module Liora: conversational answers and visitor-demand insights

[maximus]

Hi everyone,

I’m releasing Liora, a ProcessWire module that turns unanswered searches and incomplete content pages into useful conversations — and turns those conversations into actionable editorial demand.

Most site search ends in one of two places: a list of matching pages, or an empty result. That works when the visitor already knows the right words and the database contains an exact match. It is less useful when the real question is comparative, contextual or incomplete: “Which bottle fits this occasion?”, “What does this term mean?”, “What should I try next?”, or simply “I expected to find something here.”

Liora adds a second path. It can answer the question, continue with follow-ups, ground the answer in public site content, and record the complete demand signal so an editor can improve the underlying website later.

How it works in practice

I am already using Liora on LQRS, a large ProcessWire drinks-discovery site.

1. Search no longer stops at “no results”

A search such as smoky whisky from Scotland currently has no direct structured record matching the complete phrase. The normal ProcessWire results correctly say that no direct match exists.

On the same page, Liora produces a concise AI overview, explains the relevant styles and regions, and links to useful internal sources such as the Scotch Whisky collection and Highlands region. The visitor can then ask a follow-up without starting again.

This is important: Liora does not replace the deterministic search results or pretend that a generated answer is a canonical database record. The page shows both layers clearly — a conversational overview and the actual structured matches.

2. A content page can answer the next question

At the bottom of selected collection and discovery pages, Liora appears as an “Ask about this collection” surface. The widget knows the current ProcessWire page and can use relevant Atlas excerpts and published Vox discussions when those integrations are enabled.

A static collection therefore becomes a useful starting point for questions about selection, comparison, terminology or pairings, while the original page remains the source of truth.

3. A 404 becomes a recovery path

On the LQRS 404 page, Liora asks what the visitor expected to find. It can help locate a product, recipe, article or guide instead of leaving the visitor at a dead end.

4. Questions become an editorial backlog

Every tracked frontend conversation appears in Setup → Liora Insights. Editors can see repeated unmet searches, the source page, referrer, conversation history, status, model, response time, token usage, cache state, retrieval sources and failures.

This is the part I find most useful. The generated answer is immediate help for one visitor; the collected demand shows what the site should fix for everyone. Repeated questions can become better structured data, a missing product page, a clearer explanation, a new guide, improved navigation or a search synonym.

What Liora includes

• A reusable frontend conversation widget and InputfieldLiora.
• Normal JSON responses and real-time streamed responses.
• One tracked Thread per visitor conversation with chronological messages.
• Follow-up continuity that preserves recent constraints.
• Automatic page, source and referrer attribution.
• Setup → Liora Insights for review, diagnostics and editorial status.
• Repeated unmet-demand reporting.
• Optional Atlas retrieval from indexed public site content.
• Optional published Vox reviews, questions, replies and discussions.
• Optional live public search through compatible Squad providers.
• Optional coarse GeoIP enrichment.
• Validated light and dark themes.
• Safe Markdown rendering and verified same-site links.
• Localized widget text and ready-made language presets.
• A tracked JSON endpoint for completely custom frontends.
• Public PHP APIs for direct application use.

Architecture

Liora deliberately does not store provider credentials or implement low-level AI transport. That remains the responsibility of Squad.

• Liora owns conversations, tracking, retrieval orchestration, the widget and Insights.
• Squad owns providers, credentials, model discovery, requests and streaming transport.
• Atlas can supply excerpts from indexed public site content.
• Vox can supply published community evidence attached to relevant pages.
• ProcessWire remains responsible for pages, permissions, sessions, localization and persistence.

Atlas, Vox, GeoIP and live web search are optional. Liora works with Squad alone and fails back cleanly when an optional context source is unavailable.

Basic integration

<?php namespace ProcessWire;

if($modules->isInstalled('Liora')) {
    $liora = $modules->get('Liora');

    echo $liora->renderWidget([
        'context' => $page->template->name,
        'sourceUrl' => $page->url,
        'pageId' => $page->id,
        'heading' => 'Still looking? Ask Liora',
        'theme' => 'default',
    ]);
}

The tracked endpoint itself can remain deliberately thin:

<?php namespace ProcessWire;

$modules->get('Liora')->handleEndpoint();

The consuming site decides where the widget belongs. Liora never inserts itself automatically and does not take over routes, content models or publishing decisions.

Privacy and safety

• Provider credentials remain in Squad.
• Raw model HTML is never trusted.
• Atlas excerpts, community content and web results are treated as untrusted reference material.
• Same-site links are validated; external destinations can be restricted.
• CSRF protection, rate limits and conversation ownership checks apply to the tracked endpoint.
• Liora stores a hashed session owner, not a plaintext session identifier.
• Raw IP addresses and browser user agents are not stored.
• Conversation review and destructive actions use separate ProcessWire permissions.
• Stored conversations are preserved on uninstall by default unless destructive uninstall is explicitly enabled.

Requirements

• ProcessWire 3.0.210 or newer
• PHP 8.1 or newer
• An installed and configured Squad module
• A ProcessWire endpoint page for tracked frontend conversations

Installation

1. Copy the Liora directory to /site/modules/Liora/.
2. Refresh modules in ProcessWire and install Liora.
3. Configure at least one active provider in Squad.
4. Review the model, prompt, widget, retrieval, privacy and retention settings.
5. Create or reuse the JSON endpoint page and keep its URL synchronized with Liora configuration.
6. Add the widget, Inputfield or custom frontend only where it belongs in the site architecture.

Live examples

• LQRS search with an AI overview: https://lqrs.com/search/?q=smoky%20whisky%20from%20Scotland
• Page-aware collection widget: https://lqrs.com/collection/best-drinks-of-france/
• Dedicated conversation page: https://lqrs.com/ai/

Links

• GitHub: https://github.com/mxmsmnv/Liora
• Download ZIP: https://github.com/mxmsmnv/Liora/archive/refs/heads/main.zip
• README: https://github.com/mxmsmnv/Liora/blob/main/README.md
• API documentation: https://github.com/mxmsmnv/Liora/blob/main/API.md
• Integration guide: https://github.com/mxmsmnv/Liora/blob/main/docs/INTEGRATION.md
• Issues and support: https://github.com/mxmsmnv/Liora/issues

I would especially appreciate feedback about where this pattern is useful outside search: documentation, support, directories, product catalogues, knowledge bases and editorial sites — and how you would turn real visitor questions into better ProcessWire content.

[Ivan Gretsky]

Wow, @maximus! This is a great idea for a module! I can see it is on LQRS' 404 too. Going to try this!

A few little glitches I've spotted on LQRS:

1. The module fires on every search request and outputs the generated content at the top. Even when there are results. That might be distracting users from real pages.
2. LLM hallucinates links to non-existing pages when I ask for them directions based on my request. Might tell it to only give existing links.

[maximus]

Thanks, Ivan — and good catches.

The AI-first placement is intentional. Liora’s search experience was inspired by products such as Google AI Mode and Yahoo Scout, where a conversational answer becomes the primary way to explore a query and traditional search remains available as another view.

LQRS currently keeps the structured results on the same page and provides navigation between the AI overview and LQRS results. I agree that the distinction between those views could be made clearer, so I’m reviewing whether they should behave as explicit modes rather than two vertically stacked sections.

You are absolutely right about invented internal links. Liora already instructs the model to use only supplied same-site URLs, but that is not sufficient by itself. I’m adding server-side validation so that only verified URLs supplied by the site can become links. Anything else will remain plain text. I’m also reviewing inline source numbering because the verified source list should remain authoritative.

Thanks for testing it and reporting both points.

[Peter Knight]

Hi @maximus this looks brilliant and really useful, actually.

I'd like to implement this soon on my own Modules documentation. I think it could be really helpful to users.
One thing I didn't understand or didn't quite catch was where the AI fetches it's answers.

Does it use existing public pages from the LQRs site as the only source of search results and perhaps you have some editorial guidelines wrapped around that?

For example, one of my clients is a law firm, and the accuracy of the answers would need to be 100% accurate. Any hallucination or incorrect info could result in industry regulations fines etc. Ok, that's an extreme example, but it helps illustrate the point.

Cheers
P

[maximus]

Hi Peter, thank you!

Liora uses the existing website as its knowledge source. First, the site’s public content needs to be indexed using the Atlas module. When a visitor asks a question, Liora retrieves the most relevant indexed content and provides it to the LLM as context for the answer.

On LQRS, the current model is Gemini 3.5 Flash Lite via OpenRouter. The system prompt is configurable, so you can add editorial rules and instruct the model to answer only from the indexed website content.

That said, no LLM can guarantee 100% accuracy. For legal or similarly regulated content, I would treat Liora as a discovery and navigation assistant: show links to the original pages, make it acknowledge when the indexed content is insufficient, and avoid presenting generated responses as professional advice.

@Peter Knight