Current release: 2.2.01
Settings → Lumenare Search → Search & Answers brings Semantic Search and Ask Site settings into one place. Choose how search understands a query, how an answer is written and where visitors can see it. These choices are independent: an AI-written answer does not require meaning reranking, and meaning reranking does not turn on answers.
Including PDFs? Check their text, too. Lumenare supports PDFs with readable, valid text. A PDF can look correct on screen yet have damaged or missing mappings between its drawn letters and extractable text. Scanned, image-only PDFs need OCR before their contents can be searched or analyzed. If analysis stops on a document, follow the PDF inspection, repair and resume steps. A service error or a list of document IDs alone does not establish that a PDF is damaged.
Choose Search Behavior
Under Search behavior, select a default and review the preferences for each search surface. The table shows the assigned profile and the behavior available with its settings.
| Choice | What it does | AI service needed? |
|---|---|---|
| Keyword search | Matches indexed words using your weighting, synonyms and search settings | No |
| Understand questions | Removes question filler and uses approved vocabulary to improve word matching | No |
| Meaning reranking (needs AI) | Reorders existing keyword hits by comparing the query with compatible prepared content | Yes: an embedding service and prepared analyses |
| Meaning reranking + broader recall (planned) | Uses the available meaning reranking; broader candidate discovery remains planned in 2.2.00 | The same embedding requirements |
Meaning reranking can improve the order of pages already retrieved. It cannot discover a page outside that keyword result set in this release. Missing or incompatible analyses, unavailable services and request failures fall back to local search.
The Enable question understanding and vocabulary expansion switch controls the extra local interpretation used by question and meaning modes. Ask Site still performs basic question cleanup when keyword search is selected.
For a starting point without AI charges, choose Understand questions and Sources and cited excerpts, then test questions your published articles answer.
Choose Answer Style
Under Answers, choose Default answer style:
- Sources and cited excerpts returns a supported passage and source links without AI answer writing.
- AI-written answers with citations permits the WordPress AI Client to write an answer from retrieved public passages. Selecting it and saving explicitly allows enabled answer surfaces to make text-generation requests.
The default limits are 5 sources and 120 answer words. You can choose 1–10 sources and 40–240 words. The word cap also limits excerpt answers. Search Profiles can override the answer style and limits for an assigned surface.
AI writing requires WordPress 7.0 or newer with active AI Client support and an available text-generation connector with credentials and any required Lumenare Search approval. Text-generation support is checked when an answer is requested. A ready-looking status panel is not a substitute for testing a question.
The Service and Model selectors in Advanced choose the embedding model for content analysis. Answer writing uses the WordPress AI Client's text-generation service selection separately.
Configure a Content Analysis Connector
You need this setup when you want meaning reranking. Excerpt answers and local question understanding can work without it.
- Install the WordPress AI plugin and a compatible provider connector.
- Add its credentials in Settings → Connectors.
- Grant any required Lumenare Search permission under Tools → Connector Approvals.
- In Search & Answers → Advanced, open AI service and model for content analysis.
- Choose a Service, click Refresh model list, then choose a Model that the connector reports as supporting embeddings.
- Click Save Settings before starting analysis.
The picker offers locally configured and approved services. A connector that writes answers may have no embedding models; an empty model list does not mean you need to type a model ID manually.
Refresh model list contacts the selected service for its model catalog. It sends no site content and generates no answers or embeddings. Opening the settings page, saving settings and checking status do not start analysis or generation requests.
Changing service does not silently choose a replacement model. A saved choice stays visible if a catalog is unavailable; refresh to check it. If a refreshed catalog no longer contains that model, choose an available replacement and save. After changing connector credentials or account, refresh before selecting a new model.
Prepare Content for Meaning Search
First build the normal keyword index in Index Management. That index stores words used to find candidate pages. Analyze site content prepares passages and creates numerical embeddings used to compare meaning. It does not replace the keyword index, train a model or enable public answers.
After saving a valid analysis service and model:
- Click Analyze site content in the same Advanced accordion.
- Keep the tab open while Lumenare prepares public searchable posts in small batches, then sends missing or changed passages to the saved service.
- Watch Preparing posts, Analyzed passages and Pending progress.
- Use Pause to stop this browser run after its current batch. Resume analysis continues from saved progress, including after reopening the page.
- When preparation is complete, choose the desired meaning-reranking preference under Search behavior and save.
Existing compatible analyses are reused. A service or request failure stops the browser run; it does not repeatedly retry paid batches. Use Refresh status, resolve the problem and explicitly resume. Reopening the page or refreshing status does not restart analysis.
In 2.2.01, analysis validates extracted document text, repairs unusable cached text when it can, and reports unresolved documents instead of silently omitting them. For a document-read error, check the upload's existence, permissions, size and extracted text. After repairing the same attachment and checking Test Parse in Media Library, return here and choose Resume analysis. Completed compatible work is reused. If you change which document types are included, save and start a new run for that changed scope. See PDF troubleshooting for the full process.
Changing the service, model, analysis size or passage settings requires saving and preparing content for that configuration. A job stops if its saved configuration changes. Stored counts can include older configurations, so use the pending count for your selected setup rather than assuming every stored analysis is compatible.
Pause controls the browser run. Scheduled analysis may still process eligible work when meaning search is enabled. Administrators can also use the plugin's WP-CLI preparation tools for unattended work.
Choose Where Answers Appear
Under Placement, the two switches work separately:
| Placement | How to use it |
|---|---|
| Enable public Ask Site by default | Add the Lumenare Ask Site block or [lumenare_ask_site] shortcode to a page |
| Show answer cards on full search results pages by default | Show an answer above the ordinary Lumenare search results |
Both are off on a new install. Either can use the chosen answer style without enabling the other. Existing explicit choices survive an upgrade.
Review Search Profiles → Answers Per Search Profile for results-card placement, answer-style and limit overrides. Ask Site and search results may use different profiles; the surface table shows their effective choices.
Answers on Search Results
Cards use the same filtered retrieval as the visible first-page results. They retain the normal result list and pagination and label successful answers Cited excerpt or AI-written summary. There is no second search through a different Ask Site profile.
Cards appear on the first page of supported Lumenare full HTML searches. Later pages, live search dropdowns and independent or nested result loops do not receive cards. Classic themes use the main search-loop hooks; block themes use the top-level inherited Query block. A custom results template that does not use those hooks needs its own placement integration.
When a recognized question has no supported answer, an enabled first-page card shows:
I couldn’t find an answer to your question in this site’s content.
Try a more specific question.
The notice also works with zero results. Ordinary keyword searches do not receive this question notice. The Ask Site form has its own no-support message.
Sources, Citations and Privacy
Answers rely on live public body, excerpt or eligible attachment passages from retrieved site content. Drafts, private and trashed posts, password-protected body text and excluded content cannot support an answer. A title, taxonomy match or high ranking alone is insufficient.
AI writing can assess relevant retrieved passages when the local matcher cannot find an excerpt answer. It sends selected passages rather than the entire corpus and requires supplied citations. That assessment also requires exact supporting quotes from the original passages. Invalid output can fall back to a supported excerpt; results-page cards show the no-answer notice when no supported fallback is available. Ambiguous questions may produce an answer qualified to the subject covered by the sources.
Citations help readers verify an answer. Quote and citation checks establish its source text, rather than guaranteeing the model has interpreted every fact correctly.
Costs and Request Limits
| Operation | What can leave your site |
|---|---|
| Keyword or local question understanding with excerpt answers | No question or answer passages are sent to an AI provider |
| Explicit model-list refresh | A catalog request to the chosen service, without site content |
| Content analysis | Eligible public passages sent to the saved embedding service |
| Meaning reranking | An uncached query sent to the embedding service |
| AI-written answers | The question and selected public passages sent through the WordPress AI Client |
Embedding and text-generation charges depend on your provider. Choosing excerpts only stops AI answer writing; separately enabled meaning reranking can still send queries to its service.
Meaning comparisons are bounded to at most 200 existing hits and 800 passages per request. Explicit meaning reranking for live typing considers at most 20 hits. Live typing needs its own meaning-search preference; a global choice alone does not opt it into provider requests. Answer cards never appear in that dropdown.
Ask Site defaults to five requests per visitor per 60 seconds. Results-page answers use a separate allowance with the shared limits. Query embeddings use a short-lived bounded cache and additional request limits. Provider errors retain local results, and generated answers are not saved as new site content.
Upgrading from Separate Settings
In 2.2.00, the old Ask Site and Semantic Search settings entries lead to the combined Search & Answers area. Existing Ask Site shortcodes, blocks and API routes remain available.
The upgrade preserves profiles, routing, vocabulary, service/model choices and complete prior AI opt-ins. An old combined Q&A search choice becomes a search preference; it does not automatically enable public Ask Site, results cards or AI writing. Review each surface's current choices after upgrading.
Troubleshooting
- Analyze site content is disabled: confirm the connector is configured and approved, refresh and choose an embedding model, then save. Resolve any database-table error shown by the analysis status. JavaScript and administrator settings access are required.
- A connector has no analysis models: it may support text generation without embeddings. Choose a connector that offers embedding models; answer writing can still use a suitable text-generation connector separately.
- Analysis cannot read a document: inspect the listed Media Library attachment, including its file access and extracted text. A visible PDF page does not prove that its text mappings work. Follow the PDF repair steps.
- A service error lists several document IDs: those may be the sources in the entire submitted batch, including healthy documents. Diagnose the service message first; a timeout, quota, rate limit or access failure does not prove those files are damaged.
- Stored analyses exist but meaning reranking does not run: check the selected configuration's pending work, service approval and the effective surface preference. Old or changed content may need preparation again.
- Search results appear without an answer: check card placement, the assigned profile and first-page theme integration. The question also needs supporting public passages; reranking alone does not establish an answer.
- An obvious question has no answer: search for the subject directly, confirm its article is indexed and public, and check that the fact appears in the body or excerpt. Try the full subject name. AI writing can assess retrieved evidence but cannot rescue content that retrieval did not find.