Services
RAG development services: your experts' knowledge, answered in the work
When growth depends on a few experienced people, what they know becomes a business priority. We capture that expertise, get it reviewed by the people who own it, and put supported answers in front of the team where they work.
When knowledge is the bottleneck
The same three signs show up in every operation we walk. Routine questions keep landing on the same one or two people, and those people are the ones you most need on something else. Onboarding takes months because the real procedures were never written down. Work gets repeated, a site gets revisited, or a part gets ordered twice, because the person who knew was on another job.
None of that is a documentation problem you can fix with a shared drive. The knowledge is in heads, in habits, and in the way a specific site or machine behaves. Capturing it takes a method, and keeping it useful takes an owner.
Start from the objective. Protect service margin: fewer repeat visits caused by missing equipment or procedure knowledge. Support growth: bring new technicians or staff to productive faster. Free leadership capacity: stop the interruptions that reach the owner. Preserve continuity: make the business less dependent on any one person. Each of those has an operating measure behind it, and the measure is what we build toward.
What an AI knowledge system does
Capture. Guided voice or text conversations, plus photos, record how the work is actually done, including what is missing from the existing documents. Busy experts talk; the system asks the follow-up questions and drafts.
Structure. Each capture becomes a domain-specific article: a site profile, an equipment procedure, a troubleshooting guide, a vendor profile, a client protocol, a parts and substitution reference, a safety note, a service SOP. Gaps are flagged. Sources stay attached.
Review. An accountable expert corrects and approves. Versions and publication status stay visible, so everyone knows what is current and who signed off.
Retrieve. Permitted users ask questions in plain language and get an answer with the article behind it, so they can check the source. When the system does not have a supported answer, it says so and routes the question to a person.
Apply. Knowledge is connected to the property, equipment, customer, or job it belongs to, inside the tools the team already uses. Where the business case supports it, a separately scoped agent can act on it.
Maintain. Content gets updated, failed questions get investigated, retrieval gets retested as the business changes. That is a recurring service with a named owner on your side and ours.
Source to answer
Sample dataFrom one expert's head to an answer the crew can trust
1 Capture
A senior technician, on the phone from a job site: "The Unit 4 compressor trips on high discharge pressure about once a month in summer. It is almost always the condenser coil. Check the fan first, then wash the coil. Do not reset it more than twice." The assistant asks: "What pressure counts as high on this unit?" Answer: "Over 400 psi on the R-410A units."
2 Review
The draft article, "Unit 4 compressor: high discharge pressure trips", goes to the service manager. She adds one line, "Lock out the disconnect before washing the coil," and approves it. The article now carries her name and the date.
3 Answer
A newer technician asks from the truck: "Unit 4 tripped again, high pressure, what do I check?" The system answers with the four steps, the 400 psi threshold, and the lockout line, and shows the source article and who approved it. He opens it and calls no one.
When it does not know
He asks: "Is it the same for the rooftop units at the north site?" The system answers: "There is no reviewed article for the north site rooftop units. The closest is Unit 4, which is a different model. Ask Maria or the service manager." It does not guess.
Sample data. The names, the unit, and the numbers are invented to show the pattern. A real system answers only from articles your people wrote and approved.
When an existing tool is enough
If your documents are well maintained and what you need is search, a knowledge product you can buy is the right answer, and we will say so. Several established tools do verification, permissions, and cited answers well. Custom work earns its place when the knowledge has to be elicited from people rather than found in files, when it has to be structured around your real sites and equipment, when it has to sit inside a job or a work order, or when access rules and operating needs are not met by an off-the-shelf product.
The system does not know everything, does not guess when it lacks a source, and does not replace experienced judgment. It makes the judgment of your best people available to more of your team, with their review on it.
Who sees what, and how you know it is right
Every article has a reviewer and a status. Every answer carries its source. Access is by role: a field technician, an office coordinator, and a subcontractor do not see the same material, and client-specific protocols never leak across clients. Acceptance is tested against real questions before launch: does the answer match the source, does the citation support it, does the system refuse when it should, do the permission boundaries hold, does stale or conflicting content get flagged.
A focused first engagement
One knowledge domain. A small group of experts who own it. A defined set of source material and capture channels. Named user roles. At most one business-system integration, so answers land where the work happens. We deliver a working capture, review, and retrieval path, an evaluation set of representative questions, user onboarding, and an export and handover procedure, so the content is yours in a form you can take anywhere.
The assessment asks the questions that shape the scope: which questions repeatedly interrupt an owner or senior employee, and which are costly to answer wrong; what exists in documents and what must be elicited from people; who can validate and maintain the content after launch; which users may see which sources; how the system should handle missing or contradictory information; whether answering requires an action in another system. A well-defined knowledge build can proceed straight to scope without the full assessment.
Proof
A Houston security, access-control, and IT services contractor came to us with undocumented site, equipment, troubleshooting, and supplier knowledge held by a few experienced people. We designed and implemented the core of a knowledge system for them: an editor and library, voice and text capture assistants with gap tracking, versioning and publication, team and role management, and lookups into their job-management software. The client tested it against their real work and fed back specific requests and defects, which is what the parallel period is for. Full field rollout and continued content development are scoped separately, on their timeline.
Founder-built by Jonathan Klein and Chip Ray, based in Houston. Insured: cyber liability, technology E&O, and general liability.
What it costs to build a RAG application
Knowledge work is quoted in parts, so you can see what you are buying: discovery, the system, initial content capture and curation, integrations, ongoing knowledge stewardship, and third-party usage. Software maintenance does not automatically include interviewing experts or approving technical content; stewardship is its own line, with the client responsible for naming qualified approvers.
- AI Readiness Assessment: from $10,000, credited once against the first build. A well-defined knowledge build can skip it.
- Build workstreams: fixed fee, from $10,000 each, sized by domains, expert time, source quality and volume, permissions, and integrations. Quoted before work starts.
- Managed AI and knowledge stewardship: from $1,200 per month for the system; stewardship scoped by review load and capture cadence.
Prices verified September 2026. Full pricing →
What happens after go-live
Knowledge goes stale. Vendors change their APIs. New sites and new equipment arrive. The managed plan monitors the system and fixes it when something upstream changes; stewardship keeps the content current: a review queue, scheduled capture sessions, stale-content checks, failed-question analysis, and an evaluation report. How managed AI works →
Frequently asked questions
No. A chatbot answers from whatever it was trained on. This answers from your reviewed content, shows the source, and refuses when it has none. The capture and review steps are most of the work; the assistant is the last mile.
The first domain is scoped so a working capture, review, and retrieval path can be in use in weeks, not quarters. Content grows from there on a cadence you set, which is why stewardship is priced separately.
You do. Articles, sources, and the system itself transfer to you, with an export procedure that works even if you leave us. Our reusable tooling stays ours; your knowledge is yours.
Both, in stages. Assisted retrieval delivers value on its own. Where the business case supports it, a separately scoped agent can use the knowledge to take a defined action in your systems, with a person approving anything that matters.
Stewardship is scoped by how much content changes and how many experts are involved. It is a recurring service with a named owner, not an afterthought inside a support plan.
Which questions keep reaching the same two people?
Tell us what your business needs to achieve and where expertise is the constraint. We will help connect the objective to a scoped first domain.