How Liliwi works
Effortless for merchants, robust under the hood

Set up your store assistant in minutes. Behind the seamless interface lies a powerful, enterprise-ready AI pipeline built for zero-lag product discovery and real-time store actions.

5 min

Setup time to launch

100%

Data training accuracy

Zero

Hallucination rate

<800ms

Median response time

One platform, three jobs

Connect your store in three steps

No engineering team is required. Connect your store, sync your data, and your agent is ready to talk to customers.

1

Onboarding

Connect your store

Install the Liliwi app from the Shopify App Store (or our WordPress/WooCommerce plugin) and authorize read-only access to your products, collections, orders, and pages. We use official OAuth scopes; Liliwi can never see or edit your customers' payment data.

One-click OAuth install

Read-only scopes by default

Revoke access anytime

Authorize Liliwi

shopify.com/admin/oauth

Connected

2

Syncing

We index your catalog

Liliwi pulls your products, variants, collections, FAQ pages, and policies, then chunks and embeds every document into your private vector database. A typical 1,000-product store finishes syncing in under 3 minutes. New products are picked up automatically via webhooks.

Automatic webhook re-sync on product changes

Markdown, HTML & PDF knowledge ingestion

Per-tenant isolated vector collections

Sync progress
96%

1,247 products embedded

38 collections indexed

12 FAQ pages chunked

Building vector collection…

3

Training

Tune & go live

Set your agent's tone, brand voice, and objectives in plain English. Choose which actions it can take (answer only, recommend products, add to cart, capture leads). Paste the widget snippet — or flip the toggle in our Shopify app — and the agent is live on your storefront.

No-code system prompt editor

Granular action permissions

One-click widget deployment

Agent configuration

Live

Tone of voice

Friendly, concise, on-brand

Allowed actions

Answer

Recommend

AddToCart

Lead

UNDER THE HOOD

Built on a secure RAG pipeline

Every customer message travels through a retrieval-augmented generation pipeline designed for accuracy, speed, and data isolation. Here’s precisely what happens: no black boxes.

1. User asks

"Do you have this in size M?"

2. Embed query

Query → vector

3. Qdrant retrieval

Top-K relevant chunks

4. LLM inference

Together.ai / MiniMax

5. Action + reply

addToCart() + answer

Median end-to-end latency: ~650–800ms · All steps run in your isolated tenant

Context-Aware Search

When a shopper asks a question, Liliwi doesn't just look for exact word matches. It breaks down shopping intent, budget constraints, and product traits to retrieve the most relevant items across your catalog in milliseconds. This dramatically reduces hallucinations and keeps answers sourced from data you control.

Why it matters

Answers are traceable to specific products or docs. You can update knowledge and see results instantly — no retraining required.

Real-Time Data Indexing

Your embedded catalog lives in an open-source, high-performance vector database. Your product catalogs, stock availability, policy documents, and variant attributes are structured and stored in high-performance memory. Updates sync automatically so your assistant never uses outdated details.

Data privacy

Per-tenant collections, no cross-customer retrieval, and delete your data on request, instantly.

Multi-Model Intelligence

Liliwi automatically matches every customer question to the ideal processing power in real time. It uses lightning-fast responses for routine greetings and chitchat, and deploys deep reasoning engines instantly for complex product comparisons and shopping guidance.

Reliability

Automatic failover between providers, 99.9% inference uptime, and sub-second median latency.

LIVE E-COMMERCE ACTIONS

The assistant doesn't just talk; it acts

Liliwi agents can perform real actions on your storefront through a tightly scoped function-calling layer. When a customer says “add the black one in medium,” the agent calls addToCart() — and the item lands in their cart, instantly.

Every action is explicitly whitelisted by you. The agent can never perform an action you haven’t approved, and every call is logged with a full audit trail.

Direct Cart & Checkout Automation

Adds items to cart, applies relevant discount codes, and creates checkout links instantly

Instant Product & Inventory Search

Fetches live product listings with up-to-the-minute pricing and stock status.

Automated Lead Capture

Collects contact info and pushes leads directly to your CRM or email platform.

Seamless Support Escalation

Instantly creates support tickets and routes complex queries to your human team.

Function-calling layer
every action whitelisted & logged
Whitelisted
Cart & checkout automation
Audit trail

Simplified for illustration. Every tool call is validated, scoped, and logged.

SECURITY & TRUST

Designed for merchants and security teams

We know B2B buyers need more than a slick demo. Here’s the technical substance behind the trust — designed for CTOs, security auditors, and compliance teams.

Per-tenant data isolation

Every store gets dedicated vector storage and encrypted DBs. Your product data, embeddings, and chat logs are logically and physically separated from other tenants.

Never trained on your data

Your data is used solely to power your agent. We do not use your catalog, conversations, or customer data to train any foundation models, ours or our providers'. This is contractually guaranteed.

End-to-End Encryption

Data is encrypted in transit (TLS 1.3) and at rest (AES-256). Vector embeddings, conversation logs, and credentials are all encrypted with keys managed in a dedicated KMS.

Least-privilege integrations

We request only the OAuth scopes needed. Minimally privileged OAuth integratio, your AI only does what you explicitly authorize.

Complete Audit Logging

Every retrieval, generation, and function call is logged with timestamps, model IDs, and citations. Fully exportable for compliance reviews.

Deploy & convert

Built to align with SOC 2 Type II and GDPR. Supports custom DPAs, automated data deletion, and dedicated regional data residency.

Infrastructure & compliance partners

QUESTIONS

Technical FAQ

The questions your engineering and security teams will ask.

Where is my data stored, and who can access it?
Your data is stored in isolated Qdrant collections and encrypted databases hosted on our infrastructure provider. Only your tenant’s services can access your data; there is no shared collection, and our staff access is governed by strict role-based controls and audit logging.
No. The agent can add items to a cart (addToCart()), but checkout and payment always happen through your store’s native, secure checkout flow. Liliwi never touches payment data and does not have write access to payment scopes.
The agent is configured to never hallucinate. If retrieval confidence is low or no relevant context is found, it will honestly say it doesn’t know and offer to escalate to a human via a direct contact method. You control the confidence threshold and fallback behavior.
We enforce a strict data-grounding pipeline. Liliwi is restricted from generating information outside your verified store data. If a product is out of stock or missing from the catalog, the assistant directly communicates this rather than guessing.
We apply multiple layers of defense: input sanitization, system prompt hardening, tool-call validation (every action is whitelisted and schema-validated), and output filtering. The agent cannot execute actions outside your configured permissions, regardless of what a user types.