Imagine an AI that knows your procedures, contracts and product sheets as well as your best experts. That is exactly what RAG makes possible. It gives AI models such as GPT, Claude, Gemini or Mistral access to your own documents. The AI finds the right information there before every answer, and tells you its source. At Kreante, we build custom RAG systems for SMBs and mid-sized companies. We prepare your content, respect everyone’s access rights and test every answer against your real questions before going live. Your teams get an accurate, verifiable answer in seconds, and the AI becomes a real colleague they can rely on.
265+ projects in 35 countries · 4.9/5 on Clutch · teams in Paris, Madrid, Miami and Lima
Do these situations sound familiar?
An AI assistant that confidently quotes a procedure nobody ever wrote
An answer taken from a 2021 version, when the document was updated last week
An off-the-shelf tool plugged into your SharePoint that finds everything… except the right document
A first in-house test, convincing on ten questions, much less so on a hundred
The fear that an intern, by asking the right question, gets hold of information reserved for management
These problems rarely come from the AI model itself. They come from everything that happens before it answers: the state of the documents, the way they are split up, the quality of the search and access control. That is precisely where a RAG’s reliability is won or lost, and where we focus our work.
RAG stands for Retrieval-Augmented Generation. The approach was described in 2020 by researchers at Facebook AI Research[1]. In practice, four steps follow one another for every question.
The AI therefore no longer relies on its general memory, but on your up-to-date content.
Building a first RAG takes a few hours today. Turning it into a tool your teams trust every day takes real expertise, at every step.
A scanned PDF, a badly converted table or three versions of the same procedure are enough to skew the answers. So we start with your documents, before any technology.
Semantic search finds “leave” when you ask about “holidays”, but sometimes misses a product reference or an article number. Keyword search does the opposite. We combine the two, as the guide from France’s Directorate-General for Enterprise recommends[2], then re-rank the results so that only the truly useful passages are passed to the AI.
A RAG must never become a shortcut to confidential documents. The rights from your current tools are applied at the very moment of the search: a passage the user is not allowed to read is never passed to the AI.
When no document answers the question, it is better to say so than to make something up. Our assistants clearly state when the information does not exist in your content, and can point to the right person.
At Kreante, no RAG goes live without passing its exam. With your teams, we write a set of reference questions, along with the expected answers and sources. The RAG is evaluated against this set, then again after every change, so that its quality holds up over time.
| Situation tested | What we check |
|---|---|
| The answer is in your documents | The assistant finds the right passage, answers correctly and cites the right source |
| The information does not exist | The assistant says so clearly, without making up an answer |
| Two documents contradict each other | The assistant flags the discrepancy, or relies on the most recent version |
| The document is confidential | The information stays invisible to anyone without access |
| The question is vague | The assistant understands the intent, or asks for clarification |
You know, with figures to back it up, how reliable your RAG is before it even launches. And this set of questions belongs to you: it will remain your benchmark from then on.
There is no single way to connect AI to your documents. Depending on your content, your uses and your budget, the best solution is not always the most ambitious, and we steer you towards the one that will genuinely save you time.
| Criterion | Off-the-shelf tool (Microsoft 365 Copilot, ChatGPT Enterprise…) | Custom RAG | Model trained on your data (fine-tuning) |
|---|---|---|---|
| Best for | Simple searches in well-organised office documents | Many business documents, precise access rights and integration with your tools | Adopting a very specific style, format or vocabulary |
| Knowledge updates | Automatic, within the tool’s limits | Automatic, as soon as a document changes | New training required |
| Sources cited | Yes, with a search that can hardly be customised | Yes, with a search tuned to your content | No |
| Points to watch | Few settings available, quality hard to measure | Real preparation and testing work, which we take care of | Costly, with no guarantee of up-to-date answers |
For most companies, custom RAG offers the best balance: precise, up-to-date answers on documents that change every week. Fine-tuning can then complement it, to give the AI your tone or your in-house vocabulary.
Where will your data be hosted? How much will each question asked to the AI cost? We answer these two questions with you from the scoping stage, so that your RAG is both secure and cost-effective. Depending on how sensitive your data is, several options are available:
According to estimates from France’s Directorate-General for Enterprise, processing 1,000 questions with a model such as Mistral Small in the cloud costs about €7.80. That is close to €106 per user per year, at 300 questions a week[2]. Hosting on your own servers reduces this cost further, but requires a hardware investment and more time: about six months to set up, compared with two in the cloud, according to the same guide. We cost each option with you, based on your real volumes.
A RAG brings together a large part of your company’s knowledge. That is what makes it valuable, and also why its security must be flawless. The risks are well known: a booby-trapped document whose hidden text tries to give orders to the AI, a database exposed on the internet, or an assistant created in a few hours with AI and then put online without access control. And hackers now also rely on AI to spot these flaws faster.
At Kreante, security is built in from the design stage, and every RAG is reviewed by our senior developers before going live:
Access rights checked at every search, document by document
Protection against instructions hidden in documents
Data encrypted and hosted according to its sensitivity
Access keys and credentials kept out of the code
Users clearly told they are interacting with an AI, as the EU regulation has required since 2 August 2026[3]
Our AI agents write the code, our senior developers design the architecture and approve every step. Having started with platforms like Bubble, WeWeb or Xano, our teams now work with Cursor, Codex and Claude Code, and master the building blocks of a modern RAG, from PostgreSQL and Supabase to Python and LangGraph.
For REMI AI, a nutrition app for a clinic, every interpretation produced by the AI is checked by the medical team before it reaches the patient. We apply the same principle to your sensitive uses: the AI prepares and a person approves.
The scope is defined in writing before we start, and the price does not move. If new requests come up along the way, they are costed separately, fully transparently.
The code, the indexes, the prepared documents and the set of reference questions are yours.
If a tool your company already has is enough, or if your documents first need a serious tidy-up, we tell you before proposing any development.
Free
30 minutes with an expert
Fixed price, custom quote
Delivered in a few weeks on a targeted scope
Custom quote
Up and running in 1 to 3 months, depending on scope
We choose with you a first use case, the documents involved and the questions the assistant will need to answer.
We prepare your documents and build the search, respecting existing access rights from the start.
We evaluate the prototype on your reference questions. With the results in hand, you decide what comes next.
We put the RAG into service in your tools, with a first team, then roll it out further.
We monitor answer quality, add new documents and re-evaluate after every change.
Challenge
Stack
Performance and payroll management platform for AutoCartel field photographers.
Challenge
The platform ingests property data daily via Zillow scraping, runs AI-driven renovation cost and ARV estimates using OpenAI, and surfaces the highest-potential deals first. Users can review…
Stack
An AI-powered real estate intelligence platform built for a Miami-based investment team. DAVCO AI replaces manual Zillow filtering and spreadsheet-based calculations with an automated map interface that displays estimated profit margins across all active single-family listings in the Miami area.
Challenge
Tango event organizers across Europe were running everything manually — registrations, confirmations, payments, check-in. They needed one purpose-built platform to manage the full event lifecycle,…
Stack
TheFlow is a web platform built for tango event organizers in Europe. Kreante evolved it across multiple phases from a basic registration tool into a complete event management and ticketing product — covering event creation, participant management, waitlists, online payments, QR check-in, automatic invoicing, and an international tango event calendar.
Challenge
Kreante developed a comprehensive platform for Pure Engagement using NoCode technologies. Built with Bubble for the frontend and Xano for the backend, this solution facilitates organizational…
Stack
Pure Engagement needed a platform capable of handling the complexity of organizational transformation, recognizing that success depends not only on structural change but also on people’s psychological and emotional transition.
Challenge
Build an iOS app that turns iPhones plus an audio interface into a complete, professional podcast recording studio, capturing high-quality sound and video across several phones at once and keeping…
Stack
A mobile app that lets creators record, sync, edit, and export podcast audio and video using a dedicated interface and multiple iPhones working together as one studio.
Challenge
Patients leaving clinic checkups with a list of medical recommendations and lab results often had no clear, practical way to turn that information into daily meals, leading to poor adherence and high…
Stack
Kreante built REMI AI, a system that reads a patient's genetic and medical test results and automatically turns them into a personalized nutrition and supplement plan that gradually evolves over several months, with clinic staff reviewing every AI-generated interpretation before it reaches the patient.
Challenge
Stack
In the evolving landscape of web development, choosing the right approach for custom software development is crucial.
Challenge
Stack
In the sustainability-focused retail landscape, product repair and maintenance play a pivotal role in circular economy initiatives. Businesses are increasingly adopting innovative solutions to reduce waste and extend product life cycles.
Challenge
Stack
When Garde-Robe presented their innovative vision to make fashion more social and collaborative, Kreante, a leading LowCode agency, eagerly embraced the challenge.
Challenge
When INSPO approached our LowCode agency with their vision of creating the definitive platform for thought leadership, we knew this project would push the boundaries of what’s possible with…
Stack
In the rapidly evolving landscape of LowCode development, creating sophisticated applications no longer requires months of traditional coding.
Challenge
Stack
Connect with verified construction professionals, find new projects, recruit subcontractors, and grow your business with ShareLeads' trusted B2B platform.
Challenge
The basketball community in Washington D.C. lived on courts and in scattered WhatsApp groups. What was missing: a place to give it visibility, bring more people together, find players at your…
Stack
What we built In 3 months, Kreante delivered all three platforms at once. The mobile app lets you find pickup games within a 10-mile radius, rate players to compete at the right level, track your progress, and meet up to watch games together. Two tiers: a free version and a premium subscription at $39.99/season — proof that the community is willing to pay to play better. For this kind of project, visual identity isn't just a detail — it's what creates the sense of belonging from the very first open. The entire branding was conceived and built by Kreante. The landing page was built to convert: a strong visual world, a clear message, one goal — trigger the download. The e-commerce store pushed the logic even further: unapologetic street style, rooted in DC culture, to rep your colors beyond the court. Launched at a Virginia tournament, HoopSquad is now live across 13 regions in the United States — on the App Store and Google Play.
Challenge
The Meetern team had been connecting students and companies for years, through physical events. Many great encounters and opportunities later, they understood that to create real impact at scale,…
Stack
What we built Kreante designed and developed Meetern: a complete matchmaking platform, with three distinct interfaces for students, companies, and academic institutions. Students build their profile, apply, and track their applications. Companies post opportunities and receive intelligently matched profiles. Academic institutions monitor placements in real time. Across all three: integrated messaging and a matching algorithm that replaces manual introductions. The platform is multilingual, French and English at launch, Dutch on the way. A V1 delivered in 6 months, a V2 launched in March 2026 with extended language support. Today, 50 to 100 active users every day. And a team that no longer runs events, but a platform.
Challenge
Every manager knows a motivated team performs better. But motivation is invisible. And without understanding what truly motivates each person, engagement initiatives remain educated…
Stack
Kreante designed and developed Motives Solutions: a bilingual SaaS platform (FR/EN), from branding to application. Team members answer 3 anonymous questions. The app analyzes responses and generates a detailed report by motivational driver, for each individual, each team, each organization. Two modules: Motives TEAM for team-level insights, Motives ORGA for the full picture. Tracking is longitudinal: engagement evolution over time, visualized. Three questions. A clear picture of what truly moves a team forward. Measuring the invisible, so you can finally act on it.
Challenge
After a decade training thousands of women for tech jobs, Laboratoria saw its graduates stall on the way to leadership: in engineering and product, only 52 women are promoted to management for every…
Stack
Laboratoria+ is the membership platform of Laboratoria, the Latin American organization that has helped more than 14,000 women access quality jobs in tech. Built with Bubble, it brings together live courses and learning circles, 1:1 career mentoring with a network of senior women executives, online and in-person events, and a free plan alongside paid Pro and Plus subscriptions.
Challenge
Stack
Transform your organization with ROC's secure hyperautomation platform. Orchestrate workflows, governance, and access management from a living organizational chart.
Challenge
Class2Class wanted any teacher in the world to find a partner class and run a real project together, but international collaboration in schools depended on personal contacts, scattered tools and…
Stack
Class2Class is a global learning community, based in Denmark, where teachers connect their classrooms to run real-world projects around the UN Sustainable Development Goals. Kreante rebuilt the web platform with Bubble and Xano — teacher profiles, class matching, project creation and collaboration spaces — and delivered the rebranding and landing page. The community now counts more than 25,000 teachers in 190+ countries.
Challenge
In Switzerland as across the world, independent cab drivers face the same reality: ride-hailing platforms that set the rates, impose the rules, and own the client relationship. Olivier Fouvy wanted…
Stack
SmartCab is built on a radical premise: for the end customer, nothing changes. No new app to download, no account to create, no interface to learn. They send a WhatsApp message to their driver, just as they always have. Behind that message, an entire machine kicks into gear: an AI engine processes the request, calculates the route, confirms the booking, and triggers invoicing — automatically. Everything an app does. Without the app. They send a WhatsApp message. The rest happens without them. For drivers, it means freedom regained: customizable pricing, a client relationship that belongs to them, no algorithm to appease to stay visible. For fleet operators, a dispatch module handles ride assignments and real-time driver tracking. Three subscription tiers: Lite, Pro, Premium — built for every profile. What SmartCab proves is that real sophistication doesn't announce itself. It's felt — in a booking that requires zero effort, in a relationship of trust that no one gets to monetize. Available on iOS and Android.
In 30 minutes, an expert reviews your documents with you, the use case to prioritise and the right hosting options. You leave with a first estimate of budget and timeline, with no commitment.
RAG stands for Retrieval-Augmented Generation. The idea is simple: before writing its answer, the AI retrieves the useful information from a chosen set of documents, instead of relying only on what it learned during training.
It greatly reduces them, without eliminating them completely. The guide from France’s Directorate-General for Enterprise points this out: hallucinations remain possible, even if they are less frequent than with a model alone[2]. That is why a serious RAG combines several safeguards: answers limited to the documents retrieved, sources cited systematically, the ability to answer “I don’t know” and regular evaluation on reference questions.
Most of a company’s content can be:
| Type of content | Examples |
|---|---|
| Office documents | PDF, Word, PowerPoint, Excel |
| Storage spaces | SharePoint, Google Drive, file servers |
| Work tools | Confluence, Notion, intranet, ticketing tool |
| Structured data | Databases, CRM, ERP, business software |
| Paper documents | Scans and photos, after text recognition |
Documents with many tables, diagrams and images need more preparation work. We identify them during the assessment.
No, in most cases. That is the whole point of RAG: the model learns nothing from your documents, it consults them when answering. Your content can therefore be updated, added or removed at any time, without retraining.
At Kreante, every project is priced at a fixed price on a scope defined in writing, after a free assessment. The budget depends mainly on five factors:
| What drives the budget | Why |
|---|---|
| The volume and state of your documents | Scanned, duplicated or contradictory content requires more preparation |
| The number of sources to connect | Each tool needs its own connector and update rules |
| Access rights | Fine-grained rules, by department or by person, require more work and testing |
| Hosting | A model installed on your servers costs more to set up than an API |
| The interface | Teams, intranet, business software or an API for your own applications |
On top of this comes the usage cost, which depends on the model chosen and the number of questions asked. We estimate it with you before we start.
A prototype evaluated on a targeted scope is usually ready in a few weeks. Going into production then takes 1 to 3 months, depending on the number of sources, the complexity of access rights and the hosting chosen. Hosting on your own servers lengthens these timelines considerably.
The chatbot is the interface, the window where you ask your question. The RAG is the mechanism that fetches the answer from your documents. A chatbot can work without RAG, with answers prepared in advance, and a RAG can work without a chatbot, for example to automatically suggest an answer to a support ticket. See also our AI chatbot agency, AI customer service agent and AI knowledge management pages.
Beyond references, ask it five questions:
A serious agency should be able to answer each of them precisely from the very first conversation.
Several profiles coexist: agencies specialising in AI, development agencies that have added RAG to their offer, and software vendors that offer a ready-to-configure solution.
| Agency | Locations | Positioning |
|---|---|---|
| Kreante | Paris, Madrid, Miami, Lima | Custom RAG developed by AI agents and approved by senior developers, evaluated on the client’s questions, at a fixed price on a fixed scope. 265+ projects delivered in 35 countries, 4.9/5 on Clutch |
| Tensoria | Toulouse | RAG assistants for SMBs, mid-sized companies and software vendors, with hybrid search, evaluation on reference questions and sovereign hosting options |
| Premier Octet | Paris | Development agency offering custom RAG systems, built on PostgreSQL and pgvector |
| Genee | Lyon and Paris | Enterprise RAG solution with source citations, deployable on a European cloud or on your servers |
| BeTomorrow | — | Custom RAG chatbots, with an in-house tool to evaluate and monitor RAGs |
| Codialis | — | AI agency developing RAG assistants and agents, with human validation and testing before going live |
This list is not exhaustive.
Written by Jorge del Carpio, CEO and co-founder of Kreante. Since its founding in 2020, the agency has delivered 265+ projects in 35 countries.
Published Updated
From the blog
Guides, comparisons and field notes from the projects we deliver.
DevelopmentDiscover how to choose effectively between RAG and fine-tuning for your enterprise AI needs, ensuring rapid results and compliance.
DevelopmentHow to build semantic search with Supabase + pgvector + Claude embeddings in 2026. Includes Pinecone vs Weaviate vs Qdrant comparison, real Kreante client example, and 2026 pricing breakdown.
ToolsPick Lovable to get a hosted, Supabase backed live app fast. Pick Claude Design when engineering teams need code first prototypes and tighter code control.
DevelopmentMost AI automation projects show no measurable result. A Kreante project manager explains the sequence that works, with pricing and a real case study.
DevelopmentPilot AI coaching for one to three teams over 6 to 8 weeks, measure workflow KPIs, and demand embedded deliverables to keep skills in house.
Insights & TipsOutcome first AI playbook for SMBs: use a 5-criteria scoring rubric, run a 4-week before and 4-week after task audit, and pilot in 6–12 weeks to prove ROI.
