Skip to content
For Agencies Talk to Us

Knowledge systems that answer
and cite their sources.

Ask a question and get an answer with the source attached. It comes from the manuals, policies, wikis and tickets your company already has. The same system can run your site search, support tools and internal assistants. People only see what they're allowed to see.

Get a ProposalGet a Free AI Automation Review

Reply within four business hours (typical) · Respects who can see what

Brands our teams have delivered for

  • Hello Peter
  • Gruber Logistics
  • Delhivery
  • Thomson Reuters
  • Qatar Airways
  • Grundfos
  • Save
  • BERD
  • Yale University
  • Kuwait Police
  • Dubai Police
  • Panasonic
  • Infosys
  • Kia
  • Hitachi
  • Orange Business Services
What we do

PixelCrayons builds AI knowledge bases and RAG chatbots that answer plain-language questions from your own documents, with the source attached. The system finds the right passages, then writes the answer only from them. It shows people only what they may see, and we test it on real questions first. It starts at $1,500 to build and $350 a month.

  1. One system for search, support and your teamOne search system behind site search, helpdesk lookup and internal assistants in Slack or your intranet.
  2. Every answer shows its sourceEvery answer names its source documents, and with no source there's no answer.
  3. Search that respects who can see whatSearch follows your existing access rules, so nobody sees documents they can't already open.
Our work

See what we've built
and the results it brought.

Real projects for real clients, with results we can show.

All case studies →
2004
Established
500+
Agencies served
4,500+
Projects delivered
340%
Revenue growth · 7 months
eCommerce
+127%
Organic traffic · 5 months
SaaS
85%
Faster delivery · backlog cleared
via agency partner
Clutch rating 4.9 / 5, checked 15 September 2026Read our GoodFirms reviews
Google Partner
Shopify Partner
ISO/IEC 27001 certification seal (QRO), held by the Vinove groupISO 9001:2015 certification seal (QRO), held by the Vinove groupCMMI Level 3, held by the Vinove group
Technologies we work with
Commerce and web
  • Shopify
  • WooCommerce
  • Magento
  • WordPress
  • Webflow
  • Next.js
Marketing
  • Google Ads
  • Meta Ads
  • GA4
  • Klaviyo
Cloud
  • AWS
  • Microsoft Azure
  • Google Cloud
AI models
  • OpenAI
  • Anthropic
  • Google Gemini
Build and deploy
  • Docker
  • Kubernetes
  • Terraform
  • GitHub Actions
CRM and automation
  • HubSpot
  • Salesforce
  • Zapier
  • n8n

Product names and marks are trademarks of their owners. Naming a technology means our teams build with it. It does not mean a partnership, certification or endorsement.

  • Every answer cites its sources
  • Access rules respected as standard
  • No source, no answer: enforced by the system
  • Reply within four business hours (typical)
  • One system: search, support and internal tools
  • Tested before launch, logged after
In every build

What a knowledge system
actually includes.

01

One system for search, support and your team

A knowledge system isn't a chat window. It's the search system underneath. We build it once, over your documents, and use it wherever people ask questions. Site search gives answers instead of links. A lookup in your helpdesk stops agents hunting through folders. Internal assistants work in Slack or your intranet. If you want a chat window for customers, that's our chatbot service, built on this same system.

The architecture
02

Every answer shows its source

Every answer names the documents it came from: the clause, page or ticket. A person can check it in one click instead of taking the model's word. When the search finds nothing, or finds sources that disagree, the system says so instead of making something up. No source, no answer. The system itself enforces that rule, not a request in a prompt.

Every answer
03

Search that respects who can see what

The system copies your existing access rules: SSO groups, folder permissions and workspaces. So it never searches documents a person isn't allowed to see. A clever question can't show the board pack to a contractor, because it was never in their search. Access decisions are logged, and every question can be checked afterwards.

Every build
04

Testing before launch

Before anyone relies on it, we test the system on questions your team really asks. We check it finds the right documents and answers only from them. We check it refuses the questions it must refuse. You see the scores and sign off on launch. The tests run again whenever the documents change, so quality is measured, not assumed.

Before launch
05

Documents that stay up to date

Documents change every week. A system that read them once is out of date by month two. Regular updates pick up changes to your sources on a timetable you set. They flag documents that contradict each other and show the gaps people keep asking about. That list of gaps tells you what to write next. At handover your team gets the runbook (a written guide). Your materials remain yours. Bespoke deliverable rights, licences and handover are agreed before work starts.

At handover
Where answer quality is really decided

Why most bad answers
start before the model.

An internal assistant is a chain of steps with a model near the end. When answers are vague, out of date or quietly wrong, the cause is usually three or four steps earlier. That is why “try a better model” so rarely helps.

Loading documents: runs on a schedule
  1. 01

    Sources

    The documents, tickets, policies and wiki pages that count as the truth.

    What breaks here

    Nobody looks after the documents, so they quietly go out of date. The assistant then answers confidently from last year’s policy.

  2. 02

    Splitting

    Documents are split into passages small enough to find and large enough to make sense.

    What breaks here

    A split in the wrong place separates a number from its heading or a step from its warning. The passage can still be found. It is just no longer true on its own.

  3. 03

    Permissions

    Each passage carries the access rules of the document it came from.

    What breaks here

    Rules applied after the search instead of during it. The answer looks filtered, but the model has already read what it should not have.

The index

The only thing the assistant can see. It can't answer from anything outside the index, and it can answer from anything inside. So the index, not the prompt, is what keeps your data safe.

Answering: runs for each question
  1. 04

    Find

    The question is matched against the index, within what the person asking may see.

    What breaks here

    A search tuned only to find similar text returns four passages that nearly answer the question, and none that do.

  2. 05

    Answer

    The model is given the passages found and told to answer only from them.

    What breaks here

    No rule for when nothing is found. Asked something the documents don't cover, the model answers anyway. That one failure destroys trust in an internal tool.

  3. 06

    Cite

    The answer links back to the passages it came from, so a reader can check it in one click.

    What breaks here

    A citation that points to a whole 40-page document. It has a source, but nobody can check it in practice.

Ask any vendor where access rules are applied. If they say “we filter the results”, the model has already read a document it was not allowed to read.

Illustrative pipeline: stages are ours, not a vendor reference architecture

How it starts

From scattered documents
to cited answers.

01
Start

Brief and a call

We sign an NDA and map where your knowledge is kept and who can see it. We collect the real questions the system must answer.

02
Once the brief is complete

Audit and itemised proposal

We confirm the date your proposal arrives. It lists the sources in scope, where it will be used, the access rules, a demo date and the price.

03
Your decision

You sign off

You confirm the scope, the payment and the access we need. Nothing starts before this.

04
Typically 2 to 3 business days later

The team starts

We build the pilot over your own documents. Its tests start from your team's real questions.

05
On the date in your proposal

Working pilot, your documents

You ask questions of a working pilot that cites its sources and respects your access rules. Next we test it, including the edge cases of who can see what. Then it opens to real users.

Inside Prism

Your engagement, week to week,
in one workspace.

Prism is the client workspace where your requests, approvals, tasks and weekly reviews live.

Some engagements include Prism, our client workspace, and your proposal says whether yours does. Where it does, your AI build runs in a workspace you log into: requests, approvals, the task list and the weekly review, with each decision and its expected result written down.

See how Prism runs an engagement
  • 01

    Requests and approvals

    One queue, one owner, one due date.

  • 02

    Weekly review

    Each decision recorded with its expected result.

  • 03

    Actions checked against outcome

    What we did and what happened, side by side.

How we report it, in Prism

The eval scorecard,
before the pilot goes near a customer.

What you receive in month one

AI work is reported against an evaluation set built before any model is chosen. In the first month you receive the task definition with the inputs and acceptable outputs written down, the eval set of real cases with expected answers, and the scorecard showing accuracy, refusal behaviour and cost per run. The go/no-go checklist names what must hold before the pilot touches live traffic.

  • 01Task definition: inputs, acceptable outputs, and the cases the system must decline
  • 02Eval set: real cases with expected answers, reviewed by your team
  • 03Scorecard: accuracy, refusal behaviour, latency and cost per run
  • 04Go/no-go checklist: the thresholds the pilot must meet before launch
  • 05Failure log: each wrong answer, its cause, and the fix applied

Related workB2B SaaS: from invisible to answer-engine cited.SaaS · UK · A different discipline, so it is linked here rather than presented as proof of this one.

Our other AI services

Other AI services built on a knowledge system

Put a conversation on top of an AI-powered knowledge base and it is a chatbot. Feed its answers into a workflow and it is automation. Start where the problem is, or start at the AI page.

AI ChatbotsWorkflow AutomationAll AI solutions
Who is accountable

The people responsible
for your work.

Leadership and how we’re built →
Parvessh Agarwal

Parvessh Agarwal

Founder & CEO

An entrepreneur at heart. He links business analysis, strategy, clear goals and ROI to practical technology.

LinkedIn →
Saundaraya Gupta

Saundaraya Gupta

Technical Manager

Brings technical work, client updates and project management together, so work moves smoothly from brief to launch.

LinkedIn →

Every engagement

A named coordinator and escalation contact

Know who coordinates your engagement, who owns the work and how to escalate an issue.

All engagements. Roles may be combined and vary with your engagement.

Full terms →
Before you sign

Our commitments

Each one says where it applies and links to its full terms.

Response and start times

We typically respond within four business hours. Eligible engagements can typically start within two to three business days after scope, payment and required access are confirmed.

Business hours are Monday to Friday, 9:00 am to 6:00 pm IST, excluding published company holidays. Typical expectations, not guaranteed service levels.

Full terms

Clear IP & Ownership Terms

Your materials remain yours. Bespoke deliverable rights, licences and handover are agreed before work starts.

All engagements.

Full terms

Clear Delivery & Escalation Ownership

Know who coordinates your engagement, who owns the work and how to escalate an issue.

All engagements. Roles may be combined and vary with your engagement.

Full terms

How we handle claims, credentials and your data: group credentials, dated ratings, how case figures are signed off, client permissions and data handling.

Where to start

Knowledge system, chatbot,
or something else first?

Our review of your documents and questions decides what gets built. Sometimes the answer is not a knowledge system yet. These are the choices it makes, and the reason for each.

What you haveWhere we would startWhy
A support queue that needs a customer-facing chatOur chatbot serviceIt builds the conversation, the hand-over rules and the helpdesk connection. It runs on this same search system anyway.
Several teams asking questions of the same large set of documentsA knowledge system, one use firstOne system can serve site search, a helpdesk lookup and internal assistants. Each is added once it proves useful.
A plan to train a model on your data so it knows your factsSearch your documents (RAG) instead of training a modelTraining a model is costly to update, cannot cite sources and cannot respect who may see what. With RAG, updating the system means updating a document.
Dated policies, overlapping wikis and tickets that contradict the manualBuild, with sorting the sources as part of the jobWe sort trusted sources from out-of-date ones. The system flags where sources disagree instead of blending them.
Knowledge that was never written downA prioritised list of the unanswered questionsNo system can answer what no document says. The list of gaps tells you what to write next.
Confidential material that must stay insideAccess rules applied during the search, in your own cloud where neededA question only searches what that person could already open, so restricted documents are never found.

Where training a model on your tone or format is worth it, the discovery document says so.

Questions

Frequently
asked.

RAG (retrieval-augmented generation) keeps your knowledge in your documents. A question finds the right passages, and the model answers only from them and cites them. Training a model on your data (fine-tuning) is the wrong tool for facts. It's costly to update, it can't cite sources and it can't respect who may see what. With RAG, updating the system means updating a document.

A chatbot is what people talk to. A knowledge system is the part underneath that finds the answers. Our chatbot service builds a conversation, usually for customers. This service builds the search system that can also run site search, a helpdesk lookup and internal assistants. If several teams ask questions of the same large set of documents, start here.

It can make things up, and no vendor can promise it never will. A well-built system answers only from the passages it finds. It cites every answer so a person can check it. It says so when sources are missing or disagree. It's tested on your team's real questions before launch, and every question is logged after.

Access rules are built into the search itself. The system copies your SSO groups, folder permissions and workspaces, so a question only searches documents that person could already open. Every question is logged, work runs under NDA, and the system can run inside your own cloud. Rules for keeping, storing and accessing data are agreed before we load any documents.

Usually yes, and the audit will say where not. Old policies, overlapping wikis and tickets that disagree are the normal starting point. Part of the build is deciding which sources to trust and flagging where they disagree. No system can fix knowledge that was never written down. So you get a ranked list of questions no document answers.

A knowledge assistant starts at $1,500 to build and $350 a month. That covers one knowledge set of up to 1,000 documents from 2 sources, answers with sources, up to 25 users and 2,000 questions a month. Larger document sets, more access groups or more places to use it are quoted from your brief, step by step after the first one proves itself. Model use above the included questions is billed at cost.

Something we build over your own documents and hand over, with a runbook for your team. One search system serves site search, a helpdesk lookup and internal assistants, and it respects your access rules. Your materials remain yours, and bespoke deliverable rights, licences and handover are agreed before work starts.

The range

What it costs,
at the direct rate.

Priced as a build that searches one set of your documents for one use, plus a monthly upkeep retainer. Documents loaded only once are out of date by month two.

  • ProjectRAG build$1,500–$18,750fixed quote, one-time
  • RetainerMonthly maintenance$350–$1,875per month

Indicative: every figure is refined into an itemised proposal.INR figures shown excl. GST.

PricingSee indicative pricing for this serviceDirect and wholesale, in ₹ and $ →

Put your documentation
to work.

Scope a pilot: one body of documentation, one surface, a working knowledge system with citations on every answer, your permissions respected, and evaluation before anyone relies on it.

Get a ProposalGet a Free AI Automation Review

Reply within four business hours (typical) · NDA standard · Clear IP & Ownership Terms

Rankings on this topic

Scored on public evidence by the PixelCrayons Index. PixelCrayons is scored like every firm: ranked where it has a case study, listed where it has not.

Last updated