Interactive demo · synthetic dataEvery record on this page is fabricated.

Extraction Without a Schema

Records recognised by shape, not by path, from a GraphQL API that is reshaped without notice.

The production code behind this demo is my own work — design, implementation.

NOW BREAK IT

Apply one, two or all three. The response changes underneath the extractor; the records it returns do not.

THE RESPONSE AS RECEIVED · 2,011 CHARS

{
  "data": {
    "jobAd": {
      "__typename": "JobAd",
      "id": "job-4410",
      "title": "Warehouse Supervisor",
      "reference": "SN-WH-4410",
      "location": {
        "label": "Bandung, West Java"
      },
      "postedAt": "2026-02-02",
      "expiryDate": "2026-03-04",
      "applicationCount": 3,
      "applications": {
        "edges": [
          {
            "node": {
              "__typename": "Candidate",
              "id": "cand-8871",
              "fullName": "Marisol Adeyemi-Okonkwo",
              "email": "m.adeyemi@example.invalid",
              "phone": "+62 811 555 04",
              "appliedDateTime": "2026-02-11T09:12:00+07:00",
              "statusInfo": {
                "label": "Shortlisted"
              },
              "attachments": [
                {
                  "__typename": "CandidateAttachment",
                  "url": "https://files.example.invalid/a/9f2c.pdf",
                  "fileName": "Marisol_Adeyemi_CV.pdf",
                  "type": "RESUME"
                },
                {
                  "__typename": "CandidateAttachment",
                  "url": "https://files.example.invalid/a/9f2d.docx",
                  "fileName": "cover letter.docx",
                  "type": "COVER_LETTER"
                }
              ]
            }
          },
          {
            "node": {
              "__typename": "Candidate",
              "id": "cand-8872",
              "fullName": "Ignatius Broome",
              "email": "i.broome@example.invalid",
              "phone": "+62 811 555 05",
              "appliedDateTime": "2026-02-12T14:40:00+07:00",
              "applicationStatus": "Interviewing",
              "attachments": [
                {
                  "__typename": "CandidateAttachment",
                  "url": "https://files.example.invalid/a/1a04.pdf",
                  "fileName": "broome-cv-2026.pdf"
                }
              ]
            }
          }
        ]
      }
    }
  }
}

EXTRACT.TS · WALKS THE SHAPE

1JOBS

2CANDIDATES

3DOCUMENTS

Recognised by shape, not by path: a name, an id, and a signal that this is a person who applied for something.

NAIVEEXTRACT · HARD-CODES THE PATH

2CANDIDATES

3DOCUMENTS

root.data.jobAd.applications.edges

The path resolved. This is the version almost everyone writes first, and it is correct until the morning the vendor renames a key.

Marisol Adeyemi-Okonkwo
Ignatius Broome

WHAT CAME BACK

Warehouse Supervisor
job-4410 · SN-WH-4410 · Bandung, West Java · CLOSES 2026-03-04 · 3 APPLICATIONS · JobAd

CANDIDATES
NAME	ID	CONTACT	APPLIED	STATUS	JOB
Marisol Adeyemi-Okonkwo	cand-8871	m.adeyemi@example.invalid+62 811 555 04	2026-02-11	Shortlisted	job-4410
Ignatius Broome	cand-8872	i.broome@example.invalid+62 811 555 05	2026-02-12	Interviewing	job-4410

DOCUMENTS
FILE	KIND	READS AS	ATTRIBUTED TO
Marisol_Adeyemi_CV.pdfhttps://files.example.invalid/a/9f2c.pdf	CandidateAttachment RESUME	CV	cand-8871job-4410
cover letter.docxhttps://files.example.invalid/a/9f2d.docx	CandidateAttachment COVER_LETTER	COVER LETTER	cand-8871job-4410
broome-cv-2026.pdfhttps://files.example.invalid/a/1a04.pdf	CandidateAttachment	CV	cand-8872job-4410

P1 · THE ORDINARY CASE

The shape the portal returns most days. Candidates are nested inside the job and their documents are nested inside them, so every record comes out fully attributed: each CV knows which person it belongs to and which job they applied for. Nothing in the extractor knows this path exists — it is inferred from the nesting.

WHY THE TOOL EXISTS, AND ITS LIMITS

The portal deletes applicant attachments 180 days after an ad closes and has no bulk download. That constraint is the first thing the README states.
The password is typed by the human into a real browser. The script never sees or stores a credential; only the resulting session file, which is deleted after.
This is real people's personal data. Store it access-controlled, keep it only as long as hiring needs it.
If you need this supported rather than automated, the vendor sells an official partner application-export API. Use that instead.

What is real, and what is simulated

Runs in your browser

extract.ts — the 200-line production extractor, ported verbatim: the recursive walk, the key-preference lists, both regexes and every guard
looksLikeCandidate / looksLikeJob / looksLikeAttachment — the shape tests unchanged, including the rejections that throw out a company, an account, a location or a currency
Attribution by nesting: an attachment inherits the candidate it was found inside, and that candidate inherits the job
The merge by id that lets three partial responses fill in one record, the 40-level depth guard, and the WeakSet that stops a cycle spinning

extract.parity.test.ts rebuilds a 105,861-case matrix — 105,840 combinations of name, id, signal, contact, attachment and nesting shapes, plus 21 hand-written edge roots — and compares a SHA-256 of every result against a digest generated by running the original extractor in its own repository. It runs in npm run verify, so a widened character class or a reordered key list fails the build.

Simulated

The browser automation around it: the sign-in the human performs, the saved session, the walk across every job ad and applicant
Downloading a CV on sight from a short-lived signed URL, the per-job folders and the CSV index it writes
The three mutation buttons and the hard-coded-path comparison — both written for this page, neither is production code
Every payload, applicant, employer, file and address on this page. The hosts end in .invalid and can never resolve