RAG systems: AI that knows your business
Gjethe builds retrieval-augmented generation (RAG) systems that let AI answer questions from your own documents and data, with citations to the source. We handle ingestion from PDFs, wikis, help centres and databases, chunking and embeddings, hybrid search with reranking, access control by user, evaluation of answer quality and a chat or API interface.
- Free first consultation
- Written quote for your project
- Flexible payment milestones
At a glance
- Sources
- PDF, Word, wikis, help centres, databases
- Answers
- With citations to the source
- Access
- Filtered by user permissions
- Interface
- Chat, API or inside your app
Definition
What is RAG systems?
Retrieval-augmented generation (RAG) is a technique in which an AI system first searches a collection of documents for passages relevant to a question, then generates an answer using only those passages, so answers reflect your current information and can cite where they came from.
A general AI model knows a lot about the world and nothing about your contracts, policies, products or procedures. RAG connects the two: it looks things up in your material before it answers. Most of the quality comes from unglamorous work (cleaning documents, splitting them sensibly, retrieving the right passages and measuring the results) and that's where we put most of our effort. We test answers against real questions from your team before anyone relies on the system day to day.
Who is it for?
Organisations with large knowledge bases
Policies, procedures and manuals that employees struggle to search, leading to repeated questions and inconsistent answers.
Support teams
Agents and customers who need accurate answers from product documentation and past tickets.
Legal, compliance and finance teams
Contracts, regulations and reports where answers must point to the exact clause or figure.
Products adding AI search
SaaS and content products that want users to ask questions about their own data.
What's included, and what you keep
What's included
- Source inventory and access requirements
- Document parsing for PDF, Word, HTML and tables
- Cleaning, de-duplication and metadata
- Chunking strategy tuned to your documents
- Embeddings and vector index
- Hybrid keyword and semantic search
- Reranking for precision
- Answer generation with citations
- Permission-aware retrieval
- Automatic re-indexing when content changes
- Evaluation set and quality metrics
- Chat interface or API
What you receive
Handed over at the end. Yours to keep.
- Deployed RAG system with ingestion pipeline
- Chat interface or API for your applications
- Evaluation report: accuracy, grounding, citation quality
- Admin tools for sources and re-indexing
- Architecture and maintenance documentation
How does a RAG system work?
Two pipelines: one prepares your content, the other answers questions with it.
Indexing: runs when your content changes
Your sources
PDFs, docs, help centre, database rows
Chunk & clean
Split into passages, keep titles and metadata
Embed
Turn each passage into a vector
Vector store
Pinecone, Qdrant, pgvector…
Answering: runs on every question
Question
From chat, API or an internal tool
Retrieve
Hybrid search, then rerank the best passages
Generate
The model answers only from those passages
Cited answer
With links back to the source
Where does RAG quality actually come from?
Parsing: scanned PDFs, tables and multi-column layouts lose meaning if extracted naively. We use parsers suited to each format and check the output, because an answer can only be as good as the text it's based on.
Chunking and retrieval: passages need to be large enough to carry context and small enough to be precise. We combine semantic search with keyword search (essential for product codes, names and numbers) and rerank the results so the best passages reach the model.
Evaluation: we build a set of real questions with known answers and measure retrieval and answer quality before launch and after every change. Without that, improvements are guesses.
RAG or fine-tuning?
| RAG | Fine-tuning | |
|---|---|---|
| Teaches the model | Facts from your documents, at question time | A style, format or narrow task |
| Stays current | Yes, update the documents | No, needs retraining |
| Citations | Yes | No |
| Best for | Answering questions from changing information | Consistent output format or specialised behaviour |
How RAG systems works, step by step
Every step has a clear output you can review. Timing is agreed in your quote.
- 1
Source audit
What content exists, where it lives, who may see it and how often it changes.
- 2
Ingestion
Parsers, cleaning and metadata for each source, with automatic updates.
- 3
Retrieval design
Chunking, embeddings, hybrid search and reranking tuned on your content.
- 4
Generation
Answer prompts with citations, refusal rules and output formats.
- 5
Evaluation
Accuracy and grounding measured on a test set built with your experts.
- 6
Integrate and launch
Chat interface, API or embedding in your product, with monitoring.
Tools, and what shapes your quote
Pricing is based on your project and agreed in your quote. These are the things that move it.
| Factor | Why it matters |
|---|---|
| Volume and variety of sources | Many formats and systems mean more parsers and connectors. |
| Document quality | Scans, complex tables and inconsistent formatting need extra processing. |
| Access control | Respecting per-user permissions across sources adds design and testing. |
| Accuracy bar | High-stakes domains need larger evaluation sets and stricter answer rules. |
| Hosting requirements | Self-hosted models and vector stores cost more to set up than managed services. |
| Integration | A standalone chat is simpler than embedding answers inside existing tools. |
| Area | Tools |
|---|---|
| Parsing | Unstructured, LlamaParse, custom parsers, OCR for scans |
| Vector stores | pgvector, Qdrant, Pinecone, Weaviate |
| Embeddings & rerankers | OpenAI, Cohere, Voyage, open-source models |
| Orchestration | LlamaIndex, LangChain, Vercel AI SDK |
| Evaluation | Ragas, custom test sets, Langfuse |
Common mistakes to avoid
What we see go wrong most often, and how a good project avoids it.
Indexing everything as-is
Outdated, duplicated and contradictory documents produce outdated, contradictory answers. Curate first.
Vector search only
Semantic search misses exact terms like product codes and names. Combine it with keyword search.
Ignoring permissions
If retrieval doesn't respect access rights, the AI can reveal documents a user shouldn't see.
No evaluation set
Without test questions, nobody knows if the system is good, or if the last change broke it.
Frequently asked questions
What is RAG in simple terms?
RAG means the AI looks things up before answering. Instead of relying on what it learned in training, it searches your documents, reads the relevant passages and answers from them.
Is RAG better than fine-tuning?
For answering questions from your content, usually yes. RAG stays current when documents change and can cite sources. Fine-tuning is better for teaching a model a style or a narrow task.
Can it respect who is allowed to see what?
Yes. We attach permissions to documents and filter retrieval by user, so people only get answers from content they're allowed to access.
Can the data stay on our own servers?
Yes. We can run open-source embedding and language models on your infrastructure so documents never leave it.
Does it work with scanned PDFs and tables?
Yes, with OCR and table-aware parsing. These formats need extra care, which is why we check parsing quality before indexing.
How do updates to documents get in?
Sources are re-indexed automatically when they change or on a schedule, and removed documents are removed from the index too.
Can it answer in a different language from the documents?
Yes. Multilingual embeddings and models can retrieve German documents for an English question and answer in English, for example.
How accurate will it be?
It depends on your content and questions. We measure accuracy on a test set before launch and agree what level is acceptable for your use case.
Can we use it inside Slack or Teams?
Yes. The same RAG system can power a web chat, a Slack or Teams bot and an API for your own apps.
What does it cost to run?
Hosting for the index plus model usage per question. We estimate costs from document volume and expected questions.
Can answers include tables or step-by-step instructions?
Yes. Answers can be formatted as lists, tables or steps, and can quote the exact passage when wording matters, such as a contract clause.
What if the documents contradict each other?
The system can surface both sources and say they disagree, rather than picking one silently. Contradictions found this way are useful signals for cleaning up your documentation.
Can we see which questions were asked?
Yes. Questions, retrieved sources and answers are logged, with personal data minimised, so you can spot gaps and improve content.
Related services
AI Chatbot Development
Chatbots that answer from your own content.
AI Agent Development
AI that takes actions in your tools, with guardrails.
AI API Integration
Language, vision and speech models in your software.
Data Scraping & Extraction
Structured data from websites and documents.
Custom API Development
Documented REST, GraphQL and WebSocket APIs.
RAG Systems by industry
Build your RAG system.
Tell us what you need. You get an honest assessment, a clear plan and a written quote. The first consultation is always free.
- What happens next
- 1You tell us what you need
- 2We aim to reply within 24 hours with questions
- 3You get a written quote for your project