Shimira ITInternal AI knowledge assistant
Ask your company's documents a question. Get a cited answer.
Your company knows the answer. It's in a policy document, a wiki page, a past proposal, or a support article nobody can find. So people ask a colleague instead, and that colleague spends their week answering the same questions.
A knowledge assistant fixes the finding problem, not the knowing problem. It searches your own documents, answers in plain language, and cites the source so the reader can check it. Built properly, it says “I don't know” when your documents don't cover something, which is the single most important thing it can do.
What matters here
What a knowledge assistant needs to get right
Answers from your documents only
Retrieval finds the relevant passages first, and the model answers from those. It isn't free to improvise from general knowledge.
Citations on every answer
Each answer links to the passage it came from, so a reader can verify it in seconds. Without this, nobody trusts it twice.
Respects who can see what
HR documents shouldn't surface for everyone. Permissions are applied at retrieval, so the assistant can only use documents that reader is allowed to see.
Stays current
Documents change. Re-indexing keeps answers in step, and superseded versions stop being quoted as current policy.
Admits the gap
When your documents don't answer the question, it says so and points to who might. A confident wrong answer is worse than no answer.
- Next.js
- Node.js
- TypeScript
- PostgreSQL
- OpenAI
- AWS
We've shipped in this space
SnuggleRead.com
AI running in production inside strict guardrails, generating output that has to be right every time, which is the same discipline a knowledge assistant needs.
What goes wrong
Why internal AI assistants fail
Retrieval is the hard part, not the model
Most disappointing assistants have a fine model and poor retrieval: the right passage was never found, so the answer was never possible. Most of the engineering goes into how documents are split, indexed, and searched.
Messy source documents
Five versions of the same policy, three of them outdated, produces contradictory answers. Deciding what's authoritative is work your team has to do, and it's usually the bottleneck.
No permissions model
An assistant that ignores access rules will eventually quote a salary band or a disciplinary policy to the wrong person.
No measurement
Without a set of real questions and correct answers, nobody can say whether it's good. We build that set before launch and re-run it whenever the model changes.
Launched and forgotten
Real questions differ from imagined ones. Reviewing what people actually ask, and what it got wrong, is what turns a demo into a tool people rely on.
How it runs
How we build it
- 01
Feasibility
1–2 weeksReal questions from your team, tested against your real documents. We measure how often a correct, cited answer is achievable and estimate running cost.
- 02
Pilot on one domain
3–6 weeksOne body of knowledge, such as HR policies or support documentation, with citations and an evaluation harness.
- 03
Production build
2–4 monthsPermissions, more sources, sign-on, re-indexing, monitoring, and wherever your people already work — Slack, an intranet, or your own product.
- 04
Tune on real questions
ongoingReview what's asked and what's missed, improve retrieval, close the documentation gaps it exposes.
Timelines & pricing
Honest ranges, before you commit.
| Tier | What's included | Timeline | Typical cost |
|---|---|---|---|
| Feasibility assessment | Real questions tested on your documents, accuracy measured, running cost estimated | 1–2 weeks | $3,000 – $8,000 |
| Single-domain assistant | One document set, citations, one interface | 4–8 weeks | $15,000 – $45,000 |
| Company-wide assistant | Multiple sources, permissions, SSO, re-indexing, monitoring | 2–4 months | $30,000 – $90,000 |
Running costs depend on how many questions are asked and how much context each answer needs; better retrieval lowers both. Our AI app development cost guide works through a monthly estimate for an assistant answering 50,000 questions.
Before you choose
Is your knowledge ready for this?
- Your documents can be exported programmatically, not just read in a browser.
- Someone can say which version of a document is authoritative.
- The same questions genuinely come up over and over.
- You can write down 50–200 real questions with correct answers.
- You know which documents are restricted and who may see them.
- Someone will own it after launch, reviewing misses and fixing gaps.
Every “no” here is preparation work, and it's much cheaper to find out during feasibility than mid-build.
How you work with us
Three ways to engage.
Fixed price
Prototypes from $7,500
Production MVPs from $25,000
You know what you need built. We scope it, quote one price, and deliver against it, no meter running.
- One agreed price, agreed before we start
- Milestone-based payments tied to what you can see
- Change requests priced openly, never assumed
Time & materials
from $28 / hour
Blended ~$32; senior specialists to ~$45
For work that changes as you learn. You pay for the hours spent, see exactly where they go, and can steer week to week.
- Billed on real, logged hours, reviewed with you
- Reprioritise or change direction any sprint
- Start small, scale the team as it proves out
Dedicated team
from $2,800 / month
Per developer, full-time, see the rate card
Developers who work only on your product, full-time, as an extension of your team, the offshore model most of our long-term clients settle into.
- The same people, month after month
- Your tools, your standups, your roadmap
- Roughly half the loaded cost of a local hire
Full rate card and what's included on the pricing page. Or hire RAG and LangChain developers for your own team.
Questions
Before you get in touch.
That's the risk the whole design guards against. Answers are grounded in retrieved passages and cited, so a reader can check them, and the system is built to say it doesn't know rather than improvise. We measure this before launch rather than assuming it.
Almost never. Retrieval over your documents is cheaper, updates instantly when a document changes, and lets you cite sources. Training bakes information in, which makes it hard to update and impossible to attribute.
A feasibility assessment is $3,000–$8,000. A single-domain assistant runs $15,000–$45,000 and a company-wide one $30,000–$90,000, plus a running cost per question.
Yes, and it should. Permissions are applied at retrieval, so restricted documents are never used to answer someone who isn't allowed to see them.
Wherever they already work: Slack or Teams, an intranet page, your own product, or a standalone site. The interface is the easy part; retrieval quality is what decides whether it's useful.
Providers retire models regularly. Because we keep an evaluation set, switching is a measured change rather than a leap of faith: we re-run the tests on the new model and compare before switching.
Cost guides
Budgeting for this?
Real 2026 numbers, worked examples, and the hidden costs most quotes leave out.
How much does AI app development cost in 2026?
Build ranges by use case, a worked monthly model-cost example, RAG vs. fine-tuning, ROI maths, and the hidden costs most AI quotes skip.
Read the guide →How much does custom software development cost in 2026?
Real 2026 numbers: cost by project type, hourly rates across every major region, the hidden costs quotes leave out, and a free calculator to model your own estimate.
Read the guide →Not sure what you need yet? That's the usual starting point.
Tell us the problem in your own words. We'll scope it with you and put the plan in writing, free, and yours to keep either way.