# Phase 4 — Search pipeline, ranking, dedupe, validation

**Status:** Implemented  
**Date:** 2026-10-12  
**OpenAI model:** `gpt-6-luna` (cheap, accurate; `$0.10` / `$0.50` per 1M tokens)

## Pipeline

```
NL query
  → OpenAI interpretation (gpt-6-luna, reasoning_effort=none)
  → merge with user filter edits
  → Indeed fetch (paginated, max 4 pages)
  → local JobListingFilter
  → dedupe by indeedJobKey
  → availability validation (source signals)
  → optional OpenAI / heuristic ranking
  → in-memory cache (15m TTL)
  → UI results
```

## What shipped

| Component | Path |
|-----------|------|
| Orchestrator | `Services/Search/JobSearchPipeline.swift` |
| Local filter | `Services/Search/JobListingFilter.swift` |
| Dedupe | `Services/Search/JobListingDeduper.swift` |
| Criteria merge | `Services/Search/SearchCriteriaMerge.swift` |
| Ranking | `Services/OpenAI/OpenAIRankingService.swift` |
| Validation | `Services/Validation/ListingValidationServices.swift` |
| Cache | `Services/Indeed/InMemoryIndeedListingCache.swift` |
| Model default | `Services/OpenAI/OpenAIModels.swift` → `gpt-6-luna` |

## Indeed source

Live DI uses **`MockJobDataSource`** (via `JobDataSourceIndeedAdapter`) with openable Indeed **browse/search** URLs until the licensed Option C vendor is ready. Fixtures remain for mapper/contract tests.

**Why links used to look “expired”:** interim keys like `fixture-ios-…` / `ios001` are not real Indeed `jk` values. Opening `/viewjob?jk=…` with those keys shows Indeed’s expired/not-found page. The app now opens `/jobs?q=…&l=…` (live Indeed search for that role) when the key is not a real 16-char hex `jk`.

## Open on Indeed

`JobListingOpener` + `OpenIndeedLinkButton` are available on **AI Job Search** rows (and Saved Jobs). Xcode console logs (`JobOpen` / `SearchPipeline`) explain fallbacks and unavailable listings.

## Cache / TTL

| Policy | Value | Use |
|--------|-------|-----|
| Listing metadata cache | **15 minutes** | Search result pages in `InMemoryIndeedListingCache` |
| Saved job revalidation | **24 hours** | Reserved for Phase 5 persistence |

## Feature flags

```swift
FeatureFlags.isAIInterpretationEnabled  // NL → criteria
FeatureFlags.isAIRankingEnabled         // optional batch rank (degrades to heuristic)
FeatureFlags.isListingValidationEnabled // source availability pass
```

HTTP HEAD validation exists (`HTTPJobValidationService`) but is **not wired** in live DI yet — fixtures/source signals only, to avoid unsolicited Indeed probing before vendor expired-signal policy.

## Acceptance checklist

- [x] Duplicate listings collapsed by `indeedJobKey`
- [x] Shared local filter for Indeed-oriented fields (incl. freshness / exclude / exact phrase best-effort)
- [x] Open Indeed URLs remain regional Indeed hosts from fixtures/mapper
- [x] OpenAI uses `gpt-6-luna` for interpretation + ranking
- [x] Pipeline wired into `JobSearchViewModel`
- [x] Unit tests for filter, dedupe, merge, ranking order, pipeline, cache