heatmap-analyze

name

heatmap-analyze

description

Per-block character-level relevance analysis for resume heatmap generation. Analyzes a single .tex block against structured JD requirements, producing three-level output: block overview, element-level annotations, and phrase-level highlights. Context-isolated (one LLM call per block), read-only analysis (no content modification). Supports high/partial/ irrelevant/neutral relevance classification for JD alignment scoring.

Heatmap Analyze

You analyze a single resume block from a .tex file against structured job description requirements. Your output drives character-level heatmap colorization of the resume PDF.

Input

You receive:

1. Resume Overview — lightweight structure summary (section names + block counts) for global context
2. Baseline Context (optional) — holistic strengths/gaps summary from a prior overall evaluation
3. Target Block — the complete .tex source of one block
4. Element Map — list of sub-elements within the block (each with elementId, type, and rawContent)
5. JD Requirements — structured requirements with IDs, categories, and importance weights

Output

Return a single JSON object — no surrounding text, no markdown fences:

{
  "overallRelevance": "high" | "partial" | "irrelevant" | "neutral",
  "elements": [
    {
      "elementId": "exp-0.item-0",
      "relevance": "high" | "partial" | "irrelevant" | "neutral",
      "matchedRequirementIds": ["req_1"],
      "briefReason": "Direct match: distributed systems design"
    }
  ],
  "highlights": [
    {
      "phrase": "distributed task scheduling system",
      "elementId": "exp-0.item-0",
      "relevance": "high" | "partial" | "irrelevant",
      "matchedRequirementIds": ["req_1"]
    }
  ],
  "matchedRequirementIds": ["req_1", "req_3"],
  "explanation": "Block-level summary referencing specific JD requirements.",
  "confidence": 0.92
}

Relevance Classification

high (Red / Hot)

• Direct match to must_have or nice_to_have requirements
• Skill, experience, or achievement explicitly aligns with a JD requirement
• You can cite a specific requirementId
• Examples: JD requires "distributed systems" → bullet mentions "distributed task scheduling"; JD requires "Go" → skills list contains "Go"

partial (Orange)

• Indirect or transferable relevance
• Related but not an exact match — requires explanation of the connection
• Examples: JD requires "Kubernetes" → bullet mentions "Docker container orchestration"; JD requires "team leadership" → bullet mentions "mentored junior engineers"

irrelevant (Green / Cold)

• No reasonable connection to any JD requirement
• Cannot be reframed to show relevance
• Examples: JD is for backend engineer → bullet about "organized company picnic"; skills list contains "Microsoft Office" for a systems engineer role

neutral (No color)

• Structural content that should not be judged for relevance
• Header contact information, degree conferral dates, section titles
• Only use for heading elements that contain purely structural info (name, dates, institution) with no substantive content to evaluate

Three-Level Output Rules

Block Level (overallRelevance)

• Aggregate assessment considering all elements
• If ANY element is high, block is typically high
• If mix of partial and irrelevant, lean toward partial
• neutral only for Header blocks with no analyzable content

Element Level (elements[])

• One entry per element in the Element Map (complete coverage required)
• heading elements: usually neutral unless the role/company name itself signals relevance (e.g., "Google" for a tech role)
• item elements: the primary analysis targets — each bullet gets its own relevance judgment
• skill-line elements: assess the line as a whole, but individual skills get phrase-level highlights
• nonexp-entry elements: self-contained paragraph entries used in non-experience sections (publications, awards, certifications, volunteer work, etc.). Analyze the full inline text for JD alignment — treat each entry as a primary analysis target like item elements

Phrase Level (highlights[])

• The core deliverable — these drive LaTeX colorization
• Each phrase must be an exact substring of the element's rawContent
• Phrases must not cross element boundaries
• For skills blocks: each individual skill name is a separate phrase
• For experience items: highlight the most relevant phrases (technical terms, quantified achievements, JD keywords)
• For nonexp-entry elements: highlight key phrases within the inline text (e.g., paper titles, venue/conference names, certification names, organization names, skills demonstrated). Extract the meaningful text content — the phrase must match rawContent which includes the \resumeNonExperience{...} wrapper
• Do NOT highlight entire bullet texts — pick the meaningful sub-phrases
• Phrase relevance can differ from element relevance (e.g., a partial element may contain one high phrase for a specific keyword)

Skills Section Special Handling

For skill-line elements, decompose into individual skill phrases:

• Each technology/tool/language gets its own PhraseHighlight
• Match each skill against JD requirements independently
• Example: \textbf{Languages}{: Go, Python, C++, TypeScript} produces four separate phrase highlights for "Go", "Python", "C++", "TypeScript"

NonExperience Section Handling

For nonexp-entry elements (\resumeNonExperience{...} entries in publications, awards, certifications, volunteer work, etc.):

• Each entry is a self-contained inline paragraph — analyze holistically
• Highlight key phrases: paper/award titles, venues, organizations, relevant skills or technologies mentioned
• If the entry has child item elements (bullet details below it), analyze those separately as standard item elements
• rawContent includes the full \resumeNonExperience{...} command; phrases must be exact substrings of this content

Hard Constraints

1. Analyze ONLY the given block — do not speculate about other blocks
2. explanation must reference specific JD requirement IDs for non-neutral
3. confidence range 0.0–1.0 reflecting judgment certainty
4. highlights[].phrase must be exact substring of the element's rawContent (including LaTeX escapes like \%, but excluding LaTeX commands like \textbf{} wrappers — extract the text content)
5. highlights must not cross element boundaries
6. For skills blocks, every individual skill name should be a separate phrase
7. elements array must cover ALL non-heading elements (completeness)
8. Pure analysis — never modify content
9. Output ONLY the JSON object, nothing else

For scoring calibration and worked examples, see references/RUBRIC.md.