B Blengi docs
/

Build your agent

Knowledge templates

Knowledge templates are pre-structured, fill-in-the-blank scaffolds that get a new agent answering well from its very first source. Open /app/agents/{agent}/sources, click Add source, and pick the Templates tab. Choose a template, replace the [bracketed placeholders] with your own facts, and save — it indexes through the normal pasted-text pipeline. Every template is also downloadable as .md or .txt.

The eight templates

  • 🏢 Company information — name, full address, phone, email, opening hours, how to reach you (directions & parking), registration numbers, and service area. Ships with Global source suggested on, so these facts answer from every page of a conversation, not just when the visitor is on your contact page.
  • 📋 Product register — one heading per product with a fixed field list (colors, sizes, material percentages, features, image and product-page URL). The per-product headings are what let the assistant answer product facts exactly without mixing products up, and the template deliberately contains no prices — prices in documents go stale; the product page stays current. Built from the structure proven in production. On the sizes line, parenthesised text is treated as annotation, not data — write exclusions like Maten: 39-42 / 43-46 (niet in 35-38 of 47-50) freely; only the sizes outside the parentheses count toward the recorded range. Deleting a register also deletes its extracted facts, and each re-index sweeps facts left behind by previously deleted registers, so a replaced register can never keep answering under its old name. Two more line types are captured when present: a price line (Prijs:) and a caveat line (⚠️ Uitsluitingsregel:, Let op:, or Waarschuwing:). Whenever retrieved content or the visitor's own message names a register product, that product's recorded facts — sizes, material, price, and the ⚠️ caveat — are automatically attached to the model's context for the turn, so a decision-tree chunk that merely mentions a product can never leave the model to invent its properties. Caveat lines are the owner's guardrail: "Wol houdt vocht vast — geen vervanging voor een vochtafvoerende sportsok" rides along on every mention.
  • 🧭 Decision guide — teaches the assistant to advise like your best salesperson: when the guide applies, the one question that steers the advice most, explicit decision rules, never-do rules (including "never quote prices from memory" and the medical referral rule), an example conversation, and a decision tree.
  • 📦 Products & services — per item: what it is, who it is for, the options or sizes, and what makes it different.
  • ❓ Frequently asked questions — question-and-answer pairs (price, ordering, cancellation, access, payment, discounts, reaching a human, delivery).
  • 💰 Pricing — explicit price sentences, what is included, the billing period, discounts, and how to get a quote.
  • 📋 Policies — cancellation, refunds, a privacy summary, and insurance/liability.
  • 👥 Team members — per person: name, role, what they help with, and languages spoken.

Each template is authored in English and Dutch; the Templates tab defaults to the agent's language and offers an EN/NL toggle. You only fill in your content once — the embedding model retrieves across languages, so a Dutch knowledge base still answers an English visitor and vice versa.

Best practices baked in

The templates are structured the way retrieval likes best. If you write your own knowledge, follow the same rules and your bot will answer more reliably:

  • Write facts as full sentences, with the subject named. “Our address is 12 High Street, London” retrieves far more reliably than a bare “Address: 12 High Street”. A visitor asks in a sentence; the stored fact should read like one too.
  • One topic per source. Keep company info, pricing, and policies as separate sources. Cleaner sources mean cleaner citations and better ranking.
  • Make each paragraph self-contained. A retrieved chunk is read on its own, so don't let a fact depend on a heading three paragraphs up — repeat the subject in the sentence.
  • Write FAQ entries as real questions. Phrase the question the way a customer would type it; a question retrieves well against a matching question.
  • Spell out the specifics. Prices with currency and period, opening hours per day, addresses in full — no abbreviations a visitor wouldn't type.

Placeholder guard

Placeholders look like [Your address]. If you save a source with brackets still in it, Pitchbar warns you first — otherwise the bot would happily tell visitors “our address is [Your address]”. You can still save (some brackets are intentional), but the warning is there to catch the common mistake of pasting a template and forgetting to fill it in.

Google Docs & Google Sheets templates

Customers who prefer to author in Google can start from a hosted template instead of the in-app editor. The Templates tab shows a Make a copy card for each configured Google template (Company info, FAQ, Policies, Team as Google Docs; Products & Pricing as Google Sheets). Clicking it copies Blengi's template into the customer's own Drive; they replace the example content, then add the file as a source via the existing Google Doc/Sheet flow (Google must be connected in Integrations).

Setup (admin, one-time)

Cards are hidden until their copy URL is configured. Host each Google file (sharing: "anyone with the link can view") and set the matching env var to its .../copy link: 

GOOGLE_TEMPLATE_COMPANY_INFO_URL=
GOOGLE_TEMPLATE_PRODUCT_REGISTER_URL=
GOOGLE_TEMPLATE_DECISION_GUIDE_URL=
GOOGLE_TEMPLATE_FAQ_URL=
GOOGLE_TEMPLATE_POLICIES_URL=
GOOGLE_TEMPLATE_TEAM_URL=
GOOGLE_TEMPLATE_PRODUCTS_URL=
GOOGLE_TEMPLATE_PRICING_URL=