Setup & Providers

Pro feature. The AI Settings tab only appears when WP Career Board Pro is installed and active. On Free this page applies once you upgrade. (Pro features work regardless of license status - the license drives automatic updates only.)

WP Career Board Pro chooses AI providers per task: one provider for analysis & ranking (completions) and a separate provider for embedding & matching. Each has its own API key and model. You can switch at any time.

The two roles

Role	Powers	Allowed providers
Analysis & ranking (completions)	Description writer, applicant ranking, TL;DR summaries, cover-letter writer, chat assistant	Anthropic Claude, OpenAI, Ollama
Embedding & matching	Job embeddings, AI Chat Search, candidate matches	OpenAI, Ollama

Claude has no embeddings API, so it can only be the analysis provider. If you want Claude for ranking copy, pair it with OpenAI or Ollama for embeddings.

Provider comparison

Provider	Embeddings model (default)	Completions model (default)	Where it runs	Cost ballpark
OpenAI	text-embedding-3-small	gpt-4o-mini	OpenAI servers (USA)	$0.02 per 1M embedding tokens, $0.15 / 1M input + $0.60 / 1M output for completions
Anthropic Claude	Not supported	claude-sonnet-4-6	Anthropic servers (USA)	Sonnet-tier per-token pricing for completions. Requires OpenAI or Ollama for embeddings.
Ollama	nomic-embed-text	llama3	Your server / your hardware	Free (your compute)

Each provider’s model is selectable on the AI Settings tab - Claude (Haiku / Sonnet / Opus), OpenAI (completion model + embedding model), Ollama (model names). Defaults are listed above.

Recommended default: OpenAI for both roles. Best balance of quality, cost, and zero setup. Use Claude as the analysis provider when you want better description and cover-letter copy. Use Ollama when content must not leave your server.

Cost estimation (OpenAI)

Rough back-of-envelope for a typical mid-size board:

Activity	Tokens per event	Events per month	Monthly cost
Job publish (embedding)	~300 in	100 new jobs	~$0.001
AI Chat Search query	~50 in	5,000 searches	~$0.0005
Applicant ranking + TL;DR	~2500 in / ~180 out	800 scored (cached after first)	~$1.40
Job Description Writer	~500 in / ~500 out	100 generations	~$0.31
Cover-letter writer	~1500 in / ~250 out	300 generations	~$0.50
Total per month			~$2.20

Scores are cached per application, so re-opening the dashboard does not re-bill. For a large board (10x the above), expect $15-$25 / month. Set a monthly cap on your OpenAI billing dashboard - that’s the surest way to catch a runaway loop. Ollama is free but uses your CPU / GPU.

Step 1 - choose providers

1. Go to WP Admin -> Career Board -> Settings -> AI Settings.
2. Analysis & ranking provider - choose Anthropic Claude, OpenAI, Ollama, or None.
3. Embedding & matching provider - choose OpenAI, Ollama, or None.

If you can’t see the AI Settings tab, check that Pro is active.

Step 2 - enter keys (or base URL) and models

Each field on the tab links out to where to get the key, and the Ollama field notes that it is free and self-hosted.

OpenAI

1. Sign in at platform.openai.com.
2. Billing -> Payment methods - add a card. OpenAI requires a payment method even for low-volume use.
3. API keys -> Create new secret key. Copy the sk-... value immediately (you can’t view it again).
4. Paste into the OpenAI API Key field on the AI Settings tab.
5. Optionally set the OpenAI completion model and embedding model (defaults gpt-4o-mini and text-embedding-3-small).
6. Set a Monthly budget on the OpenAI dashboard. Recommendation: $20 / month for a starting board, and raise if you actually hit it.

Anthropic Claude

1. Sign in at console.anthropic.com.
2. Add a payment method.
3. API keys -> Create key. Copy the sk-ant-... value.
4. Paste into the Anthropic API Key field.
5. Optionally pick the Claude model (Haiku / Sonnet / Opus; default claude-sonnet-4-6).
6. Remember: Claude can only be the analysis provider. Set the embedding provider to OpenAI or Ollama as well, or AI Chat Search and candidate matching stay off.

Ollama (self-hosted)

1. Install on your server: curl -fsSL https://ollama.com/install.sh | sh
2. Pull the required models:

   ollama pull nomic-embed-text
   ollama pull llama3

   On a 4 GB VPS this takes a few minutes and uses about 5 GB of disk for both models.
3. Confirm Ollama is reachable: curl http://localhost:11434/api/tags should return a JSON list including both models.
4. In AI Settings, set the Ollama Base URL to http://localhost:11434 (or wherever Ollama is bound). No API key is needed.
5. Optionally set the Ollama completion/embedding model names (defaults llama3 and nomic-embed-text).
6. Server sizing: completions on llama3 require ~6 GB RAM and run slowly on CPU-only servers. For decent latency, run on a host with a GPU or use a smaller quantised variant (llama3:8b-instruct-q4_K_M).

Step 3 - save

Click Save AI Settings. Configuration is stored in these options:

Option	Holds
wcbp_ai_completion_provider	Analysis & ranking provider slug
wcbp_ai_embedding_provider	Embedding & matching provider slug
wcbp_ai_openai_key	OpenAI API key
wcbp_ai_anthropic_key	Anthropic API key
wcbp_ai_ollama_url	Ollama base URL
wcbp_ai_openai_model / wcbp_ai_openai_embedding_model	OpenAI models
wcbp_ai_anthropic_model	Claude model
wcbp_ai_ollama_model / wcbp_ai_ollama_embedding_model	Ollama models
wcbp_ai_auto_rank	Auto-score applicants on submit (on/off)

The legacy single-provider options wcbp_ai_provider, wcbp_ai_api_key, and wcbp_ai_base_url were removed in 1.3.0. AI reads only the per-task options above.

Key security

API keys are never written back into the settings page HTML. Each key field renders empty with a “saved” indicator and only updates when you type a new value. A stored key is not exposed in the page source - so an empty key field after saving is expected, not a bug.

Verifying your setup

There is no separate “test connection” button. Configuration is validated the first time a real AI call happens. To verify:

• Analysis provider: open the post-a-job form and click Generate with AI - a successful generation confirms the completion provider + key.
• Embedding provider: run Index existing jobs (below) and then search on the AI Job Search page - results confirm the embedding provider + key.

Backfilling embeddings for existing jobs

Embeddings are generated automatically at wcb_job_created time. Jobs that existed before AI was enabled have no embeddings and won’t appear in AI Chat Search or candidate matches until you backfill.

The AI Settings tab has an “Index existing jobs” button that calls AiModule::backfill_job_embeddings() and reports how many jobs were embedded. Each call is bounded (up to a few hundred jobs) so the request stays responsive; run it again to continue paging through a large catalog.

The button is disabled with a clear note when no embedding provider (OpenAI or Ollama) is set, and explains why if you try - Claude cannot index for matching.

For very large catalogs you can also loop in WP-CLI:

wp eval '
$ai = new \WCB\Pro\Modules\Ai\AiModule();
$n  = $ai->backfill_job_embeddings( 500 );
echo "Embedded $n jobs\n";
'

Run it off-hours - each call is a provider API request.

Privacy and data flow per provider

Provider	Where data goes	Logged by provider	Used for training?
OpenAI	OpenAI’s US servers via TLS	Yes (kept up to 30 days for abuse monitoring)	No, API data is opted out of training by default
Anthropic Claude	Anthropic’s US servers via TLS	Yes (kept up to 30 days)	No, API data is not used for training
Ollama	Stays on your server. Never leaves.	Only if you log it yourself	N/A

If you’re under GDPR / CCPA / HIPAA constraints, use Ollama for both roles.

Switching providers

You can change providers at any time without losing data:

• Existing job vectors are kept per provider. If you switch the embedding provider from OpenAI (1536-dim) to Ollama (768-dim) the dimensions don’t match and cosine similarity returns 0 for every comparison - so AI Chat Search returns no results until you re-run Index existing jobs against the new provider.
• Existing options are simply overwritten on save. A blank key field keeps the previously stored key (see “Key security” above).

Disabling AI

Set both provider dropdowns to None and save. Effects:

• AiModule::is_enabled() returns false.
• AI Chat Search block renders nothing for visitors (admins see a configure hint).
• The description writer, cover-letter, and ranking controls disappear.
• /ai/match, /candidates/{id}/matches, /ai/ranked-applications/{job_id} return empty lists; /jobs/ai-description and /jobs/{id}/ai-cover-letter return a 503 with wcb_ai_disabled.
• No new embeddings are generated.
• Stored wcb_ai_vectors rows and cached scores are kept (no destructive change). If you re-enable AI later, everything resumes against the existing data.

To fully clean up: deactivate Pro, and on uninstall the AI table goes with it.

Common setup errors

The top three:

• “Invalid API key” - wrong key or wrong provider selected. Re-copy from the provider dashboard. OpenAI keys start sk-, Anthropic keys start sk-ant-. Strip whitespace.
• “Connection refused” (Ollama) - Ollama isn’t running or the Base URL is wrong. ps aux | grep ollama to confirm; restart with systemctl restart ollama if needed.
• “AI is not configured” (503) - the completion provider is None, or set but missing its key. Open AI Settings and finish the analysis & ranking provider.

See 06-troubleshooting.md for the full list.

Where to go next

• 03-candidate-ai-features.md - the candidate-facing flows (AI Chat Search, matches, cover letters).
• 04-employer-ai-features.md - the employer-facing flows (description writer + ranking).