AI that reads your real data, not a sample
AI that works from your orders, catalog and customers, with your own model keys and rules you control.
Generic commerce AI answers from generic text. Kynetra works from your live catalog, order history and customer records, so its search results, recommendations and fraud flags are specific to your business.
All AI capabilities β 12 features
Why it matters that the AI works from your data
- Search results come from your own indexed catalog, so a query like "MTB under βΉ15k" can return real products with their real prices.
- Fraud rules score against your own order history, with weights and actions you set.
- The admin copilot answers from your live records by calling the admin's own read tools, as the signed-in user.
- Segments run directly on your live order and customer data, with no sync to a separate customer data platform before a campaign can go out.
Kynetra Search
NewA search index kept in your own database and updated whenever a product is saved, with fuzzy matching, synonyms, redirects and ranking you can tune.
Kynetra Search reads from a flattened product search table in your own database instead of joining raw catalog tables on every query. The database keeps that table current automatically, so saving a product updates its search row in the same step and no scheduled rebuild is needed. Queries are cleaned up, matched with fuzzy text matching, expanded with your synonyms and ranked by weighted signals: exact name matches, prefixes, matches on every word, brand, category, whether the product can be bought, and whether it has an image. In Search Studio, merchandisers pin, boost or bury products, set redirects and zero-result fallbacks, and adjust the ranking weights per business.
- Search table updated automatically when a product is saved, with no scheduled rebuild
- Fuzzy text matching so small typing errors still find the product
- Synonyms, redirects and pinned, boosted or buried products managed in Search Studio
- Zero-result fallbacks: choose the products to show for a query that finds nothing
- Ranking weights for exact, prefix, all-word, meaning-based, in-stock and image signals, set per business
- Last good results are served if the database errors, and identical queries at the same moment share one lookup
Kynetra Prime β search auto-training
NewPrime watches customer search activity and proposes relevance changes. Risky changes wait in a review queue until you approve them.
Kynetra Prime watches search activity such as queries that return nothing, clicks and add-to-carts from search, and proposes improvements: synonyms, redirects, query rewrites, zero-result recoveries and boosts. A discovery run also asks a language model to look at recent search events and suggest more. The engine applies only low-risk changes by itself. Anything riskier appears in a review queue, and it goes live only when an admin approves it. A kill switch and a percentage rollout control whether the feature runs at all, and a quality dashboard shows zero-result rate, click-through and training history.
- Learns from search events: queries with no results, clicks and add-to-carts from search
- Proposes synonyms, redirects, query rewrites, zero-result recoveries and boosts
- Risky changes wait in a review queue and go live only when an admin approves them
- Low-risk changes are applied automatically, so review the training history
- Kill switch and percentage rollout for the feature flags that gate it
- Quality dashboard with zero-result rate, click-through and training runs
Meaning-based search and related products
CoreVector search on the storefront finds relevant products when keywords match nothing, and powers related-product suggestions.
On the storefront, the shopper's words are turned into an embedding with a Cloudflare Workers AI model and matched to the nearest products in a vector index kept separately for each business. This finds relevant products even when no keyword matches. Repeated queries reuse a cached embedding. The lookup runs against a time limit behind a circuit breaker, and if it is slow or unavailable, shoppers get the keyword results instead. The same index powers a more-like-this suggestion for a product. Meaning-based fallback can be switched off per business in the search settings. Image search needs a vision and vector service to be configured for the business.
- Embeddings from a Cloudflare Workers AI model, with a separate vector index for each business
- Finds products by meaning when keyword search finds little or nothing
- More-like-this suggestions from the same index
- Query embeddings cached for repeat searches
- Time limit and circuit breaker: shoppers fall back to keyword results if the lookup is down
- Meaning-based fallback switchable per business in search settings
Command Chat β the admin copilot
CoreAsk questions about orders, catalog and analytics in plain language. It reads live data through the admin's own tools, as you, and cannot change anything yet.
Command Chat is the operations copilot inside the admin. It acts as the signed-in user on their business, so it can only reach what that user's role allows. It answers by calling the admin's own read tools, shows each tool call as it happens, and links the records it mentions. In this release it is read-only: if you ask for a change, it says so and points you to the screen where you can make it. Each tool call is recorded with the session, user, procedure, arguments, result status and time.
- Works as the signed-in user, limited to that user's role and business
- Calls the admin's own read tools, and shows each call in the conversation
- Read-only in this release; write actions are not available yet
- Each tool call is logged with session, user, procedure, arguments, result status and time
- Uses your model keys through the same AI gateway as every other AI feature
Theme copilot
CoreDescribe storefront theme changes in plain language and see them applied to the live preview, with undo.
In the theme editor, you describe a change such as a transparent header or bigger, rounder buttons. The copilot sees your current theme, returns a checked set of design-token changes with an explanation, and layers them over the theme in the editor. Because the edit goes through the editor's normal history, you can undo it, preview it and let autosave keep it. It works across several turns, so each request builds on what is already there.
- Reads the current theme, so each request is relative to what is there now
- Returns design-token changes that are checked before they are applied
- Changes flow through the editor's undo, redo, preview and autosave
- Works over several turns in one conversation
- Uses your model keys through the AI gateway
Conversation engine
A flow-based bot engine behind the WhatsApp and Telegram assistants. Language-model intent detection is built but not connected yet.
Conversations run on a session store that remembers where each customer is in a flow. A flow can search the catalog, add items to a cart, create a payment link and flag a conversation for a person to take over. A separate agent that reads a message and returns an intent (browse, add to cart, checkout, payment, complaint, return, negotiate or greeting), the items and quantities mentioned and whether to escalate has been built, but it is not part of the live flow yet, so today's bots follow the flows you design. See the conversational commerce page for the channels that are live.
- Sessions remember the customer's place in the flow
- Flows for catalog search, add to cart and payment links
- A handover flag for a person to take over
- Intent detection agent built, not yet connected to the live bot
- Used by the WhatsApp and Telegram channels
Fraud screening
Rules with weights score each new order and place it in one of four outcomes: allow, review, hold or block.
After an order is created, active fraud rules are checked against it and the weights of the rules that match are added up. A score under 30 is allowed, 30 to 60 goes to review, 60 to 90 is held, and above 90 is blocked; a rule set to block blocks the order on its own. Rules you can use include order velocity, large first orders, brand-new accounts paying with a chosen method, email address patterns, country mismatches and a blocked-payment-methods list. Held and blocked orders are put on hold with the reason, and the admin shows every triggered rule with its weight so a person can decide. The score and the rules that fired are saved with the order.
- Rule types: order velocity, order amount, account age, email patterns, country mismatch and blocked payment methods
- Four outcomes: allow, review, hold and block
- Each rule has a weight and an action you set
- Held and blocked orders go on hold with a reason
- The score and the rules that fired are saved with the order
- Review queue and rule editor in the admin
Segmentation
Build customer segments from nested AND/OR rules on lifetime value, order history, tier, points and location.
A segment is a set of rules joined by AND and OR, and groups of rules can be nested inside each other. Rules run against your live customer and order data when you ask for the segment, with no copy to a separate analytics store. Fields include email, name, phone, city, country, sign-up date, order count, lifetime value, last order date, loyalty points and loyalty tier. Segments can be used in loyalty and marketing journeys.
- Nested AND/OR rule groups
- Fields: order count, lifetime value, last order date, points, tier, city, country and sign-up date
- Rules run against your live customer and order data, with no separate analytics store
- Used by marketing journeys and campaigns in the admin
- Operators include equals, greater or less than, in, contains, starts with and is empty
- A count and sample preview before you save a segment
Marketing agent
Recommends coupons, campaigns, segments, journeys, bundles and price changes, with budget and margin guardrails. You approve each one or let safe ones apply.
The marketing agent looks at your orders, stock and customers and fills an inbox of recommendations. Triggers include ageing inventory, lapsed customers, abandoned carts, festivals, new-customer welcomes, replenishment and weekend flash sales. A recommendation can be a coupon, promotion, campaign, segment, journey, price change, bundle or referral. You can approve, reject, edit or schedule each one. Guardrails cap discounts per day and per month and protect a minimum margin, and the agent can be set to apply recommendations itself within those limits.
- Recommendation kinds: coupon, promotion, campaign, segment, journey, price change, bundle, referral
- Triggers: ageing inventory, lapsed customers, abandoned carts, festivals, replenishment, weekend flash sales
- Approve, reject, edit or schedule each recommendation, singly or in bulk
- Daily and monthly discount budgets and a minimum margin as guardrails
- Optional auto-apply within those limits; approval stays manual unless you turn it on
- Approvals and applies are limited by role permissions
Description enhancer
Generate or improve product descriptions in bulk with Claude, GPT or Gemini, and keep the drafts for review.
Choose products and a model, and the enhancer writes descriptions in the tone you pick, using the product's own details and, if you allow it, the brand's website or Google Shopping snippets. Bulk runs are background jobs with progress and per-item status, so a large batch does not block the page. You can run it in draft mode and review each result before replacing the live description. Each model shows a credit cost in the settings, and requests are recorded in the AI gateway with the model, feature and token use.
- Models: Claude, GPT and Gemini
- Tone choices such as professional, conversational, technical or minimalist
- Draft mode keeps results for review before they replace live descriptions
- Bulk runs as background jobs with per-item status
- Optional sources: the brand's website or Google Shopping snippets
- Every request is recorded in the AI gateway log with model, feature and token use
Your own keys and guardrails
Register your own provider keys, set budgets, and block personal data, risky prompts and expensive calls with rules you choose.
You can add your own provider keys in the admin. They are encrypted at rest with AES-256-GCM and only the last four characters are shown again. Calls use your key when it matches the model's provider. When a model fails, the gateway tries the fallback models you have configured. Guardrail rules run on prompts and responses: block or redact personal data patterns such as email, phone, card number, Aadhaar and PAN, flag prompt-injection patterns, set a maximum cost per call, allow or deny models, block output patterns and rate-limit calls. Spend budgets, set by scope and period, block calls once they are exceeded, and alerts fire as usage approaches the limit. Usage on your own key is billed by the provider; AI usage on the platform's shared key is billed separately from the subscription, as the pricing page explains.
- Provider keys encrypted at rest with AES-256-GCM; only the last four characters are shown again
- Uses your key when it matches the model's provider
- Fallback models if the first choice fails
- Guardrail rules: personal data, prompt injection, cost per call, model allow and deny lists, output patterns and rate limits
- Spend budgets by scope and period that block calls once exceeded, with usage alerts
- Rules can apply to chosen features only, and a test panel tries a prompt against your rules
Search resilience
ArchitectureTime limits, circuit breakers and cached fallbacks keep search answering when a dependency is slow or failing.
Slower steps such as meaning-based lookups and AI calls run against a time limit, so one slow dependency cannot hang a search. A circuit breaker stops calling a dependency that has failed repeatedly, then lets one trial request through to see whether it has recovered. Performance budgets are stored per metric, and a budget breach can switch an expensive step off so search degrades gracefully. If the database errors, the last good result for that search is served from cache. Identical queries arriving together are answered from one lookup. Individual features also sit behind kill switches.
- Time limit on slower search steps
- Circuit breaker that stops calling a failing dependency, then retries with one trial request
- Stored performance budgets that can switch an expensive step off
- Last good results served from cache if the database errors
- Identical simultaneous queries share one lookup
- Kill switches for individual search features