Answers you can trace

Senso is a context layer: one place for the facts your organization stands behind, which your people and your agents draw on. Every answer Senso writes comes from your own documents, and comes back with the passages it used.

Answering is half of it. Senso also checks what is said about you against those documents, and turns whatever is wrong or missing into a gap someone can close.

The mistake this catches is rarely an invented fact. It is a sentence that sounds right with one detail wrong: a 60-day refund window when yours is 30, or a price without the condition that limits it. Nobody notices until a customer acts on it.

senso search "How long do customers have to ask for a refund?" --output json

{
  "answer": "Customers have **30 days from the date of purchase** to request a full refund.",
  "results": [
    {
      "content_id": "6c1f2a9e-3b4d-4e8f-9a7b-1c2d3e4f5a6b",
      "version_id": "7a6b5c4d-3e2f-4a1b-9c8d-7e6f5a4b3c2d",
      "title": "Refund policy",
      "chunk_text": "Customers can request a full refund within 30 days of purchase.",
      "score": 0.87
    }
  ],
  "total_results": 1
}

Each result names the document (content_id) and the version of it (version_id) that the answer came from. Correct the document and the next answer changes with it, under a new version_id. The old text is no longer used.

Senso processes what you add and makes it searchable, with nothing for you to set up or run. Add a document with senso kb create-raw or senso kb upload, and it can be searched within seconds.

An eval splits a text into the claims it makes and checks each one against your knowledge base. Every claim gets a verdict, with the passage quoted as evidence:

senso evals text --evaluator kb_accuracy --wait \
  --text "Customers can get a full refund within 60 days of purchase. Refunds go back to the original payment method. Refunds take 5 business days to arrive."

verified means a passage states it. conflict means a passage says otherwise. unsupported means nothing on file says either way: not a falsehood, but something you cannot back up. This run scores 1 of 3 claims verified.

senso evals text checks any text, including an AI model’s answer about you. senso evals content checks your saved and generated content. A second evaluator, brand_alignment, checks text against your brand kit’s writing rules instead.

A gap is something your knowledge base should say and does not, or says differently. Senso keeps one list of them, and each one points at what to fix:

Once you fix the source, record it with senso gaps answer, or close a gap that does not matter with senso gaps dismiss. A question that found nothing closes on its own, the next time a search answers it.

A question searched once is held back until it is asked again, so a single test search does not fill the list. Keep your own tests out entirely with --no-gap-signals, or set SENSO_GAP_SIGNALS=off.

The verification loop puts these together for what AI models say about you. Senso asks AI models the questions people in your market ask, shows where a competitor is named or cited and you are not, and checks each answer against your knowledge base. You fix the source, publish content written from it, and watch whether the next answers change.

Verification Loop walks through it.

The same knowledge base can serve a team. Every teammate’s agent reads from it and adds to it, permissions decide who sees which folders, and the gap list shows the questions nobody has answered yet, so the person who knows can write the answer down once.

Shared Context walks through it.

senso search context returns the passages without an answer, so you can write the answer with your own model. Either way, run your pipeline’s answers through senso evals text to see which claims your documents back up, contradict or cannot support. That turns “our answers seem fine” into a number.