Tavus knowledge base limits vs putting the facts in a Sume script

A Tavus knowledge base is for a live agent that must answer questions from your documents. It supports English documents only, takes 5-10 minutes to process, and caps website crawling at 10,000 crawl documents per account. A Sume avatar video has no retrieval step: the facts the avatar says are the words in the script you submit.

If your content is a fixed pitch, you do not need a knowledge base. If viewers ask unplanned questions, you do.

Tavus knowledge base limits

Tavus lists supported file types as .pdf, .txt, .docx, .doc, .png, .jpg, .pptx, .csv and .xlsx. Processing can take 5-10 minutes depending on document size. For crawls it lists a maximum of 10,000 crawl documents per account, 5 concurrent crawls, a 1-hour cooldown between recrawls of the same document, and max_pages of 1-100 per crawl. The page states no size limit for documents (Tavus docs: Knowledge Base).

The scripted alternative

On Sume you submit the text. Nothing is retrieved or improvised, so a price, a date or a policy appears in the clip exactly as you wrote it. The cost is that the clip is only as current as the script: if a price changes, you render again (Generate avatar video).

For a personalised batch, pull the facts from your own database at the moment you build each script, then submit one job per recipient. Hold the source record next to the job id so you can prove what the clip said.

• Keep numbers that change out of long-lived clips, or put them in the page around the video.
• There is no retrieval layer, so the English-only knowledge base limit does not apply to a script; check the avatar docs for the languages your voice supports.
• A script over 60 seconds of estimated duration must be split into several jobs.

Decision rule

Ask whether you can list every question the viewer will have. If yes, script them and render. If no, a live agent with a knowledge base is the right tool, and the Tavus limits above are the ones to plan for.

Planning around the 5-10 minutes

A knowledge base is not instant, so a flow that creates one at signup needs a waiting state or a fallback. Process documents ahead of time and attach them when the conversation is created.

For crawls, remember the cooldown of one hour before the same document can be recrawled, so a page that changes hourly will lag. In a scripted flow there is no such lag: the script is built when you submit.

Keeping scripted facts correct

Scripts go stale in a way retrieval does not. Give each clip an owner and an expiry date in your own records, and re-render when a fact changes. Store the script text with the job id so you can find every clip that contains an old number.

For facts that change often, keep them out of the video and put them in text beside it, where an update is a page edit rather than a new render.