If you're building a recruiting tool, an ATS import or an AI screening step, the first problem is always the same: getting clean text out of a resume PDF.
-
pdftotextand most PDF libraries read two-column templates line by line across both columns, so the sidebar's skills end up glued to job descriptions. - Scanned resumes come back empty because there's no text layer.
- Sending every CV to an outside AI service means sending candidates' personal data to a third party.
- Fixed-field resume parsers give you their schema, which rarely matches your ATS.
I built an Apify Actor that handles the first step: Resume & CV Parser: PDF to Markdown & JSON. It runs IBM's open-source Docling layout model and RapidOCR inside the Actor, with the models built in. No OpenAI key is needed, and resumes aren't sent to any outside AI service. You get the whole resume as Markdown with one heading per section, and that format LLMs read well. This post shows how to use it, and how to get your own JSON fields with one extra LLM call.
What you get
One dataset row per resume:
{
"source_url": "https://api.apify.com/v2/key-value-stores/gZUDoFSN6sExqUTEm/records/sample-cv.pdf",
"status": "success",
"candidate_markdown": "## Jordan Ellis\n\n## Senior Data Analyst\n\nAustin, TX, USA | jordan.ellis@example.com | ...",
"processed_at": "2026-11-01T18:09:47.000000+00:00"
}
The Markdown for that sample CV (fictional person) starts like this:
## Jordan Ellis
## Senior Data Analyst
Austin, TX, USA | jordan.ellis@example.com | +1 555 010 0199 | jordanellis.example.com
## SUMMARY
Data analyst with 7 years of experience turning sales, marketing and operations data into dashboards, ...
## EXPERIENCE
## Senior Data Analyst, Northwind Outdoor Supply
Mar 2022 - Present
- Built the first self-serve sales dashboard in Power BI, used every week by 60 store managers.
- Designed a Python demand forecast that reduced stock-outs on the top 100 products by 18%.
## SKILLS
| Area | Tools and methods |
|----------------------|-------------------------------------------------------|
| Programming | SQL, Python (pandas, scikit-learn), R |
| BI and visualisation | Power BI, Tableau, Looker Studio |
What it does with each file:
- Layout first. Columns, sidebars, headings, lists and skill tables stay in reading order. On a two-column test CV, the sidebar came out as its own block, instead of being mixed line by line with the main column.
- OCR only where it's needed. Each page is checked for a text layer. Text pages take a few seconds, and only scanned pages go through OCR. On scanned CVs in French, German, Spanish, Portuguese and Polish the accents came through ("München" stayed "München").
-
Clear failures. A broken link, an HTML page instead of a PDF, or a Word file comes back as
"status": "failed"with a reason like"Download failed with HTTP 404.". Failed rows are not charged.
Option 1: try it in the browser (1 minute)
- Open the Actor page and click Try for free.
- The input is already filled with a sample CV. Replace it with direct links to your own PDFs, one per line.
- Click Start. A one-resume run takes about 20 seconds.
- Open the Markdown view in the Output tab, or download the results as JSON, CSV or Excel.
Links must return the file itself. For Dropbox, end the link with ?dl=1. Signed S3, Azure and Google Cloud Storage URLs work, and tokens in links never show up in the log.
Option 2: Python, from PDF to your own JSON
Install the client with pip install apify-client:
from apify_client import ApifyClient
client = ApifyClient("<YOUR_APIFY_TOKEN>")
run = client.actor("panix/resume-cv-json-parser").call(run_input={
"documentUrls": [
"https://api.apify.com/v2/key-value-stores/gZUDoFSN6sExqUTEm/records/sample-cv.pdf",
],
})
resumes = []
for item in client.dataset(run.default_dataset_id).iterate_items():
if item["status"] == "success":
resumes.append(item["candidate_markdown"])
else:
print("skipped:", item["source_url"], item["error_message"])
The Actor gives you text and doesn't guess fields. If you want fixed fields, add one LLM call with your schema. This step uses whichever LLM you choose. Here's an example with the OpenAI SDK:
import json
from openai import OpenAI
llm = OpenAI()
PROMPT = """Extract from this resume as JSON with keys:
name, email, phone, current_title, skills (list),
jobs (list of {company, title, start, end}), education (list of {school, degree, year}).
Use null when a value is missing. Resume:
"""
for md in resumes:
reply = llm.chat.completions.create(
model="gpt-4o-mini",
response_format={"type": "json_object"},
messages=[{"role": "user", "content": PROMPT + md}],
)
print(json.loads(reply.choices[0].message.content))
Clean, ordered Markdown is a much easier input for the LLM than raw PDF text, which is the point of the parsing step. The same Markdown also goes straight into a vector database (Pinecone, Qdrant, Supabase pgvector) for "find candidates who know dbt and Airflow" search.
Option 3: n8n, Make or Zapier
A typical setup parses resumes that arrive by email:
- Gmail / IMAP trigger for new messages with attachments.
- Upload the attachment to S3, Google Cloud Storage or Dropbox and get a direct link.
-
HTTP Request node:
POST https://api.apify.com/v2/acts/panix~resume-cv-json-parser/run-sync-get-dataset-items, headerAuthorization: Bearer <YOUR_APIFY_TOKEN>, body{"documentUrls": ["<the link>"]}. - An LLM node with the extraction prompt above.
- Google Sheets, Airtable or your ATS's API to store the result.
Make and Zapier have Apify apps with a Run Actor action that does step 3.
What it costs
Pay per parsed resume:
- $0.0135 per resume on Apify's Starter plan ($13.50 per 1,000), $0.015 on the free plan, $0.01 on Business.
- A start fee of $0.004 per run at the default 4 GB of memory.
- Failed downloads, web pages, non-PDF files and unreadable PDFs are free.
One resume costs $0.0175 including the start fee, and 100 resumes in one run cost about $1.35. You can set a maximum cost per run, and the run stops before the resume that would go over it.
Limitations
-
PDF only, up to 50 MB. Save Word files and photos as PDF first. Other file types come back as free
failedrows. - OCR on scanned pages covers English and other Latin-script languages. Text-layer pages work in any language.
- A scanned page takes about 40 seconds, so give big scanned batches a longer run timeout. Text resumes take about 5 seconds each after start-up.
- Output is Markdown, not fixed fields (see the LLM step above).
- Resumes contain personal data. Make sure you have a lawful basis to process them (for example under the GDPR) and delete datasets you no longer need.
If a resume layout comes out wrong, leave a comment or open an issue on the Actor's Issues tab with a sample link. I read them all.
Top comments (0)