A document-grounded RAG system at OeNB
A useful RAG system should make it obvious what supports an answer—and know when the evidence is not enough.
Explore the full pipeline.
A document-grounded AI workflow that treats retrieval, provenance, evaluation, and refusal as parts of one system.
Retrieve
Compare the question with the indexed collection, rank candidates, and evaluate retrieval behaviour with recall and NDCG.
- Input
- Question representation
- Output
- Ranked candidate passages
- Cosine similarity
- Recall
- NDCG
Which expenses require a second approval?
Turn mixed-format institutional documents into reliable context for language models while keeping answers inspectable and unsafe or unrelated requests outside the system boundary.
- 01
Built ingestion and preprocessing paths for PDF, TXT, and DOCX sources.
- 02
Connected retrieval over PostgreSQL-backed document representations to Hugging Face and vLLM-based language-model workflows.
- 03
Evaluated retrieval with recall and NDCG, and explored structured prompting, reflection, few-shot examples, and fine-tuning.
- 04
Applied cosine-similarity guardrails and built tooling that described graphs for more accessible document use.
This page is a sanitized mechanism reconstruction. It uses synthetic documents and contains no internal OeNB data, screenshots, production measurements, or confidential implementation details.
- Python
- PostgreSQL
- Hugging Face
- vLLM
- NLTK
- Stanza
- Containers