GoldenrozBook a call

A RAG chatbot that answers from your documents and shows where each answer came from

A RAG chatbot answers questions by first searching your own documents and then having a language model write a reply from the passages it found, with a citation to each source. Goldenroz builds RAG and knowledge base chatbots over SharePoint, Google Drive, wikis and help centers, with per-user permissions, citations and a test set that measures answer quality. Projects start at $25,000, fixed price per phase after a free one-hour scoping call.

From $25,000, fixed price per phase

Updated 2026-11-14. By Shai Goldenberg

What usually goes wrong

  • The answer exists and nobody can find it
    The policy is in a PDF, in a folder, in a SharePoint site two reorganizations old. People ask a colleague instead.
  • HR and IT answer the same questions all week
    Parental leave, expense limits, VPN setup, how to request a laptop. Each is documented and still arrives as a message to a person.
  • A demo that quotes the wrong version
    Someone loaded a folder into an AI tool. It answered confidently from the 2023 handbook because the old file was never removed.
  • Search that ignores who is asking
    A single shared index returns salary bands and board minutes to anyone who types the right words. The pilot stops when someone notices.

What is included

  • Content audit
    A review of the sources you want covered: which are current, which contradict each other, which have an owner.
  • Ingestion pipeline
    Connectors that read documents from SharePoint, Google Drive, a wiki or a help center, extract text from PDFs and Office files, and keep the source link and last-modified date.
  • Chunking and search
    Documents split along their own headings and sections, indexed for both meaning and exact terms, so a question about "form W-4" finds the form and not a paragraph that sounds similar.
  • Permissions per user
    The access rules of the source system travel with each passage. Search runs under the identity of the person asking and returns only what they could open themselves.
  • Citations and refusal
    Every answer links to the passages used. When nothing relevant is found, the bot says it has no source.

How a knowledge base chatbot stays accurate

Most failures of a knowledge base chatbot are content failures. The model wrote a fair summary of a document that was out of date, or of two documents that disagree. So the work starts with the content audit and with an owner for each source.

The second control is the test set. Before the build we collect a few dozen real questions and the answers your experts accept. Every change to chunking, search or prompts is scored against that set, on two things: did search find the right passage, and did the answer match it.

The third is freshness. A deleted file must leave the index, and an edited one must replace its old version. We state our evidence plainly: Shai, our CTO, built a support assistant that answers from a maintained knowledge base inside a live platform. We do not publish accuracy figures, because they only mean something on your own content.

An HR chatbot or internal chatbot for policy questions

An HR chatbot is the most common first project because the questions repeat and the documents exist. An internal chatbot of this kind usually lives where staff already are, such as Teams or Slack, covered on the Teams and Slack bot page.

HR content needs two limits. The bot answers what the policy says and does not decide an individual case: it can explain the leave policy and cannot tell someone whether their leave is approved. And questions about a person's own pay or record need a lookup in the HR system under that person's identity, which is an integration and not document search.

When RAG is the wrong tool

RAG suits questions answered by text. It is a poor fit for questions answered by data. "How many open orders does this customer have" needs a query against the order system. For that, connect the system itself through MCP integration.

It is also unnecessary for a small, stable set of documents. Some teams start with a custom GPT for business use inside their ChatGPT workspace, and that is a reasonable first step when the files are few and everyone may see all of them. The build begins to pay when you need permissions, citations, many sources or the bot inside your own tools.

How the work runs

  1. Scoping call

    Before the project

    We review real transcripts, pick the intents worth automating and agree what the bot must never do. You get a fixed price.

  2. Prototype

    About 1 to 2 weeks

    A working bot on the top few intents, tested against your past conversations in a private channel.

  3. Production build

    Roughly 4 to 8 weeks

    System integrations, fallbacks, handoff, logging and channel setup. The range depends on integrations and on WhatsApp account approval.

  4. Limited launch

    At launch

    The bot handles a share of conversations or selected hours first. Your team reviews transcripts daily.

  5. Widen

    After launch

    New intents are added in order of volume, each tested against the conversation set.

What it costs

Each phase has a fixed price agreed before work starts. Model usage and WhatsApp messaging are billed by the providers to your own accounts.

See full pricing

  • Support or ordering bot
    Fixed price per phase
    From $25,000
  • Multi-channel or multi-brand platforms
    Scoped per project
    $100,000 and up
  • Scoping call
    Ends with a written scope and a price
    $0
  • Model and WhatsApp usage
    Billed to your own accounts
    Paid to the providers

Who it is not for

  • Questions that need live figures from a database. That is a system integration, not document search.
  • Legal, medical or financial advice delivered to people with no expert review.

Who does the work

CTO and technology partner

Shai leads the technology side: business software, SaaS and mobile apps, payments and escrow systems, Postgres security, AI agents built with Claude Code and MCP, and IT and infrastructure projects with senior infrastructure engineers.

Questions buyers ask

What is a RAG chatbot?

RAG stands for retrieval-augmented generation. The chatbot searches your documents for passages relevant to the question, then a language model writes an answer from those passages and cites them. The model is not retrained on your data, so updating a document updates the answers.

How much does RAG chatbot development cost?

Projects start at $25,000 and are fixed price per phase after a free one-hour scoping call. The price depends on the number of sources, the permission model and where the bot is used.

Can each employee see only the documents they are allowed to?

Yes. Permissions from the source system are stored with each passage and checked at search time under the user's identity.

Does our data train the AI model?

No retraining is involved in RAG. Passages are sent to the model as context for a single answer. Whether a provider may retain or use that text is set by your agreement with the provider, and we document the data flow for your review.

Which document sources can you connect?

Sources with an API or export can be connected, including SharePoint, Google Drive, common wikis and help centers.

Guides on this topic

Talk to the person who will do the work

One hour, no sales team. You talk to Shai Goldenberg, the CTO, about your systems or your product, and you leave with a clear next step and a price range.

Or send the details

Book a 1-hour consultationWhatsApp us