Documentation

Knowledge base

The knowledge base is everything the agent should know about your business but is too business-specific to live in the scenario prompt. Menu items, price lists, service descriptions, opening hours, FAQ, refund policy, fitting instructions — anything that lets the agent answer "do you have…?", "how much…?", "can I…?" correctly instead of guessing.

The platform indexes the knowledge base and the agent retrieves from it in real time during the call. The retrieval is grounded — the agent quotes from your data, it does not invent facts. When the knowledge base is silent on a question, the agent says so and either escalates or captures a lead, rather than making something up.

What you can add

Three input formats:

  • Your website — paste a URL and the platform crawls the public pages, extracts readable text, and indexes it. We crawl in one go for the first index and incrementally on a schedule you set after that.
  • PDF documents — menus, datasheets, brochures, price lists, manuals. Drop the file into the Knowledge tab. We extract text (no OCR — scanned PDFs are flagged so you know to provide a text version). Up to 20 MB per file, with a soft cap of 50 files per project.
  • Direct text — paste content directly into the Knowledge tab. Useful for short FAQ entries that do not live elsewhere.

How indexing works (in plain language)

When you add content, the platform splits it into chunks at heading boundaries, computes an embedding for each chunk, and stores them in a vector database. At call time, when the agent looks up "what time do you close on Sunday?", we embed the question, find the most relevant chunks, and feed them back to the agent. Everything happens inside a tight latency budget so the agent answers fast.

You do not see this machinery. You see: "Knowledge added — searchable in a few seconds."

Refreshing content

For websites, you can:

  • Re-crawl on demand — click the Re-index button when you change prices or services
  • Schedule auto-refresh — daily, weekly, monthly; the platform fetches what changed
  • Disable auto-refresh — keep content frozen if your business is stable

For uploaded PDFs and direct text, content stays as you uploaded it until you replace it.

Boundaries — what knowledge is for vs not for

Use the knowledge base for facts the agent should know: prices, hours, menu, locations, policies, FAQ. The agent reads these and quotes from them.

Do not use the knowledge base for rules — the rules go in the scenario goal and prompt. "Do not book parties of more than 8" is a rule; put it in the goal's constraints. "Our largest available booth seats 8 people" is a fact; put it in the knowledge base.

The distinction matters because the agent treats knowledge as quotable but treats rules as binding. Mixing them weakens both.

Multilingual knowledge

If your business serves callers in multiple languages, you can upload knowledge in each language. The agent retrieves in the caller's detected language. If a fact only exists in one language, the agent translates on the fly — accurately enough for most cases, but you get better results with first-class content per language.

The default omnilingual set the platform supports: English, Spanish, Russian, French, Italian, German, Portuguese, plus 20 more. Your agent's allowed-languages list is set per project.

Privacy and storage

Knowledge base content is yours. The platform stores it inside the EU (Hetzner Germany and Finland) under the same tenant isolation as the rest of your project data. Embeddings are computed via OpenAI; only the minimum needed text leaves the platform for embedding, and we never use customer data to train shared AI models. See Security.

Capacity and limits

Today's limits, per project:

  • Up to 50 files in the knowledge base
  • Up to 20 MB per file
  • Up to 1000 indexed pages from a single site crawl (raise on request)
  • Storage is included in your wallet — see Billing — under "crawler hours" for the indexing job and standard object-storage for the cached files.

If you are bumping these caps, contact support and we will lift them per-project. They are conservative defaults, not architectural ceilings.