AI and data
RAG and company knowledge bases that answer with a source
Documents, contracts, manuals, tickets and spreadsheets hold the answers; nobody can find them. I turn them into a knowledge base you can search by meaning and plug AI into. Every answer shows which document it came from.
Who it is for: For companies with large volumes of documents or technical material (manuals, policies, contracts, catalogues) where the team or customers wait too long for answers. I also build this for myself: Leadora is exactly a RAG knowledge base for technical teams.
- From
- CZK 25,000
- Typical timeline
- 4–6 weeks
- You get
- A knowledge base, an API, quality measurement
- Updates
- New data without retraining a model
01
This is probably what hurts.
- Answering a technical question means digging through several folders and asking colleagues.
- Documents change and nobody knows which version is current.
- Full-text search finds nothing because people use different words from the document.
- You uploaded documents to a general AI tool and the answers had no source or were distorted.
- You need to expose data to an app or team via a reliable API, not a spreadsheet.
If you nodded at least twice, it is worth a conversation.
02
What we actually solve.
Typical cases clients bring. Yours may look different; we always start from what you need.
Search over documentation
Ask in plain language and get an answer plus a link to the exact page or paragraph. Search works by meaning, not just matching words.
Technical support and service
A field technician asks for a procedure or parameter and gets an answer from the current manual, including the document version.
A customer portal with AI
Customers search product documentation and guides in their own words, so support handles fewer repeated questions.
Connecting data to an application
I process data from several sources, store it cleanly in a database and expose an API that your website, app or agent draws on.
Keeping content current
New and changed documents are indexed automatically. Old versions are marked or retired so answers never rest on outdated material.
03
How it works, step by step.
For each step you see what I do, what I need from you and what you hold afterwards. No black box.
- 01
Data audit
Week 1I do
I review sources, formats and quality. I determine what can be processed automatically, what needs cleaning and what does not belong in the base.
You do
Provide the documents and say which are authoritative and which are outdated.
You get
A source inventory and a processing plan.
- 02
Preparation and indexing
Week 1–3I do
I split documents into meaningful passages, add metadata (version, department, validity) and index them for both semantic and classic search.
You do
Answer questions about unclear documents.
You get
An indexed knowledge base.
- 03
Search and answers
Week 2–4I do
I build the RAG pipeline: retrieving relevant passages, composing the answer and citing the source. I tune ranking and access control.
You do
Supply sample questions to verify against.
You get
Working search and cited answers.
- 04
Quality measurement
Week 4–5I do
On sample questions I measure whether the system finds the right source and answers faithfully. I tune where it fails and set its behaviour when evidence is missing.
You do
Judge which answers are usable.
You get
A quality report and a defined “I do not know” behaviour.
- 05
API and deployment
Week 5–6I do
I expose an API or interface, deploy into your environment and set up updates and monitoring. I hand over documentation.
You do
Name the owner of the data sources.
You get
A production solution, API and documentation.
04
What it can look like in practice.
Model situation: a manufacturer with seven hundred pages of manuals
The service team of a technical-equipment maker rummages through manuals in several versions and PDFs for every question. A new technician takes months to get up to speed.
Before
- An answer means searching several files.
- It is unclear which manual version applies.
- Experienced staff are constantly interrupted.
Process
- Manuals are split and indexed, including version.
- The technician asks in plain language.
- The system finds the relevant passages and answers.
- It shows the document, page and version.
After
- An answer is found in seconds.
- Every answer has a traceable source and version.
- Experienced staff can get on with their work.
An illustrative scenario, not a specific client and not a promised result. We measure real quality on your documents and questions in the Quality measurement phase.
05
What I build with, and why.
- Vector database and full-text
- Combining semantic and classic search beats either alone. PostgreSQL with pgvector, or a specialised store depending on volume.
- Embeddings and language models
- Chosen by language (Czech), cost and data-location requirements.
- Document processing
- PDF, DOCX, tables, websites and tickets, including scans and tables where extraction is rarely trivial.
- Backend and API
- TypeScript or Python. A clean API your website, app or AI agent can use.
- Evaluation
- A test set of questions and measured correctness. Changes are evaluated, not guessed.
What you get at the end.
- Data inventory and processing plan
- An indexed knowledge base
- A RAG pipeline with source citations
- Access control over sources
- An API and optionally a simple search interface
- Quality report, documentation and handover
06
What it costs and how long it takes.
from CZK 25,000
Typical timeline: 4–6 weeks
Database hosting and model usage are paid directly to the vendors. You get an estimate of running costs before we start.
Full price listWhat moves the price
- Volume and format of documents
- Data quality and the need for cleaning
- Access control (who may see what)
- The interface you need: API, web, agent
When it does not make sense.
I would rather say so now than have you pay needlessly.
- There are only a few dozen pages. Often a well-structured FAQ or an upload to a ready-made tool is enough.
- Sources contradict each other and nobody decides what is valid. First set responsibility for content.
- You need answers about events that are not in the documents. The system answers only from material, which is by design.
07
What people ask most.
What is RAG and why not just use ChatGPT?
RAG (retrieval-augmented generation) means AI first retrieves relevant passages from your documents and only then composes an answer from them, so each answer rests on a source you can open. A general tool offers limited access control, updating and quality verification, and cannot reliably connect to your systems.
What if the system finds no answer?
We design the behaviour for missing evidence. The assistant should admit uncertainty and offer a next step, such as forwarding the question to a person, instead of inventing an answer.
Do I need to retrain a model when documents change?
No. New and changed documents are indexed and used immediately. Not having to retrain the model is one of the main advantages of RAG.
How do you handle who may see what?
Permissions are applied at search time: a user gets answers only from documents they can access. We design the rules around your departments or roles.
Does it handle Czech and specialist terms?
Yes. I choose models that suit Czech and verify quality on your real questions, including terminology, abbreviations and internal names.
Where are my documents stored?
By agreement: with you, in a European cloud or with a provider you choose. Before we start we define where data is stored and who processes it.
Can this become an AI agent?
Yes. A knowledge base is often the foundation of an agent. We attach an assistant that answers and, if needed, performs steps in other systems.
08
Let’s talk it through.
Write a few lines about what you are dealing with. I will reply within 24 hours with a view on whether and how it makes sense, before you commit to anything.