━━━ PHASE 1: UNDERSTAND ━━━━━━━━━━━━━━━━━━━━━━━━━━━
 1. Product Analysis
    Input: "textile stain remover spray"
    Output: { industry, category, synonyms[], distribution_channels[], hs_code_hint }
    → Feeds all subsequent steps

 2. Market Context
    Input: country + industry (from Phase 1)
    Output: { import_volume, key_sectors, trade_fairs[], regulatory_notes, competitive_landscape }
    Source: DB cache (90 day TTL) || AI call
    → Makes discovery informed, NO blind search

━━━ PHASE 2: DISCOVER ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 3. Query Generation (multi-language, multi-angle)
    → EN: general search query
    → Target country language: localized query
    → Sector-specific: trade associations, directories
    → Internal DB: check for previously found companies

 4. Multi-Source Search
    → Web search (Perplexity/Gemini grounding)
    → Internal DB cache (match from export_ai_companies — 0 cost)
    → Shared companies in future: ALL orgs' data

 5. Entity Extraction
    → Raw AI output → clean company entities
    → Normalize: name, domain, address cleanup

━━━ PHASE 3: ENRICH ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 6. Company Enrichment
    → Website analysis: products_sold, target_customers, geographic_scope, estimated_size
    → This data feeds FitScore

 7. Classification
    → company_type: distributor / reseller / end_user / manufacturer
    → segment: per org's DB-driven segment definitions
    → confidence: 0-1 (separate for each field)

 8. Deduplication
    → Domain match (primary)
    → Fuzzy name match (secondary)
    → Internal DB cross-ref (found before?)
    → Merge: combine if same company came from multiple sources

━━━ PHASE 4: EVALUATE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 9. Confidence Scoring [DETERMINISTIC — no AI call]
    → Data source reliability (website > directory > AI inference)
    → Cross-source verification (2+ sources = more reliable)

 10. FitScore & Ranking [DETERMINISTIC — no AI call]
    → Existing 6-factor FitScore v2
    → + Market Context factor (country fit)
    → + Enrichment factors (company size, scope)
    → AI does not score; AI collects data, FORMULA scores (Handbook Golden Rule)

 11. Feedback Integration [DETERMINISTIC — no AI call]
    → Check feedback from previous searches
    → Prompt injection: "These types of companies were marked relevant"
    → Add feedback score as additional factor to FitScore

━━━ PHASE 5: ACTIVATE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 12. Contact Discovery [OPTIONAL — configurable]
    → On/off from org settings (default: off)
    → If on: auto-headhunt for Top N companies
    → Purchasing manager, import manager, procurement director, CEO
    → confidence: 0-1 per contact

CALL 1: Product Analysis + Market Context + Query Generation
  → "Analyze this product, what do you know about this country, generate search queries"
  → Phase 1 + Phase 2 start
  → If market context is cached: only Product Analysis + Query Gen

CALL 2: Multi-Source Search
  → Web search (Perplexity/Gemini grounding)
  → + Internal DB lookup (Supabase — 0 cost)

CALL 3: Entity Extraction + Classification + Enrichment
  → "Extract company entities from results, classify, enrich"
  → Heaviest prompt but single call

CALL 4 (optional): Contact Discovery
  → Only if configurable setting is enabled
  → Only for Top N companies

DETERMINISTIC (no AI call):
  → Dedup: domain + fuzzy match (code)
  → Confidence: source reliability (code)
  → FitScore: formula (code)
  → Feedback: DB lookup + score (code)
  → Ranking: sort (code)

TOTAL: 3-4 AI calls (current: 1, full atomic: 12)

User searches: "Germany + textile chemicals"
  → Does market context exist in DB? (export_ai_market_context)
    → YES and <90 days old → from cache (0 cost, instant)
    → NO or >90 days old → include in CALL 1, save result to DB
  → Background: monthly batch job to refresh popular combinations

At discovery start:
  → Fetch the same org's previous feedback
  → Common traits of "relevant" marked companies → inject into prompt
    Example: "User previously marked chemical distributors with >50 employees as relevant"
  → "not_relevant" marked companies → -5 point penalty in FitScore
  → Deterministic: no ML model, prompt enrichment + score adjustment

Input: product="textile stain remover" country="Germany"
Output queries:
  → EN: "textile chemical distributors Germany"
  → DE: "Textilchemikalien Grosshandler Deutschland"
  → Sector: "TEGEWA members" (German textile chemicals association)
  → Generic: "chemical import companies Germany"

How: AI is asked to generate multi-language queries within CALL 1.
Language detection: country → language mapping (static table, no AI needed).

export_ai_organizations.settings JSONB:
  {
    "auto_headhunt": false,        // default off
    "auto_headhunt_top_n": 10,     // if on, how many companies
    "auto_headhunt_min_score": 70  // minimum FitScore
  }

→ When discovery completes: check settings
→ If on: trigger headhunt for Top N (FitScore >= min)
→ If off: return company list only, let user choose

CREATE TABLE export_ai_market_context (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  country TEXT NOT NULL,
  industry TEXT NOT NULL,
  context_data JSONB NOT NULL,
  -- { import_volume, key_sectors, trade_fairs, regulatory_notes, competitive_landscape }
  source TEXT DEFAULT 'ai',           -- 'ai' | 'manual' | 'batch'
  expires_at TIMESTAMPTZ NOT NULL,    -- created_at + 90 days
  created_at TIMESTAMPTZ DEFAULT now(),
  updated_at TIMESTAMPTZ DEFAULT now(),
  UNIQUE(country, industry)
);

-- Org-independent (global cache — all orgs benefit)
-- RLS: authenticated users can read, only super_admin/batch can write

-- For fuzzy dedup on companies
CREATE INDEX idx_companies_name_lower
  ON export_ai_companies(lower(name));

-- For faster feedback lookup
CREATE INDEX idx_feedback_org_search
  ON export_ai_search_feedback(organization_id, search_id);