RAG Development: Enterprise Knowledge Search

StepTo is a 15-20 engineer team in Belgrade that builds RAG systems for enterprise knowledge search: ingestion, retrieval, permissions, citations and evaluation.

Reviewed by Igor Gazivoda, Founder & CEO of StepTo · Updated

What Is a Production RAG System?

StepTo, a team of 15-20 engineers in Belgrade, builds retrieval-augmented generation (RAG) systems that let employees ask questions in plain language and get answers drawn from the company's own documents, with source citations and per-user permissions. A production RAG system is a pipeline of ingestion, chunking, embedding, retrieval, re-ranking and grounded generation, plus an evaluation set that tells you whether it is right.

RAG, introduced in the 2020 paper by Lewis and colleagues, combines a pre-trained language model with a non-parametric memory that the model can look things up in. For a company this means the documents stay where they are, can be removed from the index, and are not trained into a model. This page takes the enterprise knowledge-search angle. For the wider LLM and agent offer, see StepTo's AI development page; for a built example, see the RAG support-automation case study.

Before you sign, ask any vendor to show the retrieval, permissions and evaluation work, not only the chat interface. The buyer's guide to RAG development partners lists the questions to ask, and the context engineering article covers how context is structured once agents are involved.

When Should You Not Build a RAG System?

A custom pipeline is not always the right answer. Three cases where a simpler option usually wins:

  • Your need is generic and sits inside a platform you already pay for, such as a help-desk tool or documentation product with a built-in knowledge assistant. A managed feature is usually enough there.
  • Your documents are few and rarely change, and nobody needs permissions, citations or measured quality. A simple search box may serve you better than a retrieval pipeline.
  • You want the model to learn a writing style or a task format rather than answer from your documents. That is a different problem, and retrieval is usually the lighter option for internal questions.

A custom build starts to pay when retrieval quality against your own, changing corpus is part of your product or your risk, and the content lives in several systems that a platform feature cannot reach.

How Is an Enterprise Knowledge Search System Built?

Six parts, each its own set of engineering decisions. A team that is strong in one will often present the whole pipeline as solved.

Ingestion and chunking

Documents from wikis, drives and ticket history are cleaned, split into chunks and tagged with metadata such as source and owner. Chunks that are too small lose the context a passage needs, and chunks that are too large dilute the signal.

Embeddings and the index

Chunks are embedded and stored in pgvector, Pinecone or Qdrant. Whether the index lives in a dedicated vector database or in the Postgres you already run is a real decision, covered in the StepTo guide to vector databases versus pgvector.

Retrieval and re-ranking

Production retrieval is rarely one similarity search. Systems that hold up on real queries typically combine lexical and vector search, then re-rank the combined results before anything reaches the model.

Permission-aware retrieval

An internal assistant searches only the documents the asking person is allowed to open. An index that pools everything into one store can surface restricted HR or board folders to people who should never see them.

Grounded answers with citations

The model answers from retrieved passages and shows its sources, so a reader can open the original page. When retrieval comes back empty, the system says so or hands off to a person instead of guessing.

Evaluation

A test set of real questions, each paired with the passage that answers it, is re-run whenever a prompt, the model or the content changes, so a drop in quality shows up before users see it.

StepTo's published RAG stack uses pgvector, Pinecone or Qdrant for vector search, LangChain or LlamaIndex for orchestration, and OpenAI GPT, Anthropic Claude or self-hosted Llama and Mistral models, chosen on your own questions. Details of the vector-store trade-off are in vector databases versus Postgres with pgvector.

Where Does the Data and the Model Run?

StepTo can deploy retrieval assistants in the customer's own cloud account or against a self-hosted model such as Llama or Mistral, with access control and logging, including for data-residency-sensitive deployments. The region or the self-hosted setup you choose decides where documents, embeddings and prompts are processed. Where StepTo's engineers sit does not.

Logging design is part of the build, and the logs themselves need access control and a retention limit. The private AI guide walks through these choices in more detail.

NDA and IP assignment terms are signed before any work begins.

How Do RAG Systems Fail in Production?

Four failure modes to ask any vendor about, including StepTo, before you sign.

Right reasoning over the wrong context

A model can reason correctly over the wrong passages, so evaluation should check retrieval first. Chunking, embedding choice, hybrid search and re-ranking are where retrieval quality is decided.

Fluent answers the sources do not support

A model can write a confident answer from retrieved context that does not back the claim. Checking that each answer is grounded in the cited passages, and escalating when it is not, is a design decision that has to be made before launch.

A poisoned or badly governed index

Retrieval indices are often filled from wikis and shared drives that many people can edit. The PoisonedRAG study reports that injecting five malicious texts per target question into a knowledge database with millions of texts reached a 90% attack success rate. Write access to the sources therefore matters as much as read access.

A corpus that keeps changing

Source documents change, so parts of the index need reindexing, and retrieval quality drifts as queries and models change. That is recurring engineering work with its own monitoring, not a one-time setup step.

Source for the poisoning figure: Zou, Geng, Wang and Jia, PoisonedRAG (arXiv 2402.07867). The StepTo article on shadow AI and RAG poisoning covers the governance side.

How Can You Engage StepTo on RAG?

StepTo's terms are the same for RAG as for any other engineering work, and they are published on the pricing page.

Add RAG engineers

Staff augmentation starts from $4,500 per developer per month, with a 3-month minimum and weekly billing.

Dedicated team

A dedicated team of three starts from $13,500 per month, billed monthly.

Fixed-scope project

Projects and MVPs start from $9,000, with the scope of a RAG build agreed after the discovery call.

Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks. It starts with a 15-minute intro call. If there is a fit, a 60-minute discovery call follows.

The rate card runs junior $25-35 per hour, mid-level $35-45 per hour, senior $50-65 per hour, and lead $65-85 per hour. For a team already inside your product, a forward deployed engineer can configure and iterate the retrieval system in your own environment.

Not sure which shape fits? The Spec a Team tool turns a role list into a priced estimate on the same rate card.

How Is a StepTo RAG Engagement Run?

RAG work follows the eight engineering practices StepTo applies on every engagement, described in full on the engineering practices page. Keeping prompts, evaluation sets and code in your own repository also keeps the exit simple if you later switch model providers.

  • Peer code review

    Every change is reviewed by a second engineer before it is merged.

  • Automated tests on every pull request

    A CI pipeline runs automated tests on every pull request.

  • Your repository from day one

    Engineers work in your repository and your tools from day one, and you own the code at all times.

  • A weekly demo or written report

    You get a demo or a written status report every week.

  • NDA and IP assignment before the start

    An NDA and an IP assignment are signed before any engineer starts, and all IP goes to you.

  • Managed, secured laptops

    Engineers work on company-managed laptops with disk encryption and two-factor sign-in.

  • Documented handover

    There is a written handover whenever an engineer is replaced or the engagement ends.

  • Senior technical oversight

    A senior StepTo engineer or lead follows each engagement.

How Do Other Vendors Position Enterprise RAG?

Two vendors that publish an enterprise RAG offer, as read on 2026-11-20. Check their pages before you rely on this summary, because both change often.

STX Next

What it reports: Lists enterprise RAG, including a platform it calls DocSpeaker, among its AI services (stxnext.com/services/rag-development-services, as of 2026-11-20).

Where it wins: STX Next reports ISO/IEC 27001 certification (stxnext.com/python-development), membership of the Claude Partner Network Services Track at Select level (stxnext.com/our-partners/anthropic), 450+ specialists (stxnext.com) and named enterprise case studies (stxnext.com/case-study/linde). StepTo cannot claim any of these. A public rate card was not found on the pages read as of 2026-11-20.

Andela

What it reports: Describes enterprise knowledge search with RAG, role-based access control, an audit log and PII redaction as one of its managed AI offerings (andela.com/deploy/managed-ai-solutions, as of 2026-11-20), with a committed cost and timeline after a two-to-four-week Inception, milestone-based billing and a range priced by scope.

Where it wins: Andela reports Select-tier partner status with Anthropic (andela.com/partners/anthropic), named client case studies and 650+ enterprise clients (andela.com). Where a buyer needs vendor partner tiers or a large enterprise footprint, it wins on those points.

StepTo is a 15-20 engineer team without certifications or vendor partner tiers to show, and it does not name clients. What it offers instead is a published rate card, engineers working in your repository from day one, and a team in Belgrade on Central European Time.

What Do Buyers Ask About RAG Development?

Which company in Belgrade builds enterprise RAG and knowledge search systems?

StepTo is a team of 15-20 engineers in Belgrade, Serbia, founded in 2020, that builds retrieval-augmented generation systems over a company's own documents. The work covers ingestion and chunking, embeddings and vector search with re-ranking, per-user permissions, source citations and an evaluation suite. Belgrade works on Central European Time, and StepTo can embed engineers in your repository and tools from day one.

What does StepTo build in a RAG engagement?

StepTo builds the full pipeline rather than a chat window around a model call: document ingestion and chunking, embeddings stored in pgvector, Pinecone or Qdrant, retrieval with re-ranking, answers with source citations, guardrails against prompt injection, and an automated test set of real questions. Where an answer cannot be grounded, the system can hand off to a person. A published example is the RAG support-automation case study on this site.

Where does our data and the model run?

That is your choice. StepTo can deploy a retrieval assistant in your own cloud account or against a self-hosted model such as Llama or Mistral, with access control and logging. Where data lives is decided by the cloud region or self-hosted setup you choose, not by where StepTo's engineers sit. NDA and IP assignment terms are signed before any work begins.

How much does a RAG project cost with StepTo?

StepTo publishes its engineering rates instead of asking for a quote first. The rate card runs $25-85 per hour across junior to lead engineers. Staff augmentation starts from $4,500 per developer per month with a 3-month minimum and weekly billing. A dedicated team of three starts from $13,500 per month, billed monthly. Fixed-scope projects and MVPs start from $9,000, and the scope of a RAG build is agreed after the discovery call.

How does StepTo check that RAG answers are correct?

StepTo keeps a test set of real user questions with known source passages and re-runs it whenever a prompt, the model or the content changes, so a drop in quality is visible. Retrieval is checked first, because an answer built on the wrong passages cannot be right. For high-stakes flows StepTo adds human handoff rather than letting the assistant guess.

Get in touch

Or reach us directly

Office hours

Monday – Friday09:00 – 18:00 CET
WeekendClosed

Sent on a weekend: we reply on the next business day.

Contracting party: StepTo d.o.o. Company details are on the Imprint page.

Send us a message

StepTo's Belgrade team replies within one business day.

It starts with a 15-minute intro call. If there is a fit, a 60-minute discovery call follows. Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks.

* required

Your details are handled as described in our Privacy Policy.

Performance-led engineering

Want senior engineers who move work forward, not just tickets?

Work with accountable, English-fluent professionals who communicate clearly, protect quality, and deliver with a steady operating rhythm. Cost efficiency matters, but performance is why clients stay with us.

Delivery signals · senior engineering team
Senior ownership
Lead-level
Delivery rhythm
Weekly
Timezone overlap
CET
1 teamaccountable for outcomes, communication, and execution