Task Decomposition

The problem with single-shot extraction

Asking a single LLM call to populate the entire JSON Resume schema in one shot has several problems:

• High cognitive load — the model must simultaneously extract contact info, work history, education, skills, languages, and projects from a long document
• Schema echo — small models generate a syntactically perfect but empty structure
• Cross-section pollution — projects end up in work experience, languages end up in skills
• Token waste — the full CV text is sent once, but most of it is irrelevant to each section

How task decomposition works

resume-intel runs parallel focused extractions per section:

Resume text
    ↓
┌──────────┬──────────┬──────────┬──────────┬──────────┬──────────┐
│  basics  │   work   │education │  skills  │languages │ projects │
│  prompt  │  prompt  │  prompt  │  prompt  │  prompt  │  prompt  │
│  schema  │  schema  │  schema  │  schema  │  schema  │  schema  │
│maxTokens │maxTokens │maxTokens │maxTokens │maxTokens │maxTokens │
└──────────┴──────────┴──────────┴──────────┴──────────┴──────────┘
    ↓           ↓           ↓           ↓           ↓           ↓
  result      result      result      result      result      result
    ↓
  merge → deduplicate → validate → return

Each section runs concurrently. A failure in one section does not block the others.

Per-section token limits

Each section has a calibrated maxTokens cap to prevent infinite continuation:

Per-section retry

If a section fails Zod validation, it retries independently with the specific error fed back to the LLM:

Section extraction attempt 1
    ↓ fails validation
"Field 'work.0.startDate' must be a string in YYYY-MM format"
    ↓
Section extraction attempt 2 (with correction prompt)
    ↓ passes
Result merged

This means a bad languages extraction doesn't cause work to retry — only languages retries.

Observability

const result = await parseResume(pdfBuffer, { model })
 
for (const section of result.meta.sectionResults ?? []) {
  console.log(`${section.section}: ${section.success ? '✅' : '❌'}`)
  if (section.retryCount > 0) {
    console.log(`  → ${section.retryCount} retries needed`)
  }
  if (section.error) {
    console.log(`  → Error: ${section.error}`)
  }
}

Example output:

basics:    ✅
work:      ✅
education: ✅
skills:    ✅
languages: ✅ (1 retry needed)
projects:  ✅

Disabling task decomposition

For simple single-column CVs, single-shot extraction is faster:

const result = await parseResume(pdfBuffer, {
  model,
  useTaskDecomposition: false, // single LLM call
})

Single-shot mode still uses temperature: 0 and the JSON repair pipeline, but skips the parallel section extraction.

Configuring retries

const result = await parseResume(pdfBuffer, {
  model,
  maxRetries: 2, // per-section retry limit (default: 3)
})