PwC India · Dec 2024, Jan 2025 · AI Engineering Intern
RAG Assistant over Internal Knowledge
PwC teams kept re-answering the same policy questions and redoing research another team had already finished, because the knowledge sat in several hundred documents that nobody could search by meaning. I worked on the retrieval-augmented assistant built to fix that, on a team shipping it for internal use.
01What the system did
The pipeline was built in LangChain: document ingestion and chunking, embedding into a vector store, retrieval, and prompt assembly for grounded answers. Two modes matched two different problems. Policy lookup returned an answer with the source passage beside it, so the reader could check it. Prior-work discovery was the more interesting one, because the useful output is not the generated text at all, it is the pointer to a document the team did not know existed.
The interface was Streamlit, which for an internal tool used by consultants was the right call: the UI could change as fast as the feedback arrived, and getting it in front of real users early mattered more than polish.
02Measuring instead of guessing
The part that shaped how I build things now was the evaluation discipline. Every chain was traced in LangSmith, and prompt changes were compared against a fixed question set offline rather than judged by reading a few outputs and feeling good about them.
In late 2024 that was not yet standard practice, and the difference between an answer that seems better and an answer that measurably is better is the whole game once a model is in the loop. I have not built anything with an LLM in it since without setting up evaluation first.
next article
HR Recruitment Platform on Salesforce →