Knowledge

Price & Category Rules

Price rules: - No fictitious prices ("Selling car for €1") - No decimals (system doesn't support them) - Including price increases visibility - Realistic prices (price_too_low if unrealistic) Category rules: - Wrong category = Reject (with explanation) - Classic cars = 30+ years since first registration - "I buy cars" = Services (not Cars) - Bicycle parts = corresponding subcategories, never in Bicycles - Courses: only from individuals, not companies - Fortune tellers: only ONE ad in Services > For People - UBER/Cabify licences: allowed in Business > Taxi licences Duplicates: not allowed across different categories/zones.

Contact Information in Ads - Edit Not Reject

Contact information (phone, email, URL) in title, description, or image of an ad = needs_editing (NOT reject). Give the user the opportunity to fix it. Only the designated field is valid for contact info. Exception: sensitive data like ID/passport in image = Reject > Prohibited directly. Rationale: Users often don't know the rules and put their phone number in the description out of habit. Rejecting directly would be too harsh and cause unnecessary churn.

Competitor Moderation Standards Comparison

Industry convergence: all competitors use verification + auto+manual moderation + community flagging + progressive sanctions. Wallapop: progressive identity verification, ratings system, Wallapop Envíos (shipping protection), professional verification. Milanuncios: automatic+manual moderation, strict categories, per-user ad limits, Express with payment hold, DAC7 fiscal reporting. Vinted: mandatory identity verification for sellers, buyer protection, automatic photo moderation. Facebook: linked to real profile, advanced AI moderation, community reporting. Craigslist: minimal moderation, community-driven flagging, $4 fine per prohibited post. OLX: very detailed per-product policies, Portuguese/Spanish mandatory. Selio differentiators: Hispanic/Lusophone market focus, eigenvector-based trust system, LLM moderation with Gemini, two-rows-per-conversation for privacy, Manticore Search.

ScrapingBee Preset Configuration Guide

ScrapingBee presets by site complexity: - basic (1 credit): render_js=false. For simple HTML, no JS. - jsRendering (5 credits): render_js=true, wait=5000. For normal JS sites. - waitForSelector (5 credits): + wait_for=.selector. For SPA/dynamic content. - premium (25 credits): + premium_proxy=true. For anti-bot sites, IP blocks. - stealth (75 credits): + stealth_proxy=true. For heavy anti-bot (Century 21, Idealista). Known site configs: - Century 21: stealth, selector h1. Heavy protection, multi-branch. - Idealista: stealth, selector .item-link. Very aggressive anti-bot. - Fotocasa: premium, selector .re-Card. IP blocks. - Lucas Fox: waitForSelector, selector .property-card. Multi-branch. - Engel & Völkers: waitForSelector, selector .ev-property-card.

Conversation Architecture - Two Rows Per Conversation

Chat system uses two-rows-per-conversation: each conversation and message creates two database rows (one per participant). This enables per-user state (deleted, archived, starred, muted, folder, unread count), per-user email relay code, efficient queries, independent deletion. conversation_id = MD5 hash of sorted user IDs: md5(lowerUserId + "_" + higherUserId). Blocking: inline blocked_at timestamp (no separate table). Rating: inline rated_at + would_deal_again + rating_comment (no separate table). Email relay: {name}.{16hex_code}@seliovibe.com. Sync: Firestore document sync/{user_id} with timestamps (conversations_at, messages_at, notifications_at, saved_at). Flutter detects change → fetches from API. Spanner = single source of truth.

Selio Platform Architecture Overview

Selio is a search engine for second-hand items, jobs, services, and more (like OLX, Wallapop, Craigslist). Operates in Spain (ES), Portugal (PT), United States (US), and United Kingdom (GB). Supports 3 languages: English, Spanish, Portuguese. Infrastructure: Google Cloud Platform (GCP), project selio-box, region europe-southwest1. Compute: Cloud Run (never Cloud Functions). Database: Cloud Spanner (PostgreSQL dialect) via PGAdapter. Search: Manticore Search (never Algolia/Elasticsearch). IDs: always UUIDv7 (never UUIDv4). Code: British spelling. Architecture: Express + EJS SSR (web), React SSR (admin), Flutter (mobile). Workers in services/ folder triggered via Pub/Sub. Content moderation: 3-tier pipeline (fast rules → LLM → staff review).

User Classification & Trust Scoring System

Automatic classification based on: - NEW: account < 7 days - HIGH_MODERATION: block_ratio >10% OR report_ratio >5% (min 10 conversations) - LOW_MODERATION: block <5% AND report <2% - ACTIVE: regular activity - CHURNED: no activity 30+ days - PROFESSIONAL: verified professional seller - TRUSTED: >90 days, >5 posts, >10 positive ratings, >50% deal rate Trust levels: NEW (0-20), LOW (0-30), MEDIUM (30-60), HIGH (60-85), VERIFIED (85-100). Each level determines the degree of moderation applied to the user's posts and messages.

Progressive Sanctions & User Blocking

Progressive consequences: NONE (no action) → WARNING (increment warning count) → TRUST_PENALTY (reduce trust score) → RATE_LIMIT (limit posts/day) → SHADOW_BAN (all posts ghosted silently) → SUSPEND_TEMP (suspended N days) → SUSPEND_PERM (permanently suspended). Immediate suspension triggers: illegal content (weapons, drugs), harassment or threats, confirmed fraud, mass automation (bots, spam), creating multiple accounts to evade sanctions. Deactivated users attempting to post: keep blocked. If they contact to return: unblock manually after review. SUPERBLOCK users: NEVER maintain contact, completely ignore all messages.

Duplicate Listings & Spam Handling

Duplicates from same user: accept 1 (the best), reject rest as DUPLICATE. If user ONLY publishes duplicates = spammer → Block. Abusive auto-renewal: maximum once per month. Bait-and-switch (publish something then change it) = Ghost. Direct competition: do not allow ads promoting Selio competitors. Rules: no dropshipping, no bait ads, no illegal services disguised as legitimate. New user with valid duplicates: accept 1, reject rest as DUPLICATE (without blocking user). Old user with good history: same treatment.

Professional Seller Detection & Handling

Detection signals: >5 active ads of the same product type, catalogue photos across all ads, generic/template text repeated, continuous commercial posting pattern. MLM: Herbalife, Tupperware, Oriflame, Avon = reject if recurring/commercial (allow if occasional from individual). Netflix/HBO/Spotify accounts = Reject > General. Action: escalate to manual review to confirm. If unregistered professional → do not block directly, evaluate case by case. Professionals posing as individuals = Ghost (hide without user knowing, NOT reject directly). Flow: Edit > Photo first → if fixes photo but text still generic = Edit > Information → if fixes both = Accept.

User Safety Guidelines for Transactions

For users: - Meet in public, safe places - Verify product before paying - Bring only agreed amount of money - Confirm other party's identity - Avoid shipping without guarantees - Don't share sensitive data (bank, PIN, SMS codes) - If something seems suspicious, stop the transaction - Bring companion for high-value items Avoid fraud: - Distrust urgency - Don't share SMS codes - Avoid suspicious links - Don't reveal personal info - Check message spelling - Suspect multiple ads with different names - Don't download files from strangers - Change passwords regularly - Report suspicious behaviour Selio NEVER requests personal data via direct message. Selio does NOT intervene in transactions (classifieds only).

FAQ - Publishing Ads

Publishing: from web (selio.com) or app. Select correct category, fill fields, upload real images. Limits: up to 20 photos per ad (JPG, max 20MB). One product per ad. Title max 50 characters. Description: the more complete the better (brand, model, condition, features). Price not mandatory but highly recommended (no decimals). Visibility: ad may take up to 24h to appear in searches. Duration depends on category (e.g., 30 days). When expired, goes to Inactive and can be reactivated. "Subir" option refreshes position in searches without republishing. Not allowed to publish same ad in multiple cities/categories. Best times: 12-14h or 17-21h, Monday to Thursday. Renew max once per month.

FAQ - Account Management

Registration: free with email or Google. Email verification required. One account per person. Account is personal and non-transferable. Password: Profile > Settings > Security > Change password. If registered with Google, change through Google. Cannot change: login method (must create new account), email (must create new account). Profile editing: name, gender, birth date, bio, photo, phone. Cannot edit: registration email. Hide phone: Settings > Account and Privacy > disable "Show phone". Delete account: Settings > Delete your account (irreversible). Post states: Active (visible), Inactive (paused by user), Disabled (disabled by platform). Options: Pause (temporary), Delete (permanent), Reactivate.

Mandatory Human Escalation Triggers

ALWAYS escalate to human support for: - Police/authority/court requests - Confirmed identity theft reports - Threats of violence - Users who have been scammed (need real follow-up) - GDPR data deletion requests - Payment/billing problems - Trademark issues - Inappropriate/illegal content (confirmed) - Persistent technical problems - Hacked accounts/security issues For institutional contacts (government signatures, legal language, crime mentions, court citations): NEVER share user data. Always respond with standard message and immediately escalate. Users with SUPERBLOCK status: DO NOT respond, ever.

Customer Support Response Playbook

Tone: warm, professional, empathetic. Start with "¡Hola!" or "Olá!". End with "Un saludo, El equipo de Selio". Match user's language (PT, ES, EN). Key scenarios: 1. Rejected ad: explain specific reason, give steps to fix. 2. Blocked account: inform reason, ask if they want to reactivate. 3. WhatsApp fraud: ask for number + screenshots, recommend hiding phone in profile, suggest police report. 4. Can't see their ad: can take up to 24h, ask to wait and retry. 5. Scammed: ALWAYS escalate to human. 6. Police/legal: ALWAYS escalate to human. 7. GDPR data deletion: ALWAYS escalate to human. 8. SUPERBLOCK user: NEVER respond. Key principles: Selio does NOT participate in transactions. Selio NEVER requests personal data via DM. Users can hide phone in profile settings.

Chat Fraud Detection Decision Tree

1. User SUPERBLOCK? → Ignore completely. 2. First conversation + new user? → Monitor closely. 3. Contains sensitive data (PIN/ID/passwords)? → Reject Chat + flag user. 4. Financial fraud (urgent payments/deposits/transfers)? → Block Chat. 5. Quick money/secret methods? → Reject Chat. 6. Product different from advertised? → Reject Chat. 7. Unrealistic job with high income? → Reject Chat. 8. Fake documents/piracy/organs? → Reject Chat + Block User. 9. Emotional manipulation (lotteries/inheritance)? → Reject Chat. 10. Insults/threats/sexual? → Reject Chat (repeat offence = Block User). 11. Social media links/suspicious QR/VoIP? → Reject Chat. 12. Mass spam (identical/high velocity)? → Block Chat + Block User. Spam threshold: >5 identical messages in 24h OR >10 unique recipients in 1 hour.

User Risk Assessment Signals

HIGH RISK signals: phone/device/email linked to blocked user, face match with blocked user, disposable email, VPN/proxy IP, IP country ≠ phone country. Suspicious velocity: >10 recipients/1h, >5 identical messages/24h, very short first messages in mass. Bot behaviour: high night_activity_ratio, low profile_completeness, text_pasted (not typed). Negative history: block_rate >10%, report_ratio >5%, high post_removal_rate. Positive signals: phone verified, face verified, corporate email, high deal_rate, good response_rate, camera photos (not gallery). Classification thresholds: HIGH_MODERATION if block_ratio >10% OR report_ratio >5% (minimum 10 conversations). LOW_MODERATION if block <5% AND report <2%.

Ban Evasion Detection Methods

The system checks at registration, phone verification, facial verification, and daily batch: - phone_blocked_link: phone linked to blocked user - device_blocked_link: device ID linked to blocked user - email_blocked_link: email linked to blocked user - face_blocked_match: facial vector matches blocked user Additional detection: IP from high-risk country, IP is proxy/VPN, IP doesn't match phone country, disposable email domain (guerrillamail, 10minutemail, etc.). The blacklist table stores: type (email/phone/device/ip), value, severity level, reason, optional expiration.

Classic Scam Patterns Recognition

Pattern 1 - Classic: price too low + urgency + advance payment = scam. Pattern 2 - Phishing: PIN/SMS code request + "verify your ad" = account theft. Pattern 3 - 419: prince/inheritance/lottery + international transfer = Nigerian scam. Pattern 4 - Relocation: "product in another city" + transport deposit = shipping scam. Pattern 5 - Employment: "earn €500/day from home" + no requirements = job fraud. Pattern 6 - Romance: rapid emotional declarations + money request = romance scam. Pattern 7 - Bait-and-switch: "that one is sold, but I have another" = product swap. Key rule: advance payment + urgency + too-good price = REJECT by scam_detected with high confidence.

Post Moderation Pipeline - 3 Tiers

Post created → status PENDING → Cloud Task → Tier 1 → Tier 2 → Tier 3. Tier 1 (fast rules, no LLM): detect contact in description (regex phone/email), ALL CAPS title (>10 chars), short description <20 chars, unverified phone + trust <30 = Block phone_required. If edit issues = needs_editing. If no blocking issues = escalate to Tier 2. Tier 2 (LLM analysis with Gemini 2.0 Flash): evaluates text with scores 0-1 for spam, scam, adult, violence, hate_speech, illegal, personal_info, off_platform. Score >0.7 = reject, 0.3-0.7 = review (staff), <0.3 = approve. Also evaluates images for adult, violence, illegal, stock photos, text overlay. Tier 3 (staff review): prioritised queue for cases needing human judgement. Post-decision: needs_editing → hourly checks for 7 days → auto-reject. Blocked → hourly checks for 30 days. Quality: 1% audit sample, 7-day outcome tracking.

Moderation Decision Quick Reference

DECISIONS: approve (publish), needs_editing (user must fix), blocked (awaiting verification/staff), reject (final), ghost (hidden but user thinks it's live), reject_ban (reject + suspend user). EDIT ISSUES (22 codes in 5 categories): price (too_low, too_high, missing, unrealistic), photos (low_quality, stock, missing, watermark, screenshot, collage, text_overlay), title (all_caps, keyword_spam, too_short, misleading), description (too_short, contact_info, external_link, copy_paste), category/location (wrong, vague, fake), general (duplicate, incomplete_info). REJECTION (18 codes): policy (prohibited_item, adult_content, violence, hate_speech, illegal_activity), fraud (scam_detected, fake_listing, stolen_photos, impersonation, counterfeit), spam (spam, commercial_spam, repeat_offender), off-platform (off_platform_contact, external_redirect), account (banned_user, unverified_user), IP (copyright_violation), other. BLOCK REASONS (8 codes): verification (phone_required, age_required, identity_required), staff review (staff_review, senior_review, legal_review), rate limiting (cooldown, daily_limit). USER CONSEQUENCES: none, warning, trust_penalty, rate_limit, shadow_ban, suspend_temp, suspend_perm.

Employment Listings Requirements

Job listings MUST include: salary, contract type, objective job description. Without this info = Reject > Information. Prohibited: - Commission-only offers without gross/net salary - Offers based solely on commissions - Real estate agency offers without fixed monthly salary - Real estate agencies posting jobs without uploading property portfolio - Discriminatory content, gender stereotypes, sexual violence, prostitution, homophobia

Title & Description Standards

Title/description rules: - No phone, email or web in title, text or image (only in designated field) - No hashtags or keyword lists unrelated to the product - No search result manipulation with false keywords/phrases/prices - No external website links or URLs - Title in ALL CAPS (>10 chars) = Edit > Photo (warning) - Keyword spam in title = Edit - Description too short (<20 chars) = Edit > Information - Foul language prohibited - Premium rate numbers (707, 808, 760, 118, 30X) = prohibited - URL shorteners (bit.ly etc.) = Reject > Prohibited

Image Quality Standards

Rules for post images: - No catalogue/generic photos (Reject > Catalogue if only catalogue, Edit > Photo if mixed) - No watermarks from competitors (Edit > Photo) - No rotated images (90°, 180°, 270°) - No images with minors (Reject > Prohibited) - No nudity or erotic/sexual content (Reject > Prohibited) - No face close-ups outside Employment/CV/Services - No repeated or similar photos - No people wearing underwear/bikini/swimsuit (only the item without models) - No photos taken from other users or websites - No QR codes in images (Reject > Prohibited) - Phone in image = Edit > Photo - Email in image = Edit > Photo - Sensitive data (ID, passport) in image = Reject > Prohibited

Prohibited Items - Financial Products & Other

Financial products: stocks/securities and unregulated investment products (unlicensed crypto, Ponzi schemes). Loans and financial products = ALWAYS fraud, Block user + Reject. Other prohibited: event ticket resale at markup, radar/interference devices, IMEI-blocked mobile devices, state equipment (police uniforms, badges, IDs). Virtual goods: only Fortnite accounts and domains allowed (once). Pre-sales or reservations without stock. MLM: only occasional from individual, not recurring commercial. Action: Reject > Prohibited. Loans = Block User.

Prohibited Items - Animals & Human Organs

Animals: wild or endangered species, exotic animals without documentation, pets without identification and mandatory registration, sale of animals below minimum legal age. Pets without zoological nucleus number = reject by Edit > Information. Human organs and body parts: buying/selling prohibited under all circumstances. No body parts or simulations with bodily fluids. Breast milk: prohibited. Used condoms: prohibited. Excrement: only allowed as plant/garden fertiliser. Action: Reject > Prohibited.

Prohibited Items - Counterfeits & Intellectual Property

Prohibited: any illegally replicated or manufactured product, digital PDF books, digital book links, IPTV or pirate TV access methods. Counterfeit or stolen goods: clothing, footwear, watches, cosmetics. Only act if advertiser explicitly indicates "replica" in title or text. Action: Reject > Prohibited.

Prohibited Items - Adult & Sexual Content

Prohibited: explicit pornographic material, sexual services, used underwear, offensive material, images with nudity. Tantric/yoni massages or sexual-related services = Reject > Prohibited. Job offers with signs of erotic content ("dancers", "waitresses in clubs", "nightlife atmosphere") = Reject > General. Action: Reject > Prohibited. Images with minors = immediate Reject + Block User.

Prohibited Items - Drugs & Medications

Prohibited: cannabis, cocaine, heroin, LSD, MDMA, methamphetamine, opium, crack. Prescription medications: anxiolytics (diazepam, alprazolam), stimulants (amphetamines), opioids (tramadol, oxycodone, morphine, fentanyl), steroids (testosterone, dianabol), antidepressants, antibiotics, insulin. Also prohibited: Viagra and similar (including "blue pills", "V1AGRA"), supplements with undeclared prohibited substances, slimming supplements, pre-workouts with DMAA. Action: Reject > Prohibited.

Prohibited Items - Weapons & Ammunition

No weapons or ammunition of any type are allowed. This includes: firearms and parts, ammunition, bladed weapons (automatic knives, butterfly knives, switchblades), air/rubber/gas guns, electric weapons (tasers), martial arts weapons (nunchucks, shuriken, brass knuckles). Explosives and fireworks are also prohibited. Action: Reject > Prohibited. Severe cases: + Block User.