What Does It Take to Build a RAG System? Three SARTECH Labs Projects

The first was client work. A thermal power company came to us with a documentation problem.
We were given technical documents for thermal power plants across several sites. Thousands of pages covering boilers, turbines, pumps, mills and fans, right down to individual parts. These documents were stored separately. Checking one against another had to be done by hand. That was slow. And easy to get wrong.
Engineers also needed answers they could trust. The exact number. The exact unit. The page it came from.
So we built a chatbot that could do that. If someone asks about a specific part, the chatbot finds the right section and gives the exact figure, along with where it came from. It also knows that "mill" and "pulverizer" mean the same thing, or that "BFP" stands for boiler feed pump.
For search, we used two methods together. Semantic search finds documents by meaning. Keyword search finds exact terms. On their own, each one misses things. Together, they don't.

The second project was a different kind of challenge altogether, the NCIIPC Startup India AI Grand Challenge.
During training, the SARTECH Labs team was given test documents in PDF format, about 24,000 of them. On the challenge day, it was close to ten lakh. Every document had to be chunked, embedded, and stored in a vector database before a single question could be answered.
The user uploads a set of questions in JSON format. For each question, the system retrieves the top 5 matching documents from the database and writes an answer using local LLMs running on the team's own PC, completely offline, no internet.
Scoring came down to one thing: accuracy. The SARTECH Labs team was one of the top 6 finalists nationwide in Stage 1.

The third was client work again, in the oil and gas industry.
This is a ship supplies business. They receive a constant stream of product request emails asking for specific items, quantities, and specifications. Matching each request and sending out quotations was costing the team time they needed elsewhere.
Using the client's own data, SARTECH Labs built a RAG system that reads incoming emails, matches them to the right product and specifications, and helps the team send quotations faster. The whole system runs privately on the client's own setup, so their data stays secure and fully in their control. Quotations that used to take hours now go out in a fraction of the time.

These three builds had nothing in common except the core idea.
The thermal power project needed precision above everything else. The challenge needed something else, a system that could handle a huge, unpredictable volume of documents, running with no internet at all. The ship supplies project needed privacy. The client's emails and product data could not leave their setup.
There's no single way to build a RAG system.
Every project starts with the same three questions:
1. What kind of documents are these: technical specs, general PDFs, contracts?
2. How will people actually ask about them: plain questions, technical shorthand, or structured queries?
3. Where does the system need to run: online, or fully offline on local machines?