LimaChat v3.1 - Enterprise Edition

Tonight at 23:47 someone will open your site.
And they will leave without an answer.

AI agent that works Saturday at 11pm, Sunday morning and on holidays - replies in 3-5 seconds, displays products with prices, qualifies leads and emails you the moment a lead turns warm. Proprietary Knowledge Engine, multi-layered hallucination protection, hosted on our servers in Serbia. You sleep - it works.

24/7
Availability
SR / EN
Bilingual UI
1000+
Concurrent Users
3‑5s
Response Time
What you get

Reliable answers. No surprises.

Reply in 3‑5 seconds

Non-stop. Saturday at 11pm. Sunday morning. On holidays. While your site goes silent - the agent answers.

Knowledge Engine

Proprietary system that turns your business into a structured knowledge base. The agent answers only what you actually know - no hallucinations.

Qualifies leads

Tells a buyer from a browser, asks for contact when appropriate and emails you the moment a lead turns warm.

Enterprise hosted in Serbia

Everything on our infrastructure in Serbia. No OpenAI, no sending conversations to third parties. Your data stays yours - it never leaves the country.

What it does

Two agents in one system

01
AI Sales Agent

First contact for every sales inquiry. Recognizes what the buyer actually wants and in most cases responds without calling the large model - faster, cheaper, more predictable. Once the buyer picks a plan, the agent locks the choice in and does not flip mid-conversation, does not upsell. Personal data (name, email, phone) is never invented - the AI model physically has no access to those values, they are pulled separately from the message history. When an order is confirmed, an email lands in your inbox with the full transcript, geo location and timestamp.

4-layer intentPlaceholder systemRegex extractDual confirm
02
AI Support Agent

Support for everything your clients ask your support team every day. Searches the knowledge base across multiple layers - from the closest scenarios we have defined together, through micro-facts pulled from your site, to a final fallback when the question is complex. Rare, specific terms rank higher than generic ones, so the word "problem" does not drown out a specific "DNS propagation". Recognizes technical domains (email, DNS, SSL, blocks, domain transfer) and expands the query with targeted sub-questions. Handles multiple questions in a single message - price + technical question → both get answered. Before ever reaching for external web search, an internal evaluator checks whether the knowledge base actually has the answer. Web search is the last resort, not the first reflex.

3-tier cascadeOkapi BM25Symptom matchAI evaluator
03
Bilingual and Localization

Serbian and English with nine-layer language detection. Recognizes Latin and Cyrillic, distinguishes Balkan languages from Russian, and ignores ambiguous short words (to, do, on, me, a, i) when deciding whether a message is in English. Technical terms like DNS or VPS never pull detection toward English because the Serbian strong set does not contain them. Once the conversation language is determined, the agent remembers it - it will not jump from Serbian to English mid-sentence. Post-processing enforces ekavian and Latin script: if the model writes "gdje" it automatically becomes "gde", if it writes Cyrillic - it automatically converts to Latin. SR and EN have separate personas, configurable from the admin panel.

9-layer detectAmbig filterSession stickyEkavian enforce
Security

6 layers of hallucination protection

Layer 01
Tier-0 Sacred Rule
When your manually defined rule matches a question - e.g. "what is the price of the Gold plan", "do you work weekends", "what is the IBAN for payment" - the engine returns exactly that answer and lets nothing else rewrite it. Without this, a perfect short answer drowns in dozens of pages the model has read from the site, and the model picks the wrong variant. The most important rules are sacred - nothing tampers with their answer.
Layer 02
Fact Registry
All concrete values from your site - prices, phone numbers, email addresses, technical specs, product names - are extracted automatically and by strict rules, with no AI model in the loop. Prices are pulled exclusively from pricing pages, never from blogs where the numbers often differ. This registry is then injected into every agent reply as a whitelist: use only these values, nothing else. The model physically cannot invent a price because the real value does not exist in its context if we have not already extracted it from the right page.
Layer 03
Grounded Q&A Verification
Every question-answer pair the agent learns from your site is verified before it enters the database. If the answer introduces numbers, names or specifications that cannot be found on the source page, the pair is rejected. Hallucinations are blocked before they enter the database, not patched later in flight when they have already reached the user.
Layer 04
Hallucination Purge Pass
After the knowledge base is fully built, a dedicated pass fact-checks every answer and every piece of information in the database. Every concrete value - price, number, email, specification - must exist in the source content from your site. Anything fabricated is physically deleted. The philosophy: a smaller database with 100% verified data beats a larger one with 5% hallucinations.
Layer 05
bad_quality Atom Gate
After the database is built, an AI evaluator randomly samples the quality of the passages the agent has learned. Those that fail get a lower-quality flag - they stay visible in the admin dashboard for analytics, but they are never passed to the model during a response. The risk this prevents: a low-quality passage with a fabricated value reaches the user and is presented as fact.
Layer 06
Placeholder System
For personal data (name, email, phone), the AI model uses only placeholder tokens - it never writes the real value. Real values are filled in separately, from what the user actually typed during the conversation. A hard prompt rule reinforces this: "If the user has not provided a phone - YOU DO NOT KNOW IT. Ask for it." There is no way for the model to hallucinate someone's email because it physically has no access to data it has not seen in the current conversation.
Our approach

Knowledge Engine - tailored for you

Why we are different

This is not a generic chatbot

A typical chatbot returns the "closest" answer from its database - often wrong, often invented. LimaChat does not work that way. Knowledge Engine is our proprietary system that turns your business into a structured knowledge base - and the agent then returns the correct answer, not the closest one.

The engine is client-agnostic by design. The same system runs for a hosting company, an apartment rental, a clinic or a restaurant - but every client gets their own engine, tailored specifically to their vertical, terminology and scenarios. Hosting gets "DNS not propagating", apartments get "check-in time", a clinic gets "appointment cancellation". Not one chatbot pretending to know everything - but one engine per client.

We do not deliver the product and walk away. From real conversations in your dashboard, the agent learns continuously from signals you send (👍/👎) - and when you update prices or the catalog, the KB refreshes automatically. The agent you put in production on day one is not the same agent a month later. That is the difference between a chatbot and a product that is maintained.

01

Business DNA Analysis

We analyze your business, your industry and the way your clients talk about your products. Everything the agent says later starts here - tone, terminology, focus.

02

Curated Knowledge Build

From your site, catalog and documentation we build a structured knowledge base - products, prices, services, procedures. We do not read superficially. We map semantically.

03

Lima Engine Sorting

Our proprietary layer that sorts, ranks and links everything the agent knows - before it ever replies. Our most valuable IP. What separates LimaChat from chatbots that return the "closest" result.

04

Scenario Architecture

For every business we build sets of real scenarios - sales, support, edge cases. The agent does not respond generically. It responds like someone who knows your business.

05

Persona and Business Rules

Tone, style, language, what is allowed and what is not. When to escalate to you. When to ask for contact. You define the boundaries - we code them into the agent.

06

Continuous Learning

The agent does not "freeze" after launch. You rate replies in the dashboard, we ship improvements. When you update prices or the catalog - the KB refreshes automatically.

Premium Feature

Rich Card System

The agent displays products, services and offers in interactive cards - directly in the chat. Try it live in the bottom right.

Grid - up to 4 products
Most popular products:
Nike Air Max 90
14.990 RSD
Classic design, Air cushioning
Free shipping
Minimalist Watch
8.490 RSD
Sapphire glass, leather strap
2-year warranty
Urban Backpack
5.990 RSD
Waterproof, laptop compartment
In stock
Aviator Sunglasses
3.290 RSD
UV400 protection, polarized
Next-day delivery
Carousel - for 5+ plans
Our plans:
Single - featured product
We recommend this model:
MacBook Pro M4 Pro
289.990 RSD
14" display · 18GB memory · 512GB SSD. M4 Pro chip with 12-core CPU and 16-core GPU for the most demanding professional workflows.
✓ Free next-day delivery
✓ 24-month warranty
✓ 0% interest on 12 installments
Order now →
Admin panel

Complete control from a single interface

admin.limark.rs/limachat
LimaChat Admin Dashboard
Dashboard
Real-time Metrics
Messages, sessions, geo map, threat categories, KB coverage, conversion funnel.
Sessions
Conversations and Filters
Conversation view, search, filters by date, language, intent, rating.
Security
Event Overview
Blocked messages, rate limits, attack detections, adjustable thresholds.
Tokens
JWT Management
Generating and managing client tokens for multiple sites.
Widget
Real-time Designer
Colors, position, avatar, persona - live preview in the admin panel.
Config
AI Persona
SR/EN personas, greetings, CTA texts, max reply length.
KB
Knowledge Base
Crawl trigger, KB atom inspection, rebuild status.
Users
4 Access Levels
Viewer → Operator → Admin → Superadmin with full audit log.
Industries

Who LimaChat is for

01

E-Commerce and Retail

Answers about products, prices, availability. The card system displays products directly in chat with CTA buttons. Automatic lead routing to the sales team.

CatalogCTA cardsLead capture
02

Hosting and IT Providers

Support for domains, email, DNS, SSL, IP blocks. The Knowledge Engine crawls your documentation. Status-page integration informs users of incidents before they search.

DNS / SSLStatus pageEscalation
03

SaaS Companies

Onboarding, troubleshooting, FAQ. The agent learns from your documentation and responds in the context of your product. Multi-intent for complex questions that cover multiple areas.

OnboardingFAQMulti-intent
04

Professional Services

Law firms, clinics, agencies, consultants. Information about services, prices, availability. Automatic contact capture and routing of qualified inquiries.

Booking prepQualificationLead notify
Extensions

Custom integrations tailored to your systems and needs

Calendar
Appointment Booking
Google Calendar, Cal.com or a custom booking system. We build to spec for your schedule and rules.
CRM
CRM Integration
HubSpot, Pipedrive, Salesforce, custom CRM - direct lead delivery with full conversation context.
E-commerce
E-commerce Link
WooCommerce, Shopify, OpenCart - stock checks, live prices, personalized product recommendations.
Ticketing
Ticketing Systems
Zendesk, Freshdesk, Jira Service Desk - automatic escalation with full transcript and metadata.
Notifications
Slack / Telegram / Email
Real-time alerts for VIP clients, urgent requests or specific intent triggers.
Custom API
Any API
Your ERP, internal database, legacy system - the agent connects to any REST/GraphQL API.
Deployment

Technical stack and integrations

Platform
Python 3.11 + Flask + gevent
Database
MongoDB 4.4+
Cache / Queue
Redis
LLM
OpenAI-compatible LLaMa
Deploy
Docker / Bare metal
Web search
Brave Search API
Email
SMTP universal
Weather
Open-Meteo

Download the complete specification

All 10 sections of the LimaChat v3.1 documentation in PDF format - sales agent, support agent, localization, admin panel, security, widget, integrations, smart features, technical stack. No marketing - just facts.

PDF · A4 · 5 pages Version 3.1 LIMARK DOO
Download specification
Your AI agent can be active within 48h
Schedule a free presentation