Knowledge

The knowledge base that actually answers

Name: Zubby AI
Price: 49 USD
Availability: InStock

Ground every conversation in your real catalog, policies, and help-center docs. Zubby retrieves the right chunks, reranks them for relevance, and cites them in the response — so shoppers trust the answer and you stay out of legal trouble.

Start free trial See live demo

zubby · live●●●

Answer engine summary

What is an AI knowledge base for ecommerce?

An AI knowledge base is a structured retrieval layer that lets an AI agent answer customer questions using your store's real content — product descriptions, shipping policies, return windows, sizing charts, and FAQs — instead of guessing from training data. Zubby's knowledge base chunks and embeds every source into a 1536-dimension vector space, retrieves the closest matches with pgvector + HNSW, reranks them with a cross-encoder, and generates an answer constrained to cite those sources. The result is grounded, auditable answers in any of 25 languages, with a feedback loop that flags coverage gaps before they become customer complaints.

Retrieval-Augmented Generation
Combine vector search with LLM generation so answers reflect your real content, not the model's training data.
Cross-encoder reranker
A second-stage ranker reorders retrieved chunks for answer-fit, not just embedding similarity.
Citation enforcement
Every response carries source pointers; low-confidence answers escalate to a human instead of hallucinating.
Auto-promote backlog
Questions with weak matches are surfaced as 'missing knowledge' so coverage improves continuously.

How it works

Ingest. Retrieve. Cite.

Three stages between a shopper's question and an answer they can trust.

Ingest

Connect your store and upload your help center. Catalog syncs via Bulk GraphQL or WooCommerce REST; docs chunk and embed in the background.

webhook + doc-embedding queue

Retrieve & rerank

When a shopper asks, we run a top-K vector search across your namespace, then re-rank the candidates with a cross-encoder for accuracy.

pgvector HNSW + reranker

Answer with citations

The agent generates an on-brand reply constrained to cite the retrieved chunks. Confidence below the threshold triggers a human handoff.

tool-grounded generation

Under the hood

Built for grounded answers, not creative writing

A retrieval pipeline tuned for ecommerce edge cases — out-of-stock variants, multi-language stores, ambiguous returns windows.

pgvector + HNSW retrieval

Semantic search over your product catalog and uploaded docs using cosine similarity on 1536-dim embeddings, indexed for sub-100ms lookups.

Cross-encoder rerank

A second-stage reranker re-orders the top-K results to maximize answer relevance — not just embedding similarity, but real semantic fit.

Inline citations

Every generated answer points back to the source chunk. Shoppers can verify; you can audit. The merchant dashboard shows source similarity scores per turn.

Document ingest

Upload PDFs, DOCX, Markdown, or paste raw text. We chunk, embed, and index automatically. Re-uploads version cleanly without breaking citations.

Auto-promote backlog

Shopper questions that score low on retrieval are auto-queued in a 'missing knowledge' panel. Add the doc, retrain in one click.

Cross-language matching

Multilingual embeddings let an English doc surface for a Spanish-speaking shopper — and vice versa — without manual translation duplication.

Tenant isolation

Every embedding row is scoped to your store_id. No cross-store retrieval, no shared training data, no leakage between merchants.

Coverage analytics

See which doc earns the most citations, which questions go unanswered, and where to invest your next hour of content work.

Inside the dashboard

See every source the agent cited

Coverage analytics, citation trails, and a missing-knowledge backlog — together in one panel.

dashboard / knowledge / coverage

live

Top cited sources (last 30 days)

94% coverage

Shipping & returns policy1842 citesfit 0.92
Sizing chart — Apparel1109 citesfit 0.88
Care guide — Leather612 citesfit 0.85
Warranty & repair FAQ487 citesfit 0.81

Missing knowledge backlog

Do you ship to Brazil?
Can I add monogramming after checkout?
What's the return window on sale items?

Auto-detected from low-similarity queries this week.

Where it runs

Ships on every ecommerce stack worth shipping on

One control plane, three deploy targets. The widget, recovery engine, and AI agent run identically across platforms.

Shopify

Supported

Native Shopify app — Theme App Extension widget, Bulk GraphQL catalog sync.

Open guide

WooCommerce

Supported

WordPress plugin signs into the same SaaS — full catalog, cart, and order events.

Open guide

Hosted Widget

Supported

Drop-in JavaScript widget for headless stacks, custom stores, and BigCommerce.

Open guide

FAQ

Frequently asked questions

Where do the answers actually come from?

Every answer is grounded in your content: product descriptions synced from Shopify or WooCommerce, plus the policy docs, FAQ articles, and PDFs you upload. The agent retrieves the most relevant chunks, then a reranker re-orders them. Generation is constrained to cite those chunks — if nothing matches the question, the agent says so rather than guessing.

What embedding model does Zubby use?

Embeddings are pinned to a 1536-dimension space (default: OpenAI text-embedding-3-small or an equivalent Azure deployment). Postgres + pgvector with HNSW indexes serves cosine-similarity retrieval. The dimension is platform-level, so all stores ride the same index strategy.

Can I see which sources the agent cited?

Yes. Every assistant message carries an attached list of source chunks. In the merchant dashboard, conversations show a 'citations' panel: which doc, which paragraph, similarity score. Shoppers can also see inline source pills in the widget if you enable them.

How fresh is the knowledge base?

Shopify and WooCommerce catalog changes flow through webhooks into the embedding worker within minutes. Document uploads embed on demand. The auto-promote queue surfaces shopper questions that returned weak matches so you can add coverage rather than wait for complaints.

Does the knowledge base work in non-English stores?

Yes. The retrieval layer is language-agnostic — it stores multilingual content and matches based on semantic similarity, not keywords. Detection is automatic per conversation, and you can upload localized doc variants which the reranker promotes when the shopper's language matches.

How big can my knowledge base get?

Plans cover hundreds of thousands of chunks comfortably. Postgres + pgvector with HNSW scales horizontally, and you can shard across tenants because every row is scoped to a store_id. If you need bigger, talk to us — the underlying primitives are the same.

What stops the agent from hallucinating shipping or returns rules?

Two layers. First, the system prompt forbids the agent from answering policy questions without a citation. Second, guardrails block the response if the model tries to claim something not in the retrieved chunks. The fallback is a 'I don't have that on file — let me get a human' handoff.

Keep exploring

Pairs well with

AI Sales Agent

The conversational layer the knowledge base feeds — discovery, recommendations, objection handling, add-to-cart.

Multi-language

25 languages auto-detected. Your knowledge base matches across languages without duplicated docs.

Security & Privacy

Tenant isolation, encrypted credentials, GDPR pipelines — your knowledge base inherits all of it.

Grounded answers

Stop hallucinating. Start citing.

Connect your store, upload your help center, watch the agent answer like it works there. Free 14-day trial. Flat, predictable pricing.

Start free trial Tour the dashboard

What is an AI knowledge base for ecommerce?

Retrieval-Augmented Generation

Combine vector search with LLM generation so answers reflect your real content, not the model's training data.

Cross-encoder reranker

A second-stage ranker reorders retrieved chunks for answer-fit, not just embedding similarity.

Citation enforcement

Every response carries source pointers; low-confidence answers escalate to a human instead of hallucinating.

Auto-promote backlog

Questions with weak matches are surfaced as 'missing knowledge' so coverage improves continuously.