Case study

AI knowledge assistant (RAG)

AI knowledge assistant (RAG)
The problem

Critical knowledge was scattered across thousands of documents — PDFs, Office files, scans — and impossible to search quickly. Generic chatbots hallucinate and can't be trusted on internal content, with no way to verify where an answer came from.

What we built

An intelligent document-processing and retrieval platform: it ingests and OCRs documents, classifies them and extracts key fields, then chunks, embeds, and indexes everything into a vector store and a knowledge graph. An agentic RAG assistant answers using hybrid search — semantic plus keyword — with cross-encoder reranking, and composes responses that cite their exact sources, declining when nothing relevant is found. Multi-tenant, with an admin to manage repositories and access.

The result

Grounded, source-cited answers over a company's own documents — every response traceable to the passage it came from, with retrieval tuned for relevance instead of guesswork.

PythonFastAPILangGraphpgvectorNeo4jNext.js

Documents run through a worker pipeline: OCR with a vision model, summarization, document-type classification and field extraction, layout-aware chunking, embedding, and entity-and-relationship extraction into a knowledge graph. At query time an agent draws on vector, keyword, and graph search, reranks the candidates with a cross-encoder, diversifies them, and generates an answer with citations back to the source passages — and refuses rather than guesses when it finds nothing relevant. The models are self-hosted, so the client’s documents stay private, and a companion module extends the same retrieval approach to video.

Have a project like this in mind?

Get in touch