How it works

Start with your documents. Get an answer you can check.

We build your knowledge base from the information your business already has. When you ask a question, it checks what it can use, finds the relevant passages and answers with sources you can open.

Step 1 · Collect

Keep the original evidence

Your documents, email, meeting notes and records from software you already use come in as raw evidence. We keep it exactly as it arrived. The system never searches it directly, so you keep the original to check against anything written from it.

  • Gmail
  • Drive
  • Notion
  • Slack
  • HubSpot
  • QuickBooks
Raw evidencekept as it arrived
Raw evidence stays unchanged and outside search.

Step 2 · Write it up

Turn evidence into documents you own

We turn the evidence into short documents in OKF, the Open Knowledge Format: plain text files you own. A small header records type, status, team or personal access, review date and sources. Every claim links to its source, so you can trace it without a proprietary database.

Original evidence

Knowledge document

type: SOP
status: stable
scope: shared
stale_after: 2027-03-01
sources: [email, 2026-09-03]
11 types, including policy and decision. You own the files.

Step 3 · Check access

Check what the answer is allowed to use

Each document has a status: draft, stable or deprecated, meaning replaced or wrong. Before ranking, the system excludes deprecated documents, those past their review date and anything you lack permission to see. A high search score cannot bring them back into your answer.

  • StatusIs it marked deprecated?
  • Review dateIs it past its review date?
  • PermissionAre you allowed to see it?
  • Repair booking guide
  • Old price listDeprecated, left out
  • Booking confirmation email
  • Winter hours 2025Stale, left out
  • Service call SOP
  • Payroll notesNo permission, left out
Example
Scope: shared with the team or private to one person.

Step 4 · Index

Give each section two ways to be found

Each section gets an embedding, a numeric fingerprint of what it means, and a word index, a lookup list of its words. It becomes searchable only when both are ready. Like a librarian, the system can then look by subject as well as by the words you remember.

Before the visit

What to record

Letting the tech in

A section needs both its embedding and its word index.

Step 5 · Search

Search three ways at the same time

Your question starts three searches at once: exact matches for names, IDs, codes or file names; keywords in the text; and meaning, using the embeddings. Each catches things the others miss, so you can ask in your own words even if you don't remember how something was written.

Fictional example: booking a repair at Miller Heating.

Step 6 · Rank and cut

Keep the passages that fit your question

The lists merge, rewarding passages that rank well in several searches. A second model rereads the best matches against your question. Scores below a fixed cutoff get dropped. The rest fit in rank order up to the answer's space limit. If none pass, it says it doesn't know instead of guessing.

  • Repair booking guide 0.91
  • Booking confirmation email 0.84
  • Service call SOP 0.72
  • Warranty terms 0.41
  • Holiday schedule 0.22

I don't know. No source passed the cutoff.

Example
A reranker is the second model that checks the best matches.

Step 7 · Answer

Follow each claim back to its source

The answer includes a footnote on every claim. Each one links to the knowledge document and the original source it came from. You can open those links to check the answer, just as you'd ask a new hire to show you where they found an instruction.

Record the customer's contact details and the heating or cooling issue.1 Add the agreed visit time to the booking.2

  1. Repair booking guide, 2026-06-08
  2. Booking confirmation email, 2026-06-09
Example
Fictional example. Each numbered footnote opens its sources.

Step 8 · Learn

Put corrections back into the knowledge base

When someone corrects an answer or makes a decision, a small model drafts a new or updated knowledge document from that exchange. It waits for review before joining the knowledge base. Your team can improve the next answer without writing the documentation themselves.

  1. 1Correction or decision
  2. 2Document drafted
  3. 3Review before adding
Back into step 2
A draft waits for review before it joins the knowledge base.

Book a call

Start with the knowledge you already have

We run our own work on this same engine. Your knowledge base can run on hardware you control. Tell us what your team keeps having to look up.