Why not just send the CV to an LLM?
Prompting a general model with the CV and a JSON schema gets you valid JSON, but
valid JSON can still be wrong. VerbatimCV uses a model too; the difference is everything around it:
layout repair before, code-owned dates and calculations after, a quote for every preference, and
null instead of a guess.
Where direct prompting goes wrong
| Problem | What VerbatimCV does |
|---|---|
| Two-column PDFs reach the model as interleaved text, and it mixes up roles or invents entries. | Detects columns and fixes reading order first; sends a page image when the layout is ambiguous. |
| Missing facts get plausible values ("Remote" because the job was in tech). | Each preference is stated, inferred or absent; absent means null. |
| "2019" becomes "2019-01-01"; durations are approximated. | Dates keep their precision; totals are computed in code with the method stated. |
| Values can't be traced back to the document. | Every preference has an evidence quote; a quote that isn't in the CV caps the confidence. |
| Hidden text tells the model to rank the candidate first. | Hidden and instruction-like text is removed before extraction and reported. |
| Output shape drifts with prompt or model changes. | A versioned schema with fixed enums; breaking changes get a new version. |
The result also lists what an agent should confirm with a person
(agent_guidance.confirm_before_acting), generated by code from the confidence scores,
never by the model or the CV.
Try it
Get a free API key (50 parses a month, no card), then:
curl https://api.verbatimcv.com/v1/parse \
-H "Authorization: Bearer $VERBATIMCV_API_KEY" \
-F file=@cv.pdf
Agents can use the parse_resume tool instead: uvx verbatimcv-mcp. See the
quickstart for Python, JavaScript and MCP setup.