They're in your website, procedures, spec sheets, contracts and product pages. We build assistants that retrieve the right information and answer from it — with the document, the section and the page attached. For the people evaluating you on the outside, and for your team on the inside.
Nobody is short of documentation. They're short of a way to get one specific paragraph out of eleven years of it — in the ninety seconds before the customer is back on the phone, or before the visitor closes the tab.
SharePoint, the shared drive, a folder of supplier PDFs, the ticket history, the product pages, and one person's inbox. The answer exists in exactly one of them and nobody remembers which.
Your search box needs the words that are in the document. People have the words the customer just used. “Can we run it wet?” never matches a heading that reads “Environmental rating, IP69K.”
Inside, the question goes to your most senior and most interrupted employee, whose knowledge leaves when they do. Outside, the visitor doesn't ask anyone — they close the tab, and you never learn what they wanted.
A model's knowledge is fixed when its training ends.2 It has never read your warranty policy or your tolerance tables, so when asked it produces something plausible and wrong. Confident and wrong is the expensive failure.
We find the passages in your material that actually address the question, and the model is only allowed to answer from those. It reads your documents at the moment of asking, every time.
Every answer carries the document, the section and the page it came from. Anyone can verify in one click — the only reason a person will ever trust it enough to use it on a live call, or act on it as a buyer.
The same retrieval system, the same citations, the same refusal to invent. What changes is who's asking, and what happens at the end of the conversation.
One question box over the SOPs, the spec sheets, the contracts and the manuals. Your staff get the answer and the page it came from, instead of walking over to the one person who knows.
Start with: sales engineering, field service, support, operations — whoever owns the specs. The team that gets interrupted most already knows which questions repeat.
People evaluating something don't ask once — they compare, push back and narrow down. The assistant answers from your published material, and when someone signals real intent, a human is told within minutes.
Start with: the pages where evaluation happens — products, specs, comparisons, pricing. That's where the questions are worth money.
A closed-book model answers from memory and hopes. This one looks the answer up in your documents first, answers only from what it found, and shows you the page. Four steps — and the third is a hard constraint, not a suggestion.
Your pages and documents are split into passages and stored so they can be found by meaning rather than by keyword. Text, tables and PDFs, including the scanned ones.
A question comes in. We pull the handful of passages that genuinely address it — matching on meaning, so the question doesn't have to use the document's vocabulary.
Those passages go to the model with strict instructions: answer from this material only. Nothing outside it is available to draw on, so nothing outside it can be invented.
A written answer in plain language, each part linked back to the passage behind it. Where the passages don't cover the question, it says so and points at a human.
Because the knowledge lives outside the model, nothing has to be retrained when your material changes. Correct a procedure this morning, re-index, and this afternoon's answers reflect the correction.2
An answer nobody can check is a rumour. Every claim is tied to the exact paragraph it came from, with a deep link into the live page or the specific page of the PDF — so the reader can verify it, and so you can audit it later.
Every conversation is scored from real signals — thumbs, fallbacks, abandonment, escalations, whether it ended in something useful. Filter to the bad sessions and read exactly what went wrong, instead of guessing whether the thing is working.
The questions your material can't answer aren't just documentation gaps — they're demand. We rank them by how often they're asked and pair each with who asked, so the right person knows what to write and, on a public site, who to call.
This is the first question every serious buyer asks, and it deserves a direct answer rather than a page of assurances. Your documents are indexed into a store you own. They are not used to train anything. Where your policy or your regulator requires it, the whole system runs inside your own infrastructure and nothing leaves it.
In both cases the assistant refuses the question it can't source, says why, and turns the dead end into a next step. That behaviour is the whole product.
A system that guesses at the second question is worse than no system, because your team will believe it once and stop believing it forever. Refusing to answer, naming the missing document, and finding the person who owns it is what makes people willing to use it on a live call.
The assistant had no sourced answer, said so plainly, logged the gap for your content team, and turned the dead end into an introduction — instead of inventing a lead time that costs you the deal three weeks later.
We'd rather you picked the right one than picked us. Here is the honest comparison, including where a cheap off-the-shelf tool beats us outright.
| Build in-house | Off-the-shelf tool | Work with us | |
|---|---|---|---|
| Time to something useful | 8–14 weeks for a median build — then it is never really finished1 | An afternoon | Two to three weeks |
| Up-front cost | $15,000–$40,000 for a basic build. The median is $75,000–$120,0001 | Nothing — though enterprise platforms want $12,000–$50,000 a year on an annual contract4 | A fixed project fee |
| What it costs to run | $150–$1,500 a month in tokens and hosting3, plus 15–25% of the build price every year in maintenance1 | $15 a month to start. $150–$1,500 once people actually use it3 | One monthly fee. Nothing metered, no usage bill. |
| Who you need | 2–4 developers with retrieval and vector database skills, at $200–$300 an hour blended1 | Anyone who can drag a file | Nobody. That's the point. |
| Handles messy PDFs, tables, scans | Eventually, if you fund it | Poorly, and it won't tell you | Yes — it's most of the work |
| Tuned to your questions | Yes, by your team, forever | No. Generic retrieval, generic results | Yes, by us, against real transcripts |
| When it gets an answer wrong | Your team debugs retrieval | You file a ticket and wait | We fix it, and tell you why it happened |
| Best when | This capability is core to your product and you're staffing for it | Your content is tidy, public and low-stakes | Answers must be right, and the documents are a mess |
If your documentation is clean, in one place, and nobody gets hurt by a wrong answer, buy an off-the-shelf tool for a few hundred a month and don't call us. We're worth the money when the material is difficult and the cost of being confidently wrong is real.
Every rollout that fails begins with “let's index everything.” Pick the smallest set of material that causes the most pain, and find out whether this works before you widen it.
Choose the team, or the pages, that get hit most. Inside, that's usually sales engineering, service, or whoever owns the specs. Outside, it's the product and comparison pages where evaluation actually happens. Both already know which questions repeat, which is the fastest tuning signal there is.
Give us the awkward documents, not the tidy ones. A clean wiki proves nothing. Hand over the scanned manuals and the spreadsheet with merged cells — if it works on those, it will work on everything else you own.
Agree what “right” means before we start. We write down twenty real questions and the answers you'd accept. That list is how the build gets judged, and it stops the project ending in vague opinions about whether the AI is any good.
Expect the gap report to sting a little. The list of questions your material can't answer is genuinely useful and mildly embarrassing for everyone. It's often worth more in the first month than the assistant is.
An assistant that isn't maintained goes stale within a quarter — your material moves and the answers stop matching it. So the build and the year that follows are sold together, and quoted together, after one call.
The build fee turns on how much material there is and what state it's in — a tidy documentation set and forty thousand scanned pages are not the same job. Twenty minutes on a call is enough for us to put a fixed number in front of you, and to tell you if the honest answer is that you don't need us.
Not ready to commit to a build? We also run a paid three-week pilot on one content set, credited in full against the build if you go ahead — and yours to keep if you don't.
We take on a small number of builds at a time, because every one is tuned by hand against your material. If that means waiting a few weeks for a slot, we'll say so on the first call.
The fastest way to work out whether this is worth your time is to give us a question your team or your website currently answers badly — and roughly which document holds the answer. That's enough for us to tell you whether this is worth doing.
A first call is twenty minutes. If your material isn't in a state where this would work yet, we'll say so and tell you what would need to change.
Sources. 1 SFAI Labs, RAG Development Costs — a development agency's own published tiers: basic $15,000–$40,000 (3–6 weeks, 1–2 developers); moderate $40,000–$120,000 (8–14 weeks, 2–4 developers); complex $120,000–$250,000; enterprise $250,000–$500,000+. Median project “$75,000–$120,000 with 8–14 weeks of development time.” Maintenance runs “15–25% of initial development costs annually,” on blended US agency rates of $200–$300 an hour. Compliance work adds 30–100%, integrations 20–60%, and a compressed timeline 30–50%. 2 indigo.ai, Retrieval-Augmented Generation — an LLM's knowledge is frozen when pre-training ends; RAG updates the knowledge base without changing model parameters, and citing sources is what makes the output checkable. 3 SpendArk, RAG System Cost — monthly running cost of a live RAG system: $150–$400 at 10,000 queries a month, $600–$1,500 at 150,000, rising to $5,000–$15,000 at a million. LLM inference is roughly 68% of that bill; the vector database is $25–$500 and embeddings $5–$50. The same workload costs ~$81 a month on a small model and ~$1,350 on a large one — which is why a subscription price and a real usage bill are not the same number. 4 Context Link, RAG as a Service — managed RAG platforms span $9–$19 a month for small-business tools, $100–$1,500 a month for platform builders, and $12,000–$50,000+ a year for enterprise infrastructure, the last “often requiring annual contracts,” “4–12 weeks of implementation” and “dedicated engineering resources.” Further reading: Forbes Business Council, Unleashing The Power Of Retrieval-Augmented Generation For Medium-Sized Businesses, also worth reading for its warning about vendors who use AI terminology to imply expertise they don't have.