← Work

Selected work

Telecom document RAG

Retrieval-augmented answers over thousands of service and topology documents, served by local models.

Problem

Operational knowledge — service records, topology, tickets — is scattered across systems and exports. Engineers waste time hunting for context that already exists somewhere.

Approach

Documents are ingested and chunked into a vector store with a local embedding model. A GPU-served LLM answers questions with citations back to the source material, and a curated knowledge layer keeps the corpus organised. The LLM is never placed on the critical path of ingestion, so retrieval stays available even when generation is down.

Stack

OnyxOllamapgvectorGPU servingPostgreSQLRAG

Outcomes

  • Natural-language answers grounded in the actual document corpus
  • Cited sources, so every answer is verifiable rather than hallucinated
  • Search across a large internal corpus without manual indexing
  • Fully self-hosted — no internal data leaves the network