A generic hiring ad like «پرستار درمانگاه» was published as a dated SHIFT with an invented date
(«فردا») and default hours («۰۸:۰۰–۱۴:۰۰») the source never stated — because classification defaulted
to Shift. Now a dated Shift is only produced when the text carries an explicit shift signal
(شیفت/آنکال/کشیک/نوبت); everything else is an ongoing hiring post → Job (no date to invent). Fixed in
both the parser default and the Publish branch (so an AI mislabel can''t force a shift either).
ReclassifyMisclassifiedShiftsAsync (in the post-ingest auto-cleanup) converts the existing signal-less
aggregated shifts into jobs in place — copies the content to a JobOpening and archives the old shift
(its URL 410s). After one pass it''s a no-op since new ads no longer become shifts.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parser: most jobs read «توافقی» because the amount extractor only saw 6–10 digit numbers, missing
the way Iranian ads actually state pay — «۱۵ تومان»، «۴۰ تا ۵۰ تومان»، «۲۰ میلیون»، «۲۰م» all mean
MILLIONS of toman. Add colloquial detection (1–3 digit number + تومان/م/میلیون → ×1,000,000, lower
bound of a range), guarded so it never matches dates/hours or a long literal-toman figure. Also: a
stated amount now wins over «توافقی» (ads often say a number AND «… بقیه توافقی»).
Backfill: BackfillPayAsync re-parses existing aggregated jobs/talent that have no salary and fills
it in place (no AI, no ID/URL change) — wired into the post-ingest auto-cleanup and exposed as an
admin button. Existing «توافقی» listings with a stated number get their salary; genuinely-negotiable
ads stay توافقی. Also improves the baseSalary in JobPosting rich results.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The person-name extractor was grabbing the word after a title even when it was a verb/filler/
availability/role word, producing garbage headings like «خانم هستم»، «دکتر ام»، «دکتر داروساز
آماده». Stop collecting at a NameNoise word (هستم/ام/آماده/جویای/role words…), so a real name
(«دکتر سپیده علیزاده») still works but these fall back to the role heading. New ingests only;
existing rebuild via the talent reprocess button.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
(1) Specialist guard: the AI sometimes labels a clearly-specialist ad («پزشک متخصص گوش و
حلق و بینی»، «فلوشیپ»، «فوق تخصص») as «پزشک عمومی», so an ENT post published as
«استخدام پزشک عمومی». When the primary role is GP but the ad text names a specialist, swap
it to «پزشک متخصص» (the subspecialty stays as a tag).
(2) Phone type: the landline regex 0\d{2,3} also matched 09xx MOBILE numbers and labeled them
«تلفن ثابت». Iranian landline area codes are 0[1-8]xx (021/026/…), never 09 — restrict it so
mobiles are no longer mislabeled as landlines.
Both apply to new ingests; existing mislabeled rows correct on turnover/reprocess.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cleans up the crawl-generated facility table that surfaced garbage on /Facilities
(«بیمارستان هستم», «... از مدجابز», bare «کلینیک», «سازمان برنامه جنوبی» x3):
- FacilityMatcher.IsJunkName: shared detector for non-names — bare type words, cores
made only of filler/verb tokens, and leaked crawl-source/placeholder text. Added
داروخانه/آسایشگاه to the generic type words so bare ones are caught and dedupe better.
- HeuristicListingParser.ExtractFacilityName now rejects junk candidates (and emoji), so
new ingests fall back to the shared placeholder instead of forging a fake facility.
- IngestionService.MergeAndCleanFacilitiesAsync (+ admin button): folds junk facilities
into the placeholder and merges Persian-fuzzy duplicates into one keeper, repointing
their shifts/jobs first. Hard guard: only purely crawl-generated, unmanaged facilities
are removed — employer-owned and verified facilities are never touched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Tags: parser extracts cert/skill keywords (mmt, ICU/CCU, دیالیز, اتاق عمل,
اورژانس, مسئول فنی, پروانهدار…) + role + city into TalentListing.Tags
(+ migration); shown as chips on cards.
- Deep search on /Talent: «جستجوی عمیق» box does Postgres ILIKE across
tags, description, person, area, role, city (every term must match);
matches are highlighted with <mark> via SearchHighlight.
- Never delete: ShiftStatus.Archived + the admin «بایگانی گروهی» action now
ARCHIVES aggregated posts (hidden from site, kept in DB) and leaves the
raw crawl rows intact — a permanent archive for future analytics.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- ContactMethod entity (Type + Value + SortOrder) 1→N on TalentListing (+ migration).
- Parser extracts ALL contacts: multiple phones + landlines, email, and
socials (Instagram/Telegram/Bale/WhatsApp/website) via URLs and Persian
keyword cues; primary Phone kept for cards.
- ContactInfo helper: per-type label/icon/clickable href (tel:/mailto:/t.me/…).
- Ingestion attaches contacts to each (fanned-out) talent listing; manual
Review re-parses to attach them + the admin-typed phone.
- Talent details renders the full contact list as buttons; falls back to the
single phone, then the Divar source link.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
An ad like «استخدام پرستار سالمند و کودک و همراه بیمار» names several roles;
we kept only the first. Now:
- Parser collects ALL roles (ParsedListing.RoleNames): exact taxonomy
matches (substring-deduped so پرستار⊂پرستار سالمندان) plus synonyms
(سالمند→پرستار سالمندان, کودک/همراه بیمار→پرستار, اتاق عمل→تکنسین اتاق عمل…),
capped at 4.
- Ingestion publishes one Shift/Job/Talent per resolved role (AI role +
parser roles, distinct, capped), so each role is independently
browsable and filterable. RawListing dedupe is unchanged (one raw → N posts).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Parser now reads the currency: ریال amounts (incl. «میلیون ریال» and
numbers with no تومان unit but ≥200M) are converted to تومان (÷10), so
«۴۰۰٬۰۰۰٬۰۰۰ ریال» shows as ۴۰٬۰۰۰٬۰۰۰ تومان instead of 400M.
- Aggregated facility fallback name no longer embeds the source
(«مرکز درمانی (از مدجابز)» → «مرکز درمانی (نامشخص)»).
- Talent details only ever names Divar as a fallback source (when the
number couldn't be extracted); medjobs/telegram are never shown publicly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- HarvestPhones was run over the whole page, so Medjobs' own header/footer
number (09101016110) was appended to every ad. Now harvest only the ad's
description region in Medjobs + Website sources; the protected number
still comes from the reveal call. No more duplicate number across ads.
- The amount extractor read phone digits as a Toman price
(۹,۱۰۱,۰۱۶,۱۱۰ تومان). The parser now strips «شماره تماس…» lines and
mobile/landline numbers before extracting money, and only accepts 6–10
digit numbers with no leading zero (phones/ids start with 0 or are 11+).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
AI (when enabled, now that the server proxy is up):
- AiStructured gains phone, personName, yearsExperience, isLicensed.
- The auditor appends an authoritative output-schema to the admin prompt
so classification stays correct even with an older stored prompt — it
now classifies kind as shift|job|talent and extracts the contact phone
and talent details.
- Ingestion publish prefers the AI's tags (kind/role/city/facility/phone +
talent fields) over the heuristic parser when present.
- Default prompt updated to describe the three kinds + new fields.
Phone extraction from websites (Medjobs / generic sites), where the
number sits behind a "تماس با این آگهی" reveal:
- HtmlUtil.HarvestPhones scans the full markup for tel: links, JSON-LD
"telephone", data-*phone* attributes, and inline Iranian mobile/landline
numbers (Persian digits folded), normalized (mobiles 09…, landlines 0…).
- Medjobs + Website sources append harvested numbers to the ad text so the
parser/AI capture them; manual review then prefills the phone too.
- Parser phone extraction now also captures a landline as a fallback.
Note: if a site loads the number purely via XHR (not in HTML), a
per-source reveal endpoint would be a follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a third listing kind alongside Shift/Job for healthcare staff who
advertise their own availability (very common in Iranian medical
channels, e.g. "دندانپزشک آماده همکاری… ۵۰٪ تسویه"). These have no
facility; the contact phone is the key field.
- Model: TalentListing (role, person name, years, licensed, city/district,
area note, availability, gender, comp, phone) + ListingKind.Talent +
RawListing.LinkedTalentId + DbSet/relations/indexes + EF migration.
- Parser: detect آمادهبهکار/جویای کار → Kind=Talent; extract person name,
years of experience, licensed flag, area («منطقه ۱»), phone. Facility
name extraction now skipped for talent.
- Validator: talent path scores role + phone + medical (no facility/pay
required).
- Ingestion auto-publish: creates a TalentListing for talent kind.
- Review (manual publish): Talent option + talent fields; publishes a
TalentListing without a facility. Shift/Job facility now falls back to a
shared «نامشخص / ثبت نشده» record when the ad names none — publishing
never fails on a missing facility.
- Browse /Talent (indexable, filters: city/district/role/gender),
details /Talent/Details (noindex — personal contact, tel: call button),
_TalentCard, badge-talent, nav link, home section.
- Sitemap includes /Talent; robots disallows /Talent/Details. Archiver
expires stale talent listings.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When publishing a scraped listing we now look for a facility we already
have that is exactly or closely the same, and only create a new one when
there is no match — avoiding duplicates like «بیمارستان میلاد» vs «میلاد».
- ListingParser: extract a facility name (keyword + distinctive words) from
the post and surface it in the parser notes.
- FacilityMatcher: Persian-aware normalization (ي/ك, ZWNJ, punctuation),
type-word stripping for a "core" name, contains + Levenshtein similarity,
and FindBest (same-city exact → any-city exact → same-city fuzzy → fuzzy).
- Review (manual publish): auto-select a matching facility or prefill the
new-facility name; resolve-or-create uses fuzzy match; dropdown preselects.
- IngestionService (auto-publish): reuse FacilityMatcher against a run-wide
facility list (grows as new ones are created) instead of exact-name only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Analyzed live Divar (POST search) and Medjobs (ad_listing sitemaps) data — both are free Persian text. Tighten the medical-relevance gate (drop generic «استخدام»/«شیفت» that match retail/restaurant ads; add clinical terms: بهیار/اتاق عمل/بیهوشی/رادیولوژی/آزمایشگاه/دیالیز/فوریت/تریاژ/… ) so off-topic Divar jobs get flagged, not treated as medical. Add clinical role synonyms in the heuristic parser (بهیار/کمکپرستار/سالمند→پرستار, اتاق عمل→تکنسین اتاق عمل, فوریت→فوریتهای پزشکی, آزمایشگاه→کارشناس آزمایشگاه, فوقتخصص→پزشک متخصص…). Result on live data: Medjobs now yields ~9/30 queue-ready healthcare listings; Divar correctly flags ~72/75 noise for manual review.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
soroush.asadi