Knowledge AI
Every business has a Dave — the one who knows what the handbook actually says, why you switched suppliers, and what you did at that customer’s place back in 2019. A knowledge assistant reads everything Dave knows into a system your whole crew can ask, in plain English, at 2 a.m., with the page it came from attached.
Try It
Pick a question below and watch what actually happens between the question and the answer. This is the same shape of flow that runs on a real deployment — the difference is that yours reads your documents.
Choose a question on the left to run it through the pipeline.
>
Passages retrieved from your documents
Grounded answer
Demonstration The documents, passages, names and figures above are fabricated to illustrate the flow — they are not real policies, contracts or customer records. A real assistant is built against your documents and cites your pages.
Under the Hood
Here’s every stage of what you’d be getting — and beside each one, where that stage breaks once your documents are real, messy, and scanned in crooked. That second column is the reason to bring someone in: eight specific ways these systems fail quietly, learned by building them rather than reading about them. Keeping your project clear of all eight is the job I’d be doing for you.
Pull documents out of wherever they live — shared drives, email attachments, that filing cabinet you scanned in 2017 — and turn PDFs, spreadsheets, drawings and photos into clean, machine-readable text.
Where it breaks Scanned pages with no text layer, multi-column forms that read out of order, and tables that flatten into gibberish. Most failed projects fail right here.
Cut the text into passages small enough to search precisely but large enough to still make sense on their own — and keep the trail back to the source document and page number.
Where it breaks Cut badly and a price table gets severed from the heading that says which year it applies to. The answer looks confident and quotes last year’s number.
Convert each passage into a numeric fingerprint that captures meaning, so a search for “vacation” also finds the page that only ever says “paid time off.”
Where it breaks Generic models don’t know your trade’s shorthand, your customers’ names, or what your crew calls things. Getting your vocabulary to land in the right place is tuning work, not a checkbox.
Keep those fingerprints in a searchable index that returns the closest matches in milliseconds, on infrastructure you control, with your documents never leaving your tenant.
Where it breaks Permissions. If the index doesn’t know that payroll folder is off-limits, the assistant will cheerfully quote it to whoever asks.
Pull the candidate passages for a question, then run a second, sharper pass that reorders them so the genuinely relevant page beats the one that merely shares vocabulary.
Where it breaks Skipping the rerank. It’s the single cheapest upgrade to answer quality, and it’s the step most DIY builds leave out.
Hand the model the question plus only the retrieved passages, and require every claim to trace back to one — so answers come with the document and page attached, and “I don’t know” is an acceptable answer.
Where it breaks An assistant that would rather invent a plausible number than admit the manual doesn’t cover it. Making refusal the safe default is deliberate engineering.
Build a scored test set from real questions your people actually ask, so every change is measured instead of eyeballed — and you can prove the thing got better, not just different.
Where it breaks There is no evaluation. The demo dazzles, month three quietly degrades, and nobody can say when or why.
New revisions land, procedures change, people leave. The index has to keep up, superseded documents have to stop being quoted, and someone has to own that.
Where it breaks The assistant confidently cites Rev B for two years after Rev C shipped. Stale answers erode trust faster than no answers.
Straight Talk
A knowledge assistant demos beautifully on three clean documents. The trouble starts at twelve years of real files — scans with no text layer, four revisions of the same procedure, a spreadsheet where the important number lives in a merged cell. Closing that gap is the whole job, and it’s the part I’ve already spent the time learning.
That’s the work you’re hiring for: getting your specific documents to parse cleanly, teaching the system your trade’s vocabulary, tuning retrieval until the right page wins, building a test set from the questions your crew actually asks, and making the whole thing refuse to guess rather than invent. It’s measured in your documents, not in a diagram.
And you’d be working with the person who does it. No account manager, no handoff to a team you never meet, no enterprise platform priced for a company ten times your size. I’m in North Dakota — I can come look at the actual filing cabinet, and you have my number when something breaks.
If an assistant isn’t the right answer for your business, I’ll tell you that in the first conversation instead of selling you one. Advice like that means more coming from the person who’d be paid to build it.
Tell me the question your crew asks most often and where the answer currently lives. I’ll tell you straight whether an assistant would help, what it would take, and whether it’s worth it. Free, no obligation.
Get In Touch