Your content already
has the answers
They're in your website, spec sheets, procedures and contracts. We build an assistant that finds them and answers in plain language, with a link to the exact page it used. Put it on your website for buyers, or use it inside for your team.
- Answers cite the document and the page
- Runs in your own cloud where compliance requires it
- Looked after by the engineers who built it
“The AI RAG Agent Sova built for our site is exactly what we hoped for.
Read the full quoteHide the full quote
It works like a site concierge for visitors who’d rather ask a question than dig through pages themselves, by pulling answers straight from our own knowledge base and pointing people toward deeper resources on the site. It’s helped increase site engagement, lengthen time on site, pages viewed, and encourages conversions, all while making Top10ERP.org genuinely more searchable and user-friendly for anyone who prefers a guided experience.
The custom dashboard has been just as valuable, giving us real insight into what visitors are asking so we know exactly where to expand our content next.
It truly feels like Sova brought our site into the future.”
Top10ERP.org helps businesses compare ERP software. Their assistant answers visitors' questions from Top10ERP's own vendor and product content, links to the page each answer came from, and sends every question it can't answer to a dashboard their content team works through.
You have the answers.
Nobody can find them.
Most companies have years of documentation. The hard part is finding the one paragraph you need while a customer waits on the phone, or before a website visitor gives up and leaves.
Six places to look
Your website, the shared drive, a folder of supplier PDFs, old support tickets, the product pages and someone's inbox. The answer is in one of them. Nobody's sure which.
Search wants the exact words
A search box only finds the words in the document, but people search with the words their customer used. Type “can we run it wet?” and you'll never land on a page titled “Environmental rating, IP69K.”
Somebody pays for it
Inside the company, the question lands on your most experienced (and most interrupted) person, and what they know leaves when they do. On your website, visitors don't ask at all. They close the tab, and you never find out what they wanted.
ChatGPT hasn't read your manuals
An AI model only knows what it learned in training. It's never seen your warranty policy or your tolerance tables, so it fills the gap with something that sounds right and isn't. Those are the mistakes that cost money.
Look it up, then answer
Our assistant searches your material for the passages that answer the question, and it's only allowed to answer from those. It checks your documents fresh every time someone asks.
Every answer shows its source
Each answer links to the document, section and page it came from, so anyone can check it in one click. That's what makes people trust it enough to use on a live call, or to buy on.
It checks the book
before it answers
A general AI model answers from memory and hopes it's right. Ours looks the answer up in your documents, answers only from what it finds, and shows you the page. Here's how, in four steps.
-
01
Index
We break your pages and documents into short passages and store them so they can be found by what they mean, not just the words they use. Text, tables and PDFs, even scanned ones.
Once, then kept up to date -
02
Find
When a question comes in, we pull the few passages that actually answer it. We match on meaning, so the person asking doesn't need to use the same words as the document.
Under a second -
03
Stick to the source
Those passages go to the AI with one rule: answer from this material and nothing else. It has nothing else to draw on, so it has nothing to make things up from.
On every answer -
04
Answer
You get a clear answer, with each point linked to the passage it came from. If your documents don't cover the question, it says so and points the person to someone who can help.
Sources included
The knowledge lives in your documents, not inside the model, so nothing needs retraining when things change. Fix a procedure in the morning, re-index it, and the afternoon's answers use the fix.2
For your customers,
and for your team
It's the same assistant either way: sources on every answer, and no making things up. What changes is who's asking, and what happens when the conversation ends.
The question your best person gets asked twice a week
One place to ask about your SOPs, spec sheets, contracts and manuals. Your team gets the answer and the page it came from, without tracking down the one person who knows.
- Answers cite the document, section and page
- Scoped to the material you choose to index
- Unanswered questions routed to the document owner
- Self-hosted where policy or a regulator requires it
Start with: sales engineering, field service, support or operations, whichever team owns the specs. The team that gets interrupted most already knows which questions keep coming up.
The visitor who won't fill in a contact form
People shopping around don't ask just one question. They compare, push back and narrow things down. The assistant answers from what you've published, and when someone looks ready to buy, your team hears about it within minutes.
- Deep links into live pages and PDF pages
- Spots buying intent during the conversation
- Routed to CRM, Slack or inbox with the full transcript
- Indexes whitepapers behind a form, without breaking the form
Start with: the pages where people make up their minds, like products, specs, comparisons and pricing. Questions asked there are worth money.
Two conversations.
Watch the second answer.
Both times, the assistant gets a question your documents can't answer. It says so, explains why, and passes the person to someone who can help, instead of guessing.
If it guessed at that second question, your team would believe it once and never trust it again. Saying “I don't know,” naming the missing document and finding the person who owns it is what makes people comfortable using it on a live call.
It had no source for that answer, so it said so, logged the gap for your content team and offered to put the visitor in touch with your team. A made-up lead time could have cost you the deal three weeks later.
Search that understands
the question
Most site search matches the words people type, not what they mean. Ours runs on the same engine as the assistant and looks at both, so visitors can type the question they actually have and get results that fit.
It only matches words. The spec sheet says “IP69K” and “high-pressure caustic cleaning,” so the right product is there but never shows up.
- 400 Series stainless motorIP69K · 316 stainless housing
- 450 Series hygienic gearmotorIP69K · food-grade lubricant
- Wash-down maintenance scheduleGuide · §3.2 cleaning intervals
It understands what the visitor means, and still catches exact terms like part numbers, model names and ratings, which is what keyword search was good at.
- Words and meaning
- Keyword search and meaning-based search work together, so “IP69K” and “survives a pressure washer” find the same product.
- Ask it like a person
- Visitors can type their whole question, details and all, instead of guessing which two words your site happens to use.
- A cited answer on top
- If you want, a short answer appears above the results with a link to its source. The usual results stay below for people who like to browse.
- Every miss is a to-do
- Searches that come up empty go into the same gap report as the assistant’s unanswered questions.
It replaces the search box you have now, with or without the assistant. Fewer dead ends means more pages viewed, longer visits and more people reaching the pages that convert.
What you can see,
and what stays yours
Down to the paragraph
Every claim links to the exact paragraph it came from, whether that's a page on your site or page 7 of a PDF. Readers can check it for themselves, and you can review any answer later.
- Sources down to the paragraph
- Deep links into live pages and PDF pages
- Full audit trail of every answer and its sources
You can see when it's struggling
Every conversation is scored on what actually happened: thumbs up or down, fallbacks, people leaving, escalations, and whether it led anywhere useful. Filter to the bad ones and read exactly what went wrong.
- Good / neutral / bad session labels, with reasons
- Drill from any chart straight into the transcripts
- Token and cost usage tracked per conversation
See what people can't find
The questions your documents can't answer show you what people want. We rank them by how often they come up and show who asked, so you know what to write next and, on a public site, who to call.
- Ranked unanswered topics, clustered automatically
- Monthly digest of what's missing and what's rising
- Which documents get cited, and which never do
It stays yours
It's usually the first thing buyers ask, so here's the short answer. Your documents are indexed into a store you own. They're never used to train anything. If your policy or a regulator requires it, the whole system runs inside your own infrastructure and nothing leaves it.
- Your content is never training data
- Self-hosted deployment where it's required
- Retention rules and transcript redaction you set
- No third-party cookies, and passes security review
Three ways to get this done
We'd rather you pick the right option than pick us. Here's a straight comparison, including where a cheap off-the-shelf tool beats us.
| Criteria | Build in-house | Off-the-shelf tool | Work with us |
|---|---|---|---|
| Time to something useful | 8–14 weeks for a median build, and it's 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 charge $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, starting at $250. 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 | No one on your side |
| Handles messy PDFs, tables, scans | Eventually, if you fund it | Poorly, and it won't tell you | Yes. It's most of what we do |
| 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, all in one place, and a wrong answer won't hurt anyone, buy an off-the-shelf tool for a few hundred a month. You don't need us. We're worth it when the material is messy and a confident wrong answer would cost you.
One fixed build fee,
then a month at a time
An assistant nobody looks after goes stale within a few months, because your material keeps changing. So every build comes with a monthly plan that keeps it accurate. We quote both after one call, and there's no long-term contract.
What’s included
- PDFs, spec sheets, scans and gated documents
- Citations and strict grounding on every answer
- Internal deployment, public widget, or both
- CRM, Slack and calendar handoff
- Judged against twenty questions you choose
- Gap report on what your material can't answer
What’s included
- Hosting and all running costs included
- Re-indexing as your documents change
- Monthly tuning against real transcripts
- Wrong answers fixed, with an explanation of why
- Model upgrades handled, at no extra cost
- Monthly report: questions, quality, gaps, cost
What’s included
- Monthly subscription from go-live, no minimum term
- Cancel any time, with no exit fee or balance owed
- Half the build on signature, half on go-live
- Export your content, conversations and leads any time, including if you leave
- No annual contract to negotiate, budget for or escape
What it costs depends on what you've got
The build fee depends on how much material you have and what shape it's in. A tidy set of manuals and forty thousand scanned pages are very different jobs. A 20-minute call is enough for us to give you a fixed price, or to tell you honestly if you don't need us.
We only take on a few builds at a time, because each one is tuned by hand. If there's a wait for a slot, we'll tell you on the first call.
Tell us how we can help your team
Where do people get stuck? Maybe it's the same few questions landing on one person's desk, or buyers who can't find the spec they need. A few lines is plenty, and it helps us bring something useful to the first call.
The first call takes twenty minutes. If your material isn't ready for this yet, we'll tell you and explain what would need to change.
Sources and further reading
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.
Ask Sova
Pick a question. Each answer comes from this page, with the section it came from.
Demo: answers are written from this page, not a live model.