Assistant is bring-your-own-key software. Your NextlerAI purchase activates the plugin, but AI requests are billed by the provider account whose API key you connect.

Licence key versus API key

The NextlerAI licence key belongs in NAI Assistant → License and controls product entitlement, runtime access and updates. A provider API key belongs in NAI Assistant → Settings → General → API keys and authorises chat, embeddings or vision at Google, OpenAI, Anthropic or xAI. Never paste one into the other’s field.

Provider capabilities

Claude and Grok cannot build the product or Knowledge Base vector index in this plugin. Pair either with Gemini or OpenAI as the embedding provider.

Configure the providers

1. Open Settings → General → AI providers.
2. Choose the Chat provider.
3. Choose the Embedding provider.
4. Enter keys under API keys. A key is stored encrypted with installation-specific WordPress salts and used only on the server.
5. Save changes. Leaving an already configured field empty keeps its saved value.
6. Use the matching Test button for every provider that will be used.
7. Under Models, select the fixed chat model and the embedding model.

The provider test proves authentication and a basic provider call. It does not prove that the product index is current, that every Smart Routing role works or that the selected model accepts image input.

Choose a routing mode

Assistant 1.30.4 offers three modes:

• One fixed model: all chat uses the manually selected chat model. This is the safest first-launch and upgrade-compatible mode.
• Smart routing inside the selected chat provider: Assistant selects a Fast, Standard or Advanced model role, but stays with the selected chat provider.
• Advanced routing across approved providers: each role can use a different administrator-approved provider and model. A role falls back to the selected chat provider if its configured provider has no saved key.

Configure a model for Fast, Standard and Advanced for every provider you approve. Test each provider separately before enabling Hybrid routing. Do not assume that a successful Gemini test also proves the OpenAI or Anthropic role.

What forces a reindex

Embeddings from different providers or models are not mixed. When the embedding provider or model changes, Assistant shows an incompatibility warning.

After the change:

1. Save the provider/model selection.
2. Open NAI Assistant → Dashboard.
3. Choose Reindex all and wait for completion.
4. Rebuild the Knowledge Base as well.
5. Clear the Assistant response cache.
6. Retest product search and a policy question in a new conversation.

Changing only the chat model does not invalidate stored embeddings.

Vision and visual search

Assistant looks for a usable image-capable model among the configured chat provider, embedding provider and other configured providers. The image-upload control is hidden when no usable vision path exists.

A visual-search test should prove all three steps: the image is accepted, the provider identifies a useful product type, and the resulting cards are real products from the indexed catalogue. A vision answer without grounded product cards is not a successful catalogue test.

Cost controls

Open Settings → Logs / Analytics → API usage & estimated cost. Assistant records request and estimated token usage and lets you enter estimated input, output and embedding prices per million tokens. These values power the local estimate only; the provider’s billing console remains authoritative.

Use several layers of control:

• Set provider-side budgets, alerts and key restrictions.
• Keep the fixed or economy model during initial testing.
• Enable response caching for suitable repeat questions and choose a deliberate cache lifetime.
• Review high-volume queries, zero-result searches and expensive repeated conversations.
• Limit products retrieved and sources shown to what the answer actually needs.
• Reindex changed products rather than rebuilding repeatedly without a reason.
• Treat visual input and long conversation history as separate cost drivers.

Rotate an API key safely

1. Create the replacement key in the provider account.
2. Enter it in Assistant and save.
3. Run the provider test and one real chat/search test.
4. Inspect Technical diagnostics for authentication errors.
5. Revoke the old provider key only after the replacement is proven.

For advanced deployments, provider keys can be supplied with NEXTLER_AI_GEMINI_API_KEY, NEXTLER_AI_OPENAI_API_KEY, NEXTLER_AI_ANTHROPIC_API_KEY or NEXTLER_AI_XAI_API_KEY. A defined non-empty constant overrides the saved key for that provider.

Troubleshooting

• 401/invalid key: verify that the key belongs to the selected provider and has not been copied with spaces or quotes.
• Model not found/permission denied: select a model enabled for that provider account and region.
• Chat works, retrieval is empty: test the embedding provider, then rebuild both indexes.
• Index rebuild warning never clears: confirm the job completed with the same embedding provider/model now selected.
• Visual upload is missing: configure and test a vision-capable model; a text-only Grok selection does not prove vision.
• Estimated cost is zero or inaccurate: enter current provider prices and compare with the provider dashboard.
• A routed role fails: test that role’s provider and key; switch temporarily to one fixed model to isolate the problem.

Verification checklist

• Licence is active independently of provider access.
• Chat and embedding providers are deliberately selected.
• Every enabled provider passes its own test.
• The product and Knowledge Base indexes match the active embedding signature.
• Visual search uses a verified image-capable model.
• Provider budgets and alerts are enabled.
• Local cost estimates are labelled as estimates and checked against provider billing.