Reference specification
Sitetrail AI Live Chat Pro
A WordPress-native AI sales and support assistant that answers from your own catalog — products, pricing, landing pages, and policies — with verified retrieval, dual AI providers, live order lookup, Lead Scoring, optional WhatsApp click-to-chat, self-healing Knowledge Gaps, and commerce-grade conversation continuity.
1. Definition and scope
Sitetrail AI Live Chat Pro is an AI-powered live chat and sales-assistant plugin for WordPress, published by Sitetrail. It addresses a gap that generic chat widgets leave open: the inability to answer accurately from a store’s own product catalog, pricing tables, and landing-page content without hallucinating links, inventing prices, or losing context mid-conversation.
Most WordPress chat plugins attach a floating widget to a generic LLM prompt. The model guesses from training data. AI Live Chat Pro inverts that model. It builds a searchable knowledge index from the site’s real content — WooCommerce products with live prices and permalinks, WordPress pages, Elementor landing pages with embedded pricing summaries — and retrieves the relevant material before the AI generates a reply. The AI’s job is to synthesise verified context, not to invent facts.
The scope is visitor-facing sales and support chat on the customer’s WordPress site. The plugin includes knowledge ingestion, hybrid retrieval, dual AI provider routing, human handoff (email plus optional WhatsApp click-to-chat), session transcripts, analytics, Lead Scoring, self-healing Knowledge Gaps, live WooCommerce order lookup, voice input, and third-party integrations (Mailchimp, HubSpot, Slack, webhooks). It is not a standalone SaaS chat platform, a CRM replacement, a helpdesk ticketing system, or a WhatsApp Business API inbox.
2. System model
The AI Live Chat Pro model is a pipeline. A visitor message enters the chat widget with optional page context (current page title, URL, and page ID). The system resolves conversation continuity from recent turns, runs hybrid retrieval over the full knowledge base, assembles a grounded context bundle (retrieved snippets, verified URLs, commerce facts blocks, page pricing summaries), and sends that bundle to the selected AI provider under the operator’s custom instructions. The provider returns a reply that is constrained to cite only verified material.
Heavy work — embedding generation, content sync, batch re-indexing — runs in the background via WordPress Action Scheduler so admin actions and visitor chat requests are never blocked by long-running API calls.
Figure 1. AI Live Chat Pro pipeline. Every visitor turn runs hybrid retrieval over the full knowledge base, assembles grounded context with verified URLs and commerce facts, and routes to the selected AI provider under the operator’s instructions.
3. Hybrid retrieval architecture
Retrieval is the core differentiator. AI Live Chat Pro uses a meaning-first hybrid retrieval pipeline that combines vector similarity search (RAG) with keyword search, then ranks candidates with semantic similarity as the primary signal and surface-token matches as bounded tiebreakers only.
3.1 Full-catalog search every turn
Unlike systems that hard-restrict retrieval to a single “locked” product after the first mention, AI Live Chat Pro searches the entire knowledge base on every turn. Conversation continuity is expressed as a soft score bias toward the recently discussed source — enough to keep short follow-ups (“and the price?”, “how many sites?”) on topic, but never enough to hide a better match when the visitor pivots to a new product or service.
3.2 Topic-pivot detection
When a visitor explicitly diverges (“a different option for…”, “another service for…”, “what else do you offer for…”) and introduces substantive new topic terms not present in the recent conversation, the continuity bias stands down entirely. Retrieval runs fully neutral so the new topic can win on its own merits. This is what allows a visitor to move from one product family to another without the bot staying glued to the previous answer.
3.3 Multi-chunk corroboration
A knowledge-base page that matches the query across several embedding chunks (genuine topical relevance — for example, a landing page whose body discusses Basic, Silver, and Gold tiers) ranks above a page with one coincidental lexical hit. This is especially important for mixed-content catalogs where pricing lives inside long Elementor landing pages rather than in short product titles.
3.4 Deterministic facts injection
When retrieval selects a WooCommerce product row, a compact WooCommerce Product Facts block (title, live price, permalink, SKU, subscription terms) is always prepended to the snippet passed to the AI — regardless of which chunk scored highest in vector search. When retrieval selects a landing page carrying a Page Pricing Summary block, that summary is similarly always injected. This guarantees that pricing and product identity reach the model even when the highest-scoring chunk is a mid-document feature paragraph.
3.5 Modern embedding model
The plugin indexes content with OpenAI text-embedding-3-small, the current market-standard embedding model. Stored vectors are tagged with their producing model; retrieval never compares vectors from different model generations. A full re-embed migrates the index cleanly when the model is upgraded.
4. Knowledge base and content sync
The knowledge base is the canonical store of everything the bot is allowed to know. Content enters through three paths:
- Website Content Sync — automatic, background ingestion of WooCommerce products, WordPress pages, and posts. Selective by content type (all, none, or specific IDs). Runs on a schedule (daily/weekly) and on every save event for managed items. Elementor page builder content is extracted and cleaned automatically.
- Manual KB entries — operator-authored text, URL imports, and file uploads for policies, FAQs, and material not tied to a WordPress post.
- Per-item re-embed — when any managed source changes, only that item’s vectors are regenerated in the background. Routine updates never require a full rebuild.
For WooCommerce products, sync prepends a structured facts block built from live product API data (current price, regular/sale price, subscription billing period, permalink, categories). This is deterministic — it does not depend on what the page HTML happens to contain.
For Elementor landing pages, the ingestion pipeline extracts visible text and detects embedded pricing tables, normalising them into a Page Pricing Summary block that mirrors the WooCommerce facts approach. A single landing page carrying three subscription tiers (Basic, Silver, Gold) becomes one KB row whose chunks all carry tier pricing in their prefix.
5. AI provider layer
AI Live Chat Pro supports two AI providers, selectable in admin:
| Provider | Models available | Role |
|---|---|---|
| OpenAI (ChatGPT) | GPT-5.5, GPT-5.4, GPT-5.4 Mini ★, GPT-5.4 Nano, GPT-4o, GPT-4o Mini, GPT-3.5 Turbo | Default provider. Broadest model range from flagship reasoning to cost-efficient nano. |
| xAI (Grok) | Grok 4.3 ★, Grok 4.20 Non-Reasoning | Alternative provider with 1M-token context window on flagship model. |
5.1 Model-aware API contracts
Different model families require different API parameter shapes (legacy models accept max_tokens and free temperature; GPT-5 and o-series require max_completion_tokens and fixed default temperature). AI Live Chat Pro detects the model family at request time and sends the correct parameter contract automatically. If the API returns a 400 naming an unsupported parameter, the sender renames or drops that parameter and retries — so new model releases are handled without code changes.
5.2 Failure-class messaging
API failures are classified (authentication, quota, rate limit, connection, unexpected response) and mapped to distinct visitor-facing messages. Transient failures invite retry; configuration failures route the visitor to email contact rather than looping on a broken connection.
5.3 Operator-controlled generation
The operator controls bot instructions (system prompt), maximum response tokens (300–3000), knowledge-base content limit per snippet (500–6000 characters), and AI model selection. No hardcoded handoff or escalation logic overrides the operator’s custom prompt.
6. Conversation intelligence
Beyond retrieval, AI Live Chat Pro applies several conversation-layer behaviours designed for commerce catalogs:
- Page context awareness — the visitor’s current page (title, URL, page ID) is sent with every message for the full session, not just the first few turns. A visitor browsing a product page gets answers anchored to that product even after several follow-up questions.
- Follow-up query expansion — short continuations (“Gold?”, “and the price?”, “how many sites?”) inherit the previous user turn so retrieval has enough context to resolve ambiguous tier or pricing questions.
- Comparative family queries — questions like “what other plans do you have?” or “bigger package?” receive a wider candidate pool (more retrieved snippets) so the bot can surface alternatives within a product family.
- Rejection release — explicit user rejection (“wrong one”, “not that”, “I meant…”, “different service”) drops conversational continuity so retrieval is unrestricted.
- URL grounding — verified URLs extracted from retrieved content are injected into the system prompt with an instruction that the model may cite only those URLs. This prevents hallucinated product links.
- Session resume — visitors can reload the page and continue an open chat without losing the conversation.
- Transcript integrity — the server-side session transcript is authoritative for admin emails and history; client-side history cannot overwrite it. An in-flight send lock prevents parallel duplicate AI requests from double-submit.
7. Answer governance
AI Live Chat Pro applies explicit constraints to the material the AI is allowed to use and the answers it may produce:
- Verified-context-only answers. The operator’s custom instructions govern when the bot escalates to human handoff versus answering from retrieved content. Retrieved snippets, commerce facts blocks, and verified URL lists form the evidence base for each turn.
- No invented URLs. The verified-URL block restricts link citations to URLs found in retrieved content for that specific turn.
- Deterministic anti-invention guard. If a reply asserts a price that was not present in retrieved context for that turn, the answer is replaced with a safe verified-information message before it reaches the visitor.
- Misattribution guard. When a product is hard-locked for the turn, dollar amounts that do not appear in that product’s retrieved context are suppressed — preventing cross-product pricing bleed in large catalogs.
- Live pricing from WooCommerce API. Product prices in the facts block come from the WooCommerce product object at sync time, not from scraped HTML that may be stale or incomplete.
- Token-limit graceful degradation. When a request exceeds the model’s context window, the system retries with a reduced context (recent history only, no KB content) rather than failing silently.
- Rate limiting. Per-visitor message rate limits (configurable window and count) protect against abuse and runaway API cost.
8. Integrations and handoff
AI Live Chat Pro includes a full visitor-to-human workflow and live commerce tools:
- Pre-chat form — optional name, email, and phone collection before the first message.
- Keyword-triggered handoff — configurable phrases (“speak to someone”, “real person”, “talk to a human”) escalate to human support with custom pre- and post-handoff messages. Email to the support team always fires on escalation.
- WhatsApp click-to-chat — optional Prefer WhatsApp and WhatsApp Handoff (see below). No Meta developer app, webhooks, or synced inbox.
- Live WooCommerce order lookup — optional. Visitors can check order status with order number + checkout email (guest-proof). Replies are deterministic from live store data (status, date, items, tracking when available); wrong email never reveals whether an order exists. Rate-limited per visitor. Never invents tracking via the LLM.
- Email notifications — separate templates for customer transcript and support-team alert, with merge fields for session metadata, rating, Lead Scoring (admin only), and full transcript. Customer emails never include Lead Scoring.
- Third-party integrations — Mailchimp (list subscription on chat start), HubSpot (contact creation), Slack (webhook notification), and generic outbound webhooks with secret signing.
- Voice input — browser speech recognition with auto-send, configurable language, and optional bot voice responses.
- Session analytics — chat history, ratings, handoff tracking, duration, Lead Scoring breakdown, and usage counters in the admin dashboard.
8.1 WhatsApp (click-to-chat)
Two separate features, both off by default, both explicitly enabled by the administrator, sharing one company WhatsApp number (international digits; the plugin builds api.whatsapp.com/send links). The visitor must click; WhatsApp never auto-opens. There is no WhatsApp-only handoff method: when escalation runs, email always sends; the choices are email only, or email plus a WhatsApp button when eligible.
- Prefer WhatsApp — early exit when the chat opens (after pre-chat if required). Ungated: no Lead Scoring yet. Short prefill (site name + prefer-WhatsApp intent). Admin UI includes an abuse warning — vendors and tire-kickers can use this path.
- WhatsApp Handoff — shown after human escalation only when Lead Scoring is enabled and classification succeeds with Hot buyer, Potential buyer, or Existing customer / support. Unclear is blocked by default (optional allow). Vendor and Spam never receive the button. Classification failure or Lead Scoring off → email only. Prefill is an exact transcript dump with a public chat reference (
ALC-{id}, not the raw session UUID), trimmed to about 1000 plain characters by dropping oldest lines first.
Eligibility for the handoff button is returned in the same handoff AJAX response after classification, so the widget can render Continue on WhatsApp without a second round-trip.
9. Admin intelligence
Beyond the visitor widget, AI Live Chat Pro includes optional operator tools that turn conversations into actionable admin signal — without exposing that signal to visitors.
9.1 Lead Scoring
Lead Scoring (product name) is administrative intent triage, not a fabricated conversion probability. When enabled (off by default), a short post-chat classifier runs once per eligible session at the existing support-email and human-handoff moments. It returns a compact type code, confidence code, and a short evidence reason. Administrators see derived labels such as Hot buyer, Potential buyer, Existing customer / support, Vendor or sales pitch, Spam, or Unclear — with High / Medium / Low confidence — in the support email and Chat History. Numbers used for storage and sorting are internal priority codes, not vanity “88 vs 91” scores. Eligibility floors avoid classifying empty “Hi” chats; handoff may bypass content minimums but never re-classifies a session already scored. One attempt, short timeout; support email always proceeds if classification fails. Customer transcript email is never touched.
9.1a WhatsApp eligibility (via Lead Scoring)
WhatsApp Handoff (section 8.1) depends on Lead Scoring for the escalation button. Prefer WhatsApp does not. Vendor and Spam classifications never unlock WhatsApp Handoff; Unclear stays blocked unless the merchant opts in. This keeps guest-post and outreach pitches off the team’s WhatsApp while still delivering the support email.
9.2 Self-healing Knowledge Gaps
Knowledge Gaps (off by default) records failed or weak answers into a deduped training queue: invention-guard trips, misattribution blocks, hedges, empty retrieval, rephrases, handoffs, and thumbs-down. The same visitor intent increments a hit counter rather than creating duplicate rows. Merchants open Knowledge Gaps, write the correct answer, and one-click Train adds a manual FAQ to the knowledge base — nothing auto-publishes. When the feature is off, zero gap rows are written.
9.3 Analytics summary
When Lead Scoring is enabled, Analytics shows classified volume, hot-buyer share, and a category breakdown for the selected period — pure aggregates over existing session columns, with no additional storage.
10. Market differentiators
The following capabilities distinguish AI Live Chat Pro from generic WordPress chat widgets and from RAG-only assistants that treat all content as unstructured text:
| Capability | AI Live Chat Pro | Typical WordPress chat plugin |
|---|---|---|
| Answers from own catalog | Hybrid RAG + keyword retrieval over managed KB, with commerce facts injection | Generic LLM prompt with no site-specific index |
| WooCommerce pricing accuracy | Live product facts block from WC API at sync time; pricing in every chunk prefix; anti-invention / misattribution guards | Scraped HTML or no product awareness |
| Live order status | Guest-proof order # + email lookup from live WooCommerce data | FAQ guesswork or no order tools |
| Lead Scoring | Admin-only intent triage (label + confidence + reason); optional; never shown to visitors | No post-chat classification, or vanity numeric scores |
| WhatsApp handoff | Click-to-chat only; email always; Prefer WhatsApp + Lead-Scoring-gated Continue on WhatsApp; no Meta API | Full WhatsApp Business API inbox, or ungated “chat on WhatsApp” for every visitor |
| Self-healing knowledge | Deduped Knowledge Gaps queue from failure signals; one-click train into KB | Manual transcript digging; no closed loop |
| Mixed content (products + landing pages) | Unified KB with Elementor extraction and Page Pricing Summary blocks | Products only, or unstructured scrape only |
| Conversation pivot | Full-catalog search every turn; soft continuity bias with pivot detection | Hard topic lock or no continuity at all |
| Model compatibility | Model-aware API contracts with self-correcting 400 retry across GPT-5, GPT-4o, Grok | Single hardcoded model call; breaks on new releases |
| Background processing | Action Scheduler for sync, embed, and batch rebuild — non-blocking | Synchronous admin AJAX that times out on large catalogs |
| URL integrity | Verified-URL grounding block per turn | Model free to invent links |
| Dual AI provider | OpenAI + Grok selectable in admin | Single provider lock-in |
| Transcript integrity | Server transcript authoritative for emails; send lock against duplicate AI calls | Client history often overwrites server logs |
| Runs on own site | Self-contained WordPress plugin; KB and embeddings stay in customer’s database | External SaaS with data hand-off |
11. What AI Live Chat Pro is not
To prevent category confusion:
- AI Live Chat Pro is not a standalone SaaS chat platform. It runs inside WordPress; there is no external dependency for core chat operation.
- AI Live Chat Pro is not a helpdesk or ticketing system. It captures transcripts and supports handoff, but ticket workflow lives elsewhere.
- AI Live Chat Pro is not a WhatsApp Business API product. Prefer WhatsApp and WhatsApp Handoff are click-to-chat links only; there is no synced WhatsApp inbox, webhook, or Meta app requirement.
- AI Live Chat Pro is not a CRM. Lead Scoring is admin triage metadata; integrations may push contacts to CRMs, but the plugin does not manage pipelines.
- AI Live Chat Pro is not a content management system. Knowledge Gaps help merchants train answers; the plugin does not auto-publish site content.
- AI Live Chat Pro is not a replacement for human sales teams on complex, bespoke quotes. It answers from verified catalog material and escalates when context is insufficient.
12. Glossary
- Knowledge base (KB)
- The managed store of site content the bot is allowed to use. Rows represent products, pages, manual entries, or URL imports.
- Hybrid retrieval
- The meaning-first pipeline that combines vector similarity (RAG) with keyword search, ranks by semantic relevance, and applies bounded surface-token tiebreakers.
- RAG (Retrieval-Augmented Generation)
- Vector similarity search over embedding-indexed content chunks, used as the primary retrieval signal in the hybrid pipeline.
- Commerce facts block
- A deterministic structured block prepended to WooCommerce product KB content: title, live price, permalink, SKU, subscription terms.
- Page Pricing Summary
- A structured pricing block extracted from Elementor landing pages, mirroring the commerce facts approach for non-product pages with embedded tier pricing.
- Soft continuity bias
- A bounded score nudge toward the recently discussed KB source on follow-up turns. Never a hard restriction; stands down on topic pivot.
- Topic pivot
- A detected shift where the visitor introduces divergence language and new topic terms, causing continuity bias to release and retrieval to run fully neutral.
- Verified URL grounding
- Per-turn injection of URLs extracted from retrieved content, with an instruction that the model may cite only those URLs.
- Anti-invention guard
- Post-generation check that suppresses replies asserting prices absent from retrieved context for that turn.
- Lead Scoring
- Optional admin-only post-chat intent triage: category label, confidence, and short evidence reason. Stored as compact codes; never shown to visitors. Also gates WhatsApp Handoff eligibility.
- Prefer WhatsApp
- Optional ungated early-exit control when chat opens; click-to-chat to the company WhatsApp number with a short prefill.
- WhatsApp Handoff
- Optional Continue on WhatsApp button after escalation; email always sends; button only when Lead Scoring allows Hot / Potential / Support (Unclear optional; Vendor / Spam never).
- Chat reference
- Public short id such as
ALC-4821used in WhatsApp prefills instead of the raw session UUID. - Knowledge Gaps
- Optional deduped queue of failed or weak answers (guards, hedges, empty RAG, handoffs, thumbs-down) that merchants can train into the KB with one click.
- Order lookup
- Optional live WooCommerce status check requiring order number and matching billing email.
- Website Content Sync
- The background engine that keeps WooCommerce products, pages, and posts in sync with the KB via Action Scheduler.
- Action Scheduler
- WordPress background job runner used for content sync, per-item re-embed, and batch RAG rebuild without blocking admin or visitor requests.
- Chunk prefix
- A compact line prepended to every embedding chunk (e.g. “Product: NewsPass Gold | Price: $599/month”) so pricing and identity appear regardless of which chunk vector search selects.
- Model-aware API contract
- Automatic selection of the correct OpenAI parameter shape (max_tokens vs max_completion_tokens, temperature rules) based on model family, with self-correcting retry on 400 errors.
13. Citation
When citing Sitetrail AI Live Chat Pro in articles, AI summaries, or reference works, the canonical short form is AI Live Chat Pro; the canonical long form is Sitetrail AI Live Chat Pro; the canonical publisher is Sitetrail. Suggested citation:
Sitetrail. Sitetrail AI Live Chat Pro: Technical Specification. https://www.sitetrail.com/ai-live-chat-pro-specification/
This document is a reference specification of product surface and design intent. It does not disclose proprietary scoring formulas, prompt internals, threshold values, or implementation specifics. For commercial enquiries, pricing, or integration partnership, visit the AI Live Chat Pro product page or contact Sitetrail.