before the server is contacted. Digits
// only (parseInt read "5m" as a 5 ms timeout) and at most 2^31-1 ms:
// AbortSignal.timeout() fires a longer delay after 1 ms, and throws past 2^32-1.
const timeoutRaw = (process.env.OLLAMA_TIMEOUT_MS || '300000').trim();
const timeoutMs = /^\d+$/.test(timeoutRaw) ? Number(timeoutRaw) : NaN;
if (!Number.isSafeInteger(timeoutMs) || timeoutMs <= 0 || timeoutMs > 2_147_483_647) {
console.error(`❌ Invalid OLLAMA_TIMEOUT_MS: "${process.env.OLLAMA_TIMEOUT_MS}" — must be a positive integer of at most 2147483647 (milliseconds).`);
process.exit(1);
}
let saveReport = true;
for (let i = 0; i < args.length; i++) {
if (args[i] === '--file' && args[i + 1]) {
const filePath = args[++i];
if (!existsSync(filePath)) {
console.error(`❌ File not found: ${filePath}`);
process.exit(1);
}
try {
jdText = readFileSync(filePath, 'utf-8').trim();
} catch (err) {
console.error(`❌ Could not read file: ${filePath}`);
console.error(` ${err.message}`);
process.exit(1);
}
} else if (args[i] === '--model' && args[i + 1]) {
modelName = args[++i];
} else if (args[i] === '--url' && args[i + 1]) {
baseUrl = args[++i].replace(/\/$/, '');
} else if (args[i] === '--posting-url' && args[i + 1]) {
postingUrl = args[++i];
} else if (args[i] === '--no-save') {
saveReport = false;
} else if (!args[i].startsWith('--')) {
jdText += (jdText ? '\n' : '') + args[i];
}
}
if (!jdText) {
console.error('❌ No Job Description provided. Run with --help for usage.');
process.exit(1);
}
// A posting URL is the tracker's deterministic dedup key, so it is taken only in
// a form that can actually become one. Parsed, not prefix-matched: `https://`
// satisfies a prefix test and merge-tracker.mjs:697 would then classify it as
// the URL extra, but normalizeUrl yields no key for it -- so it would sit in the
// URL column looking like a key while deduping nothing. A placeholder written
// there would be worse still, handing every such row the same key.
if (postingUrl && !isPostingUrl(postingUrl)) {
console.error(`❌ --posting-url must be a complete http(s) URL: "${postingUrl}"`);
process.exit(1);
}
// ---------------------------------------------------------------------------
// File helpers
// ---------------------------------------------------------------------------
/**
* Read a file and return its trimmed contents, or a placeholder if missing.
* Emits a console warning when the file is absent so the user knows context is incomplete.
* @param {string} path - Absolute path to the file.
* @param {string} label - Human-readable label used in the warning and placeholder.
* @returns {string} File contents or a "[label not found]" placeholder.
*/
function readFile(path, label) {
if (!existsSync(path)) {
console.warn(`⚠️ ${label} not found at: ${path}`);
return `[${label} not found — skipping]`;
}
return readFileSync(path, 'utf-8').trim();
}
// ---------------------------------------------------------------------------
// Loopback guard — cv.md + full JD are sent to this endpoint.
// A remote URL would silently exfiltrate private data.
// ---------------------------------------------------------------------------
{
let hostname;
try {
hostname = new URL(baseUrl).hostname;
} catch {
console.error(`❌ Invalid OLLAMA_BASE_URL: "${baseUrl}"`);
process.exit(1);
}
const isLoopback = hostname === 'localhost' || hostname === '127.0.0.1' || hostname === '::1';
if (!isLoopback && process.env.OLLAMA_ALLOW_REMOTE !== '1') {
console.error(`
❌ Remote Ollama endpoint detected: ${baseUrl}
Your CV and job description would be sent to a remote server.
This tool is designed for local use only.
If you intentionally want to use a remote endpoint (e.g. tunnelled
Ollama on a home server), set:
OLLAMA_ALLOW_REMOTE=1 node ollama-eval.mjs ...
`);
process.exit(1);
}
}
// ---------------------------------------------------------------------------
// Check Ollama is reachable before burning time on prompt assembly
// ---------------------------------------------------------------------------
try {
const probe = await fetch(`${baseUrl}/api/tags`, { signal: AbortSignal.timeout(5_000) });
if (!probe.ok) throw new Error(`HTTP ${probe.status}`);
} catch (err) {
console.error(`
❌ Ollama not reachable at ${baseUrl}
1. Install Ollama: https://ollama.com
2. Start server: ollama serve
3. Pull a model: ollama pull ${modelName}
`);
process.exit(1);
}
// ---------------------------------------------------------------------------
// Load context files
// ---------------------------------------------------------------------------
console.log('\n📂 Loading context files...');
const sharedContext = readFile(PATHS.shared, 'modes/_shared.md');
const ofertaLogic = readFile(PATHS.oferta, 'modes/oferta.md');
const cvContent = readFile(PATHS.cv, 'cv.md');
const profileContent = readFile(PATHS.profile, 'modes/_profile.md');
// _custom.md is optional: an absent file is a valid end state, so no warning.
const customContent = existsSync(PATHS.custom) ? readFileSync(PATHS.custom, 'utf-8').trim() : '';
const profileYml = readFile(PATHS.profileYml, 'config/profile.yml');
const languageInstruction = outputLanguageInstruction(parseOutputLanguage(profileYml));
// ---------------------------------------------------------------------------
// Build system prompt with token budget management
// ---------------------------------------------------------------------------
const { contextBody, budgetReport } = buildBudgetedPrompt({
sharedContent: sharedContext,
ofertaContent: ofertaLogic,
cvContent,
profileYml,
profileContent,
customContent,
jdText,
jdInContext: false, // The user message carries the JD; still budget it once.
maxTokens: numCtx, // matches options.num_ctx below
});
if (budgetReport.compressed) {
console.log(`📊 Token budget: ${budgetReport.beforeTokens} → ${budgetReport.afterTokens} tokens (saved ${budgetReport.beforeTokens - budgetReport.afterTokens})`);
console.log(` Trimmed sections: ${budgetReport.removed.join(', ')}`);
if (budgetReport.overBudget) {
console.log(` ⚠️ Still ${budgetReport.afterTokens - budgetReport.budget} tokens over budget after compression`);
}
} else if (budgetReport.overBudget) {
console.log(`⚠️ Token budget: ${budgetReport.totalTokens} tokens exceeds ${budgetReport.budget} limit by ${budgetReport.totalTokens - budgetReport.budget}`);
} else {
console.log(`📊 Token budget: ${budgetReport.totalTokens} tokens (within ${budgetReport.budget} limit)`);
}
const systemPrompt = `You are career-ops, an AI-powered job search assistant.
You evaluate job offers against the user's CV using a structured A-G scoring system.
Your evaluation methodology is defined below. Follow it exactly.
${contextBody}
═══════════════════════════════════════════════════════
IMPORTANT OPERATING RULES FOR THIS SESSION
═══════════════════════════════════════════════════════
1. You do NOT have access to WebSearch, Playwright, or file writing tools.
- Block D (Comp research): use training-data salary estimates; note them as estimates.
- Block G (Legitimacy): analyze JD text only; skip URL/page freshness checks.
- Post-evaluation file saving is handled by the script, not by you.
2. ${languageInstruction}
3. Generate Blocks A through G in full.
4. At the very end, output this exact machine-readable block:
---SCORE_SUMMARY---
COMPANY: <company name or "Unknown">
ROLE: <role title>
SCORE: <global score as decimal, e.g. 3.8>
ARCHETYPE: <detected archetype>
LEGITIMACY: <High Confidence | Proceed with Caution | Suspicious>
---END_SUMMARY---
`;
// ---------------------------------------------------------------------------
// Call Ollama
// ---------------------------------------------------------------------------
const endpoint = `${baseUrl}/api/chat`;
console.log(`🤖 Calling Ollama (${modelName})... this may take a minute.\n`);
let evaluationText;
try {
// Privacy gate: inspect the complete serialized request before the network
// call. Ollama may be local, but it may equally be a remote host set via
// OLLAMA_BASE_URL — either way the payload is checked.
const requestBody = JSON.stringify({
model: modelName,
messages: [
{ role: 'system', content: systemPrompt },
{ role: 'user', content: `JOB DESCRIPTION TO EVALUATE:\n\n${jdText}` },
],
// stream:true so response headers arrive with the FIRST token. With stream:false
// Ollama sends nothing until the whole report is generated, and Node's undici client
// gives up at its own 300s headersTimeout — a deadline neither OLLAMA_TIMEOUT_MS nor
// AbortSignal.timeout controls, which surfaced as a bare "fetch failed" at 5:01.
stream: true,
// Ollama's native /api/chat reads generation params from `options` only.
// This call targets that endpoint (NOT the OpenAI-compatible /v1 route,
// which ignores `options` and has no num_ctx equivalent), so both the
// deterministic temperature and the enlarged context window actually take
// effect. Without num_ctx here Ollama defaults to a 2048-token context and
// silently truncates the prompt; without temperature it runs at 0.8.
options: { temperature: 0.4, num_ctx: numCtx },
});
assertLlmRequestAllowed(endpoint, requestBody);
const res = await fetch(endpoint, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: requestBody,
signal: AbortSignal.timeout(timeoutMs),
});
if (!res.ok) {
const body = await res.text();
console.error(`❌ Ollama API error: HTTP ${res.status}`);
console.error(` ${body.slice(0, 300)}`);
process.exit(1);
}
// Streamed /api/chat is newline-delimited JSON: one object per token, the last carrying
// done:true and the token counts.
let acc = '', buf = '', promptCount = 0, evalCount = 0;
const decoder = new TextDecoder();
for await (const chunk of res.body) {
buf += decoder.decode(chunk, { stream: true });
let nl;
while ((nl = buf.indexOf('\n')) !== -1) {
const line = buf.slice(0, nl).trim();
buf = buf.slice(nl + 1);
if (!line) continue;
let obj;
try { obj = JSON.parse(line); } catch { continue; }
if (obj.error) {
console.error(`❌ Ollama error: ${obj.error}`);
process.exit(1);
}
if (obj.message?.content) acc += obj.message.content;
if (obj.done) {
promptCount = obj.prompt_eval_count ?? 0;
evalCount = obj.eval_count ?? 0;
}
}
}
// Flush a final line that arrived without a trailing newline. Ollama terminates every
// chunk with one, but a body that ends mid-line would otherwise be dropped silently.
const tail = buf.trim();
if (tail) {
try {
const obj = JSON.parse(tail);
if (obj.error) {
console.error(`❌ Ollama error: ${obj.error}`);
process.exit(1);
}
if (obj.message?.content) acc += obj.message.content;
if (obj.done) {
promptCount = obj.prompt_eval_count ?? promptCount;
evalCount = obj.eval_count ?? evalCount;
}
} catch { /* a truncated final line is not recoverable; the empty-response check below reports it */ }
}
evaluationText = acc.trim();
// Native /api/chat reports tokens as prompt_eval_count / eval_count, not an
// OpenAI-shaped `usage` object; map them through the shared normalizer.
const usage = normalizeOpenAIUsage({
prompt_tokens: promptCount,
completion_tokens: evalCount,
total_tokens: promptCount + evalCount,
});
tracker.record('evaluation', usage);
if (!evaluationText) {
console.error('❌ Ollama returned an empty response.');
process.exit(1);
}
} catch (err) {
if (err.name === 'TimeoutError') {
console.error(`❌ Request timed out after ${Math.round(timeoutMs / 1000)}s.`);
console.error(` Try a smaller/faster model, or increase OLLAMA_TIMEOUT_MS.`);
} else {
console.error(`❌ Ollama API call failed: ${err.message}`);
}
process.exit(1);
}
// ---------------------------------------------------------------------------
// Display evaluation
// ---------------------------------------------------------------------------
console.log('\n' + '═'.repeat(66));
console.log(' CAREER-OPS EVALUATION — powered by Ollama (' + modelName + ')');
console.log('═'.repeat(66) + '\n');
console.log(evaluationText);
// ---------------------------------------------------------------------------
// Parse score summary
// ---------------------------------------------------------------------------
const summary = parseScoreSummary(evaluationText);
const company = summary?.company || 'unknown';
const role = summary?.role || 'unknown';
const score = summary?.score || '?';
const archetype = summary?.archetype || 'unknown';
const legitimacy = summary?.legitimacy || 'unknown';
// An addition for company and role "unknown" would merge as a real application.
const trackable = Boolean(summary?.company && summary.role && summary.role.toLowerCase() !== 'unknown');
// ---------------------------------------------------------------------------
// Save report
// ---------------------------------------------------------------------------
if (saveReport) {
let reservedNumbers = [];
let saveStage = 'report';
try {
if (!existsSync(PATHS.reports)) {
mkdirSync(PATHS.reports, { recursive: true });
}
reservedNumbers = await reserveReportNumbers(1, { rootDir: DATA_ROOT, reportsDir: PATHS.reports });
const num = formatReportNumber(reservedNumbers[0]);
// LOCAL calendar day (#3070). This one value becomes three things that have to
// agree with each other and with the user's calendar: the report FILENAME
// ({num}-{slug}-{today}.md), the report's own `**Date:**` header, and the date
// column of the tracker row written for it.
//
// On the UTC day an evaluation run on a Sunday evening in the Americas produces
// 042-acme-2026-08-18.md, dated the 18th, in a tracker row dated the 18th —
// while every other date the user sees, and every date the other scripts now
// stamp, says the 17th. The filename is the part that cannot be corrected
// later: reports are addressed by it.
const today = localToday();
const companySlug = slugifyCompany(company);
const filename = `${num}-${companySlug}-${today}.md`;
const reportPath = join(PATHS.reports, filename);
const reportContent = `# Evaluation: ${company} — ${role}
**Date:** ${today}
**Archetype:** ${archetype}
**Score:** ${score}/5
**URL:** ${postingUrl || '(pasted)'}
**Legitimacy:** ${legitimacy}
**PDF:** pending
**Tool:** Ollama (${modelName})
---
${evaluationText.replace(/---SCORE_SUMMARY---[\s\S]*?---END_SUMMARY---/, '').trim()}
`;
writeFileSync(reportPath, withJdArchiveSection(reportContent, jdText), 'utf-8');
console.log(`\n✅ Report saved: reports/${filename}`);
if (!trackable) {
console.warn(`\n⚠️ The reply has no ${summary ? 'company and role in its SCORE_SUMMARY' : 'SCORE_SUMMARY block'}, so no tracker row was written.`);
console.warn(' Re-run the evaluation, or use a model that follows the report format.');
} else {
// AGENTS.md Pipeline Integrity rule 1: never hand the user a row to paste
// into data/applications.md. Evaluations persist as a tracker addition and
// merge-tracker.mjs applies dedup, status validation, report-link
// normalization and the tracker lock. A pasted literal skipped all of that,
// and at 8 cells it was also silently dropped by every reader's width guard.
// Field order is the TSV contract's -- status BEFORE score; merge-tracker
// swaps them into the tracker's own column order, resolved by name.
const additionName = `${num}-${companySlug}.tsv`;
const trackerFields = [
String(parseInt(num, 10)),
today,
tsvSafe(company),
tsvSafe(role),
'Evaluated',
normalizedTrackerScore(score),
'❌',
`[${num}](reports/${filename})`,
tsvSafe(`Ollama evaluation (${modelName})`),
];
// Optional tenth field, labelled in the header below so it resolves by name.
// Pass 0 can then match on it instead of waiting for --backfill-urls.
if (postingUrl) trackerFields.push(tsvSafe(postingUrl));
// Header row first (#3517/#3706): merge-tracker resolves the fields by name,
// so this row cannot be ingested into the wrong columns. The optional URL
// needs its own label -- values are read BY label, so a tenth field the
// header does not name is not mis-mapped, it is dropped.
const additionHeader = postingUrl ? `${TSV_ADDITION_HEADER}\turl` : TSV_ADDITION_HEADER;
saveStage = 'tracker addition';
mkdirSync(PATHS.trackerAdditions, { recursive: true });
writeFileSync(
join(PATHS.trackerAdditions, additionName),
`${additionHeader}\n${trackerFields.join('\t')}\n`,
'utf-8',
);
console.log(`\n📊 Tracker addition saved: batch/tracker-additions/${additionName}`);
console.log(' Run `node merge-tracker.mjs` to merge it into the tracker.');
}
} catch (err) {
console.warn(`⚠️ Could not save evaluation (${saveStage}): ${err.message}`);
process.exitCode = 1;
} finally {
if (reservedNumbers.length > 0) {
try {
await releaseReportNumbers(reservedNumbers, { reportsDir: PATHS.reports });
} catch (err) {
console.warn(`⚠️ Could not release report reservation: ${err.message}`);
}
}
}
}
console.log('\n' + '─'.repeat(66));
console.log(` Score: ${score}/5 | Archetype: ${archetype} | Legitimacy: ${legitimacy}`);
console.log('─'.repeat(66) + '\n');
console.log(formatBreakdown(tracker, modelName, 'ollama'));