Point it at your website. Your agent knows your business by the time you finish your coffee.

Four ways to feed the knowledge base

1. Crawl a website — give us URLs; we crawl them (with SSRF protection, depth caps, rate limits, DNS TTL caching — we respect your server).
2. Upload documents — PDFs, DOCX, Markdown. We extract text, chunk it, and index it the same way we index web content.
3. Free-text paragraphs — paste product descriptions, pricing explanations, service details — anything you want the agent to cite.
4. Q&A pairs — write exact question + answer pairs for the things you want the agent to nail every time (pricing questions, hours, service areas, insurance).

All four sources feed one knowledge base per deployment. You can mix and match.

Hybrid retrieval, not single-vector search

Most chatbot platforms use plain vector search: embed the query, find the nearest document chunks by cosine similarity, pass them to the LLM. That misses a lot — vector search is good at semantic match but bad at exact-term match (codes, prices, product SKUs, names).

Wengrow uses hybrid retrieval: we run both a Voyage-embedded vector search AND a BM25 lexical search, then re-rank the combined results using Reciprocal Rank Fusion. The practical effect is that asking "what's the price of the Pro plan" surfaces the exact pricing sentence from your site, not a vaguely related paragraph about growth plans in general.

This is the retrieval stack Anthropic recommends for production RAG. It's more accurate than any single method alone, and it's what every tier gets out of the box.

Per-deployment knowledge namespaces

Each deployment gets its own isolated knowledge base. If you run a multi-brand business, a multi-language site, or an agency with multiple clients, each deployment's agent only sees its own namespace. No cross-contamination. No confused answers.

This is how white-label agencies run Wengrow: 125 client deployments, 125 isolated knowledge bases, one admin.

See how resellers use Wengrow →

Scheduled auto-refresh (Growth and above)

Your website changes — you launch a new product page, update pricing, publish a case study. Scheduled auto-refresh re-crawls your site on a cadence you set (daily, weekly, or on manual trigger). The knowledge base stays in sync without anyone remembering to do it.

Audit: see which chunks the bot used

Every conversation in the lead management panel shows which knowledge base chunks the agent cited to answer. This is how you tune the knowledge base over time: if the bot keeps citing the wrong chunk for a question, you know exactly which passage to improve.

See the lead management panel →