---
phase: 03-tier-3-9-providers
plan: 02
subsystem: providers
tags: [providers, tier-3, specialized, voice, image, embeddings]
requires: [PROV-10-schema, embed-loader]
provides: [PROV-03]
affects: [pkg/providers/registry, engine/detector]
tech_stack_added: []
patterns: [dual-location-yaml, keyword-only-fallback, tight-prefix-regex]
files_created:
  - providers/perplexity.yaml
  - providers/you.yaml
  - providers/voyage.yaml
  - providers/jina.yaml
  - providers/unstructured.yaml
  - providers/assemblyai.yaml
  - providers/deepgram.yaml
  - providers/elevenlabs.yaml
  - providers/stability.yaml
  - providers/runway.yaml
  - providers/midjourney.yaml
  - pkg/providers/definitions/perplexity.yaml
  - pkg/providers/definitions/you.yaml
  - pkg/providers/definitions/voyage.yaml
  - pkg/providers/definitions/jina.yaml
  - pkg/providers/definitions/unstructured.yaml
  - pkg/providers/definitions/assemblyai.yaml
  - pkg/providers/definitions/deepgram.yaml
  - pkg/providers/definitions/elevenlabs.yaml
  - pkg/providers/definitions/stability.yaml
  - pkg/providers/definitions/runway.yaml
  - pkg/providers/definitions/midjourney.yaml
files_modified: []
decisions:
  - "Providers without documented key prefixes (You.com, Unstructured, Runway, Midjourney) use keyword-only detection (no regex) to avoid Phase 2 false-positive regression."
  - "Providers with documented prefixes (Perplexity pplx-, Jina jina_, Voyage pa-, Stability sk-) use tight regex with high/medium confidence."
  - "ElevenLabs/Deepgram/AssemblyAI use hex alphanumeric patterns with low confidence + entropy_min 4.0 — keyword pre-filter guards against noise."
  - "Midjourney has no official API; verify block uses empty URL as sentinel (no active verification possible)."
metrics:
  duration_seconds: 70
  tasks_completed: 2
  files_changed: 22
  completed_at: "2026-04-05T11:42:06Z"
---

# Phase 3 Plan 02: Tier 3 Specialized Providers Summary

11 specialized Tier 3 LLM/AI providers added (search, embeddings, voice, image/video) across dual-location YAML, bringing total Tier 3 count to 12 with pre-existing huggingface.

## What Was Built

### Task 1: Search + Embeddings (commit `7ad9588`)

Added 6 providers covering search APIs and embedding/document-processing services:

| Provider | Type | Detection |
|----------|------|-----------|
| Perplexity AI | Search LLM | `pplx-[A-Za-z0-9]{48,}` (high) |
| You.com | Search | keyword-only |
| Voyage AI | Embeddings | `pa-[A-Za-z0-9_\-]{40,}` (medium) |
| Jina AI | Embeddings | `jina_[A-Za-z0-9]{40,}` (high) |
| Unstructured.io | Doc processing | keyword-only |
| AssemblyAI | Voice (STT) | `[a-f0-9]{32}` (low) |

### Task 2: Voice + Image/Video (commit `0ac12e5`)

Added 5 providers covering speech, image, and video generation:

| Provider | Type | Detection |
|----------|------|-----------|
| Deepgram | Voice (STT) | `[a-f0-9]{40}` (low) |
| ElevenLabs | Voice (TTS) | `[a-f0-9]{32}` (low), `XI_API_KEY` |
| Stability AI | Image | `sk-[A-Za-z0-9]{48}` (medium) |
| Runway | Video | keyword-only |
| Midjourney | Image | keyword-only (no official API) |

All 11 provider YAMLs dual-located (`providers/` + `pkg/providers/definitions/`) to satisfy the embed loader contract.

## Key Decisions

- **Keyword-only where no documented format exists.** Per Phase 3 lessons-learned, providers without distinctive prefixes (You.com, Unstructured, Runway, Midjourney) rely solely on keyword pre-filtering to avoid false positives.
- **Tight regex for documented prefixes.** Perplexity (`pplx-`), Jina (`jina_`), Voyage (`pa-`), Stability (`sk-`) use prefix-anchored regex with high/medium confidence.
- **Low-confidence hex patterns backed by keyword pre-filter.** ElevenLabs, Deepgram, and AssemblyAI use hex-alphanumeric regex (32 or 40 chars) with `confidence: low` and `entropy_min: 4.0` — the Aho-Corasick keyword filter ensures these only fire on matched contexts.
- **Midjourney verify sentinel.** Midjourney has no first-party API; VerifySpec uses empty URL/status fields as a sentinel for "cannot actively verify."

## Verification

- `go test ./pkg/providers/... -count=1` → **PASS**
- `go test ./pkg/engine/... -count=1` → **PASS**
- `diff providers/<name>.yaml pkg/providers/definitions/<name>.yaml` for all 11 providers → identical
- `grep -l 'tier: 3' providers/*.yaml | wc -l` → **12** (PROV-03 satisfied)

## Deviations from Plan

None — plan executed exactly as written.

## Requirements Satisfied

- **PROV-03**: 12 Tier 3 Specialized providers (11 new + pre-existing huggingface)

## Self-Check: PASSED

- All 22 files present on disk.
- Commits `7ad9588` and `0ac12e5` exist on current branch.
- `go test ./pkg/providers/... ./pkg/engine/...` green after each task.