A retrieval-augmented assistant for documents, databases, images and video

The problem

Organisations sit on knowledge in four incompatible shapes: documents nobody can search, a database only two people can query, images and scans that are invisible to search entirely, and video nobody will ever rewatch. The question a user has usually spans more than one of them.

A general-purpose chatbot cannot answer those questions. It has never seen the data, and when it guesses it does so fluently — which is worse than not answering at all.

---

Approach

Each source type gets its own ingestion path into a common retrieval layer: documents are parsed and chunked, scans and images go through OCR, video through transcription, and database access is handled as structured querying rather than pretending a schema is prose.

Answers are generated only from retrieved context and carry citations back to the source, so a user can verify a claim instead of trusting it. When retrieval returns nothing relevant, the correct output is saying so.

---

What was built

• Multi-format ingestion — documents, spreadsheets, images, video and SQL sources
• OCR pipeline for scanned and image-based content
• Chunking, embedding and vector indexing with metadata filtering
• Retrieval and answer generation with citations to source material
• Conversational interface with session history
• Background processing queue for large uploads and re-indexing