- RAG
- EMBEDDINGS
- LLM
RAG Document Assistant
A retrieval-augmented generation assistant over a document corpus: indexing, semantic retrieval and answers with source citations. Code at github.com/delcenjo/rag-document-assistant.
Retrieval-augmented generation over a folder of documents: plain-language questions and answers built only from what it found, citing the files they come from. The repo ships with the documentation of a made-up product (Nimbus Notes) so anyone can clone it and get a working pipeline without supplying data.
- Recall@40.88
- Answerscited
- DemoHF Spaces
Retrieval gets its own evaluation
Before judging the generated answer you need to know whether the system finds the right chunks. On a hand-labelled set of questions it measures recall@4: for what fraction of questions does the correct source document appear in the top 4 results. On this corpus: 0.88.
Every answer with its source
Answers cite the file and score of every chunk used (pricing.md at 0.55, security.md at 0.31...). The goal is for every claim to be verifiable in one click, not to sound convincing.
Try it live
There is an interactive demo on HuggingFace Spaces with the sample corpus loaded: you ask a question and see the answer with its citations and scores. It is the fastest way to understand what the pipeline does inside.
RAG Document Assistant
Interactive demo hosted on HuggingFace Spaces
Next projects:
Credit Risk Platform
End-to-end MLOps platform for credit-default scoring: reproducible training with a versioned model registry, a FastAPI inference API with input validation, PSI drift monitoring and a Prometheus and Grafana observability stack, all containerised with CI. Code at github.com/delcenjo/credit-risk-platform.
Transformer from scratch
A GPT-style language model implemented from scratch in PyTorch: multi-head attention, causal masking and residual blocks written by hand, plus a byte-pair tokenizer and an ablation study. Code at github.com/delcenjo/transformer-from-scratch.
Built with
