AI search and answers over your enterprise documents.
RAG Development Services for Enterprises
Supertype is a RAG development company. We build systems that find, cite and extract information from your own documents and data, respect who is allowed to see what, and run in your cloud or on-premise.
Our teams in Singapore and Jakarta have built RAG and document AI for public institutions and enterprises.
What is enterprise RAG?
Retrieval-augmented generation (RAG) connects a language model to your own documents and databases. When someone asks a question, the system retrieves the most relevant passages first, then has the model answer from them, with citations.
Enterprise RAG adds what a business needs on top: access control, evaluation, audit logs and deployment inside your environment. Without retrieval, a general-purpose model cannot see your private data and will guess.
0 vectors
The six stages of a RAG system
- 01
Access and sources
We start by listing your documents and databases, what shape they are in, and who is allowed to see what. Permissions go into retrieval from day one. A filter added at the end tends to leak.
- 02
Parsing
This is where most RAG systems go wrong. A scanned page or a table turns into garbled text, and nothing later in the pipeline can repair it. We use OCR and table extraction, and keep every passage linked to its document and section.
- 03
Chunking and metadata
Small chunks lose context and large ones bury the answer. We split documents so each piece still makes sense on its own, then attach dates, owners and document types for retrieval to use.
- 04
Retrieval
Vector search misses exact terms such as names and clause numbers, so we pair it with keyword search, rerank the results and filter by the user's permissions.
- 05
Generation with citations
The model answers from the retrieved passages only, cites them, and says so when they don't contain the answer.
- 06
Evaluation and monitoring
We write a test set from real questions and score retrieval and answers separately, which tells us which stage failed. After launch we keep an eye on the queries that cost the most, come up the most, run slowest or get the worst answers.
Most hallucinations trace back to retrieval, so we start with the pipeline, not the prompt. Our book The Enterprise AI Agentic Stack covers ETL, vector databases and observability in more depth.
What we build
Enterprise search and Q&A
Staff ask a question in plain language and get an answer with links to the exact passages behind it. The hard part is that people search in two ways: by meaning ("what are our obligations if a supplier defaults?") and by exact term (a contract number, a name). We run both kinds of search together, so neither gets missed.
At BPK, officials and researchers search past audit findings and recommendations without knowing which report to open.
What you get
- Hybrid retrieval with reranking
- Permission filters per user and role
- Citations on every answer, in a UI or over an API
- A test set and quality scores you can rerun
Document extraction
For collections of reports, contracts or filings, a chat box is the wrong tool. What you want is the content turned into data: fields, tables and tags you can filter, count and analyse. Scanned pages and inconsistent layouts are the usual obstacles, so the pipeline checks its own output and flags low-confidence results for a person to review.
BPK's PDF audit reports now sit in a queryable database, each one tagged by content, metadata and location.
What you get
- An extraction schema agreed with your team
- OCR and table handling for scanned and complex PDFs
- Confidence scoring and a review queue
- Loading into the database or warehouse you already use
Knowledge assistants
A chatbot over your wikis, SOPs and databases, for staff or customers. Adoption depends on where it lives, so we put it in the tools people already use, such as Slack, WhatsApp or the intranet, instead of asking them to learn a new one. Every answer names its source, so a person can check it.
The chatbot is one interface on top of the retrieval layer. Our AI chatbot development service builds the two together.
Who can build AI tooling into our business processes?
Supertype. We're the AI developer trusted by Indonesia's largest enterprises.
What you get
- A chatbot connected to your document sources
- Integration with the channels your people use
- Handover to a human when the assistant is unsure
- Logs of questions and answers for review and audit
Agentic RAG
Some questions need more than one lookup. An agentic system breaks the question into steps, retrieves for each, calls tools such as a SQL query or an internal API, and combines the results. It also costs more per question, since each step is another model call. We route simple steps to cheaper models to keep the running cost down.
passages with citations
structured figures
live system data
What you get
- An agent with the tools it needs and nothing more
- Guardrails and approval steps for actions that matter
- Tracing of every step, so failures can be debugged
- Cost per question measured and tuned
Pilot audits
If you already have a RAG pilot that hallucinates, misses documents or costs too much, we review it before anyone rebuilds it. We collect the questions that fail, trace each one to the stage that caused it, and rank the fixes by effort and impact.
Most often the fix is in parsing or retrieval, not in the model.
What you get
- A test set built from your real questions
- Each failure traced to a pipeline stage
- A ranked list of fixes
- A fixed-scope proposal if you want us to do the work
Case study: AI search over 1,500 audit reports
PDF audit reports extracted into a queryable database
of data under AI-powered search
The Audit Board of Indonesia (BPK) audits the management of state finances. Its BIDICS platform holds audit reports as PDF documents: annual, financial, performance, compliance and special audits.
We built the AI layer as part of our LLM development work. LLMs extract each report into a queryable database, tag and categorize it by content, metadata and location, and map it to the UN Sustainable Development Goals so BPK can track progress against them.
Custom RAG or a RAG platform?
Not every company needs a custom build. Compare the options on control, setup time and where the system runs, and weigh them against lifetime cost, not just the first build. Our build, buy or partner guide explains why.
| RAG platform | Cloud knowledge base | Custom RAG system | |
|---|---|---|---|
| Best for | Company-wide search across common workplace apps | Teams already standardized on one cloud | Specialized documents, strict access or deployment rules, or measurable accuracy |
| Setup | Fastest | Fast, within that cloud | Longest, designed around your data |
| Control over retrieval and evaluation | Limited | Partial | Full |
| Deployment | Vendor-hosted | That cloud | Your cloud, private network or on-premise |
RAG or fine-tuning?
Most enterprise use cases start with RAG. We add fine-tuning, which is part of our LLM development services, only when it clearly improves quality, cost or latency.
| RAG | Fine-tuning | |
|---|---|---|
| What it changes | What the model can see | How the model behaves |
| Stays current with your documents | Yes, as the index updates | No, it needs retraining |
| Cites its sources | Yes | No |
| Suited to | Facts, documents, changing data | Format, tone, domain vocabulary |
Frequently Asked Questions
What is RAG in business?
Retrieval-augmented generation (RAG) lets an AI system answer questions from a company's own documents and data. It retrieves the relevant passages first and then generates an answer from them with citations, so the answers reflect current, private information instead of the model's general training data.
What does a RAG project involve?
A typical project covers reviewing data sources and access rules, parsing and indexing the documents, building retrieval and generation with citations, integrating with your systems and users, and evaluating quality before and after launch.
Our work for the Audit Board of Indonesia (BPK) involved 1,500+ PDF audit reports, around 1TB of documents, so parsing and indexing at that scale was a real part of the job.
Is RAG still relevant now that models have long context windows?
Yes, for most enterprise use cases. Long context helps with a few documents at a time, but RAG is still how you search thousands of documents, enforce access permissions, keep answers current and cite sources at a reasonable cost.
Agentic systems often use RAG as one of their tools, which is how it shows up in our AI agents work.
Can we use RAG on sensitive data we cannot send to an external AI provider?
Yes. We can deploy the whole system in your cloud account, private network or on-premise, using open-source models that run inside your infrastructure.
Access follows your existing permissions, so people only get answers from documents they are already allowed to see.
How do you handle long or complex documents like scanned PDFs and tables?
Parsing is where most RAG systems fail, so we treat it as its own step: OCR for scanned pages, table extraction, and metadata that keeps each passage tied to its document and section.
We have done this at scale on 1,500+ audit report PDFs for the Audit Board of Indonesia, where scans, tables and long reports were the normal case.
How do you measure RAG quality?
We build an evaluation set from real questions and score both retrieval (did it find the right passages) and answers (are they correct and faithful to the sources). Every change is tested against it, and production is monitored after launch.
Watching quality at run-time, not only at build-time, is a theme in our book, The Enterprise AI Agentic Stack.
Our RAG pilot hallucinates on internal documents. Can you fix it?
Usually, yes. Hallucination in RAG is most often a retrieval problem: poor parsing, chunking or search. We audit the pipeline, measure where it fails and fix those steps first.
Changing the model is rarely the first fix, because a better model still cannot answer well from passages that were never retrieved.
Can RAG work with our existing data sources and systems?
Yes. We connect RAG systems to document stores, wikis, databases, data warehouses and business applications, and keep the index in sync as content changes.
For companies running Odoo, we have a separate page on AI for Odoo companies.
Do we own the RAG system?
Yes. The pipeline, prompts and evaluation set we build for you are yours, and we can hand them over to your team or keep running them for you.
How much does RAG development cost?
Cost depends on the number and quality of data sources, how complex the access rules are, how much evaluation the use case needs, and where the system must be deployed.
We scope these first and send a fixed proposal. If you are unsure whether to build this in-house, the build vs buy vs partner guide covers the cost trade-offs.
How is this different from your AI chatbot development service?
RAG is the retrieval engine; a chatbot is one interface on top of it.
If you need a conversational assistant, see our AI chatbot development service. If you need search, extraction or a retrieval layer for other systems, this is the right service.
Send us the RAG pilot that is not working
Tell us what your documents look like and what people need to find in them. We will tell you what we would change.
What to send
A sample of the documents, the questions people ask, and the answers you get today.
What you get back
Where retrieval is failing, which stage to fix first, and a fixed-scope proposal.