AI & automationshipped
Telusur
Internal knowledge assistant (RAG) that answers from company documents with citations, only from documents the asker is allowed to read, and refuses when no source exists. English and Indonesian.
- Timeline
- 5 milestones
- Role
- Full Stack & AI Engineer
- Team
- 1 person
- Languages EN · ID
- Demo 2:54
- Answers Cited

Overview
Telusur is a multi-tenant RAG assistant for company knowledge. Staff ask a question, get a streamed answer with citations, and a click on a citation opens the source PDF at the quoted page. Access is enforced per collection and per document, so an answer never draws on something the asker cannot read.
The Challenge
- Most RAG demos stop at "it answers".
- A real deployment needs permission-aware retrieval, answers that admit when nothing supports them, an audit trail, and evidence that quality is measured instead of assumed.
The Solution
- Ingestion handles PDF, DOCX and Markdown with structure-aware chunking and document versioning.
- Retrieval is hybrid (pgvector HNSW plus full-text, fused with RRF, then reranked) with local multilingual embeddings.
- Tenancy uses Postgres row-level security, and an access-check simulator shows exactly why a user can or cannot see a document.
- Unanswered questions are grouped by topic so admins can fill knowledge gaps, every action lands in an append-only audit log, and an evaluation harness reports measured results.
