Skip to content

Managed knowledge infrastructure

One layer over every system you already have.

Your records and your documentation, joined into context an AI can actually use — without replacing a system or building any of it yourself.

Book a demoStart free
  • free to start
  • no card
  • your first answer today

The question

Can we turn on SSO for our team?

One record, one document, two different systems. Neither contains the answer.

[1]Billing · record

plan: Team Annual seats: 240 sso_entitlement: false renews: 2027-03-01

account-4471
[2]Help centre

SAML SSO is available on Business and Enterprise plans. Team plans use email and password, with optional two-factor.

Single sign-on setup

The fact the graph holds

Team Annualexcludes →SAML SSO

confidence 0.93

The answer

Not on your current plan — account 4471 is on Team Annual1, and SAML SSO starts at Business2. On Team you can turn on two-factor instead, or move to Business to enable SAML2.

Your CRM does not know how SSO works. Your help centre does not know what this account pays for. The answer needs one fact from each, and it exists in neither.

The problem

Your systems are each right. The gaps between them aren't.

Nobody has a knowledge problem because nobody wrote anything down. They wrote it all down — in six different applications, in two different shapes.

  • The facts and the explanations live in different apps

    Who the customer is, what they bought, what they are entitled to — records, in your CRM and your billing system. How any of it actually works — prose, in your docs, tickets and courses. A real question needs one of each, and no single system holds both.

  • The same fact lives in five places and nothing says which is right

    Pricing is in the billing system, on the pricing page, in a sales deck and in two help articles. Roadmaps ship daily now, so those copies drift apart faster than anyone can reconcile them by hand.

  • Nothing connects them at the moment someone asks

    Distribution is not the problem — each of those systems is purpose-built and correct for its own domain. The problem is that when a customer actually asks, there is no unified view, so a person assembles the answer by hand or gets it wrong.

What changes

Three things start working.

  • Nobody assembles the answer by hand

    Today someone opens the CRM, then the docs, then a ticket, and stitches the three together. That is the job that disappears — and the answer stops depending on whoever happened to know where to look.

  • You can put it in front of a customer

    Because every claim links to the document behind it, an answer becomes something you can ship rather than something you have to go and double-check. That is the whole difference between an internal experiment and a feature.

  • You find out what you haven't written

    It says when it does not know. Those questions accumulate, in the exact words real people used, and that list is a content roadmap nobody had to commission a survey to get.

Why it works

Trustworthy. In your control. Built to last.

Three claims, each one something you can check rather than something you have to believe.

How it finds things

One search isn't enough, because the question rarely matches the document.

So GRaaS runs three at once and merges what they find.

  • always runs

    What you literally asked

    The straightforward search. It runs on every question at full strength, so you can never end up worse off than an ordinary search engine.

  • follows the connections

    What you meant

    Your question says renewal; your content says uplift, true-up, escalator. This pass follows those connections and reaches documents your words would never have matched.

  • ranked by agreement

    What backs it up

    Once it knows which facts are relevant, it goes and finds every document that states them, ordered by how many agree. This is the pass that finds support for a point rather than text resembling a phrase.

Nothing here can make your search worse. The plain search always runs at full strength. The other two only ever add to it — if they find nothing, you still get an ordinary good result.

How the search actually works

Ask it something hard

Watch it show its work.

Not a black box — every answer arrives with the passages behind it and how many sources agree.

Does this account's plan include the instructor-led lab?

  1. direct lane142 ms

    9 passages, hybrid dense + keyword

  2. query entities96 ms

    resolved: Professional plan, Certification track, Lab

  3. expanded lane318 ms

    2-hop traversal, 14 facts, 5 new documents

  4. evidence lane201 ms

    6 sources across CRM, billing and catalogue

  5. rank fusion3 ms

    4 above threshold, from 3 systems

  6. answer synthesis890 ms

    1 answer, 3 citations, 0 dropped markers

Answer

1650 ms total

The account is on the Professional plan [1], which includes the certification track but not the instructor-led lab inside it [2] — that lab is a Premium add-on, billed separately [3].

  • [1]Salesforce — Account record1 source
  • [2]Plan Entitlements Matrix3 sources
  • [3]Course Catalogue 20262 sources

Why you can trust it

Every answer shows its work.

Not one source — all of them, and which system each came from. When a wiki page, a support ticket and a course all say the same thing, you see three, and you can open any of them.

  • You can check it in seconds — which is the difference between an answer you act on and an answer you go away and verify.
  • You can see when something is thin. One source behind a claim looks different from five, and the interface tells you which you are looking at.
  • You can see which system each source came from — the billing record and the help article side by side. We do not quietly pick a winner between them, and we do not yet flag the disagreement for you either: that call is yours, with the evidence in front of you.
Onboarding completion—reduces→Ninety-day churn3 sources

Asserted by

  • Q3 Retention Review

    notion

    0.95 confidence

  • Support Themes 2026

    zendesk

    0.88 confidence

  • Activation Playbook v4

    lms_content

    0.81 confidence

The part nobody else does

It tells you when it doesn't know.

Which is the only reason you can believe it when it does.

Most AI search will answer anything you put to it. That is a design choice, and it is the wrong one for a system whose answers you intend to show someone. GRaaS is built to say plainly that your content does not cover a question — and we test for it, using questions we know the content cannot answer, scored so that answering them counts as a failure.

“Answer only from the numbered sources, and say plainly that they do not contain the answer rather than filling the gap from the model's own knowledge.”
from the instruction every answer is generated under

Control

Your team decides what the system understands.

Most systems make you name your categories on day one, before you have seen what is actually in your content — and then make changing your mind so expensive that nobody ever does.

  1. It learns as it reads

    Every new kind of thing it finds in your content gets recorded. That list is an honest picture of what you actually have, rather than what you assumed when you set it up.

    always on, never edited

  2. Your team curates a version

    Rename things. Merge the six words your teams use for one concept. Drop the ones that turned out to be noise. Nothing you do here touches your content until you say so.

    a draft, not a live edit

  3. See the cost before you pay it

    Before anything changes, you are told exactly how much of your content each change would affect. Not a warning dialog — a number, per change.

    an impact report, not a confirmation

  4. Change your mind freely

    Nothing is ever deleted, so any change can be undone. And when your content moves on past your vocabulary, you are told it has gone stale — before your answers do.

    the reason people actually curate

How the vocabulary works

What you learn

The questions are worth as much as the answers.

Every question anyone asks is kept. Most teams find this is the part they did not know they needed.

  • What people actually ask

    Every question is kept, in the words it was asked in. The gap between what you thought your users wanted and what they type is usually a surprise.

  • What your content misses

    Questions it could not answer, listed. That is your content roadmap, and it costs nothing to collect.

  • Why an answer was what it was

    Open any past question and see which documents were found, how, and what the answer was built from. Most bad answers turn out to be explainable.

  • Whether it is getting better

    How much of your content is connected up, how well it is holding together, and how much of it bridges more than one of your systems — the number that tells you the join is real.

Connecting it

Two calls. No migration, no crawler, nothing of yours to connect.

You send us content. That is the whole integration — which also makes the security conversation short: we hold nothing you did not hand us, and we cannot reach anything you did not.

Most teams wire this into whatever already fires when their content changes. Sending the same document twice replaces it rather than duplicating it, so retries and backfills are safe by default.

For your engineers
POST /v1/ingest
curl -X POST https://api.graas.ai/v1/ingest \
  -H "x-api-key: $GRAAS_KEY" \
  -H "content-type: application/json" \
  -d '{
    "sourceSlug": "billing",
    "externalId": "account-4471",
    "title": "Account 4471 — Team Annual",
    "textContent": "Plan: Team Annual. Seats: 240.
                    SSO entitlement: false. Renews: 2027-03-01."
  }'

# 202 Accepted
# { "documentId": "...", "externalId": "account-4471",
#   "status": "extracting", "chunksCreated": 1,
#   "statusUrl": "/v1/ingest/.../status" }
POST /v1/search
curl -X POST https://api.graas.ai/v1/search \
  -H "x-api-key: $GRAAS_KEY" \
  -H "content-type: application/json" \
  -d '{ "query": "does onboarding reduce churn?" }'

What we measured, and what we haven't.

Graph expansion contributes to 201 of 231 question runs on our evaluation corpus. On the question sets we can currently score, it has not been shown to answer a question that ordinary search missed.

We are not going to dress that up. What it demonstrably adds is the cross-document connection, the evidence trail behind every answer, and a vocabulary you can govern. If you are buying a percentage point of accuracy, we cannot yet prove you should buy it here.

Read the full methodology

Reliability

Nothing you send us goes missing.

An earlier version of our pipeline silently dropped four percent of a corpus — accepted the documents, returned success, lost them. No error anywhere. We found it, fixed it, and published the write-up.

Documents lost
9 in 205improved tozero

Across the same corpus, after moving orchestration onto durable steps with per-step retries.

Extraction p90
~103 simproved to5.2 s

Per document, sharded and run concurrently under a per-organisation ceiling.

Throughput
3.0 / minimproved to17.1 / min

Documents fully processed per minute, same corpus, same hardware shape.

Worst pipeline step
348% of budgetimproved to4%

As a share of the load balancer timeout. Above 100% the request dies before it finishes.

What broke, and how it was fixed

If your customers will be behind it

Built for putting your customers behind it.

The first thing anyone will ask is how your customers are kept apart. They are kept apart by a wall, not a filter.

  • A wall, not a filter

    Each of your customers gets their own space, and there is no query that reaches across. Not a field everyone has to remember to filter on — an actual boundary.

  • Keys you control

    Issue as many as you need, each limited to reading or writing. Rotate or revoke them yourself, without asking us.

  • Your identity system

    Single sign-on, directory sync, and removing someone ends their access immediately rather than whenever their session happens to expire.

What a security review will find

Find out in an afternoon.

Free to start, no card, no call. Point it at a slice of real content and ask it the question your current search gets wrong.

  • no card required
  • first workspace is free
  • your content stays yours
Talk to usStart free