Document RAG
What I built
Markdown ingestion, multiple ways to split documents, metadata, combined keyword and meaning-based search, surrounding context, model answers, citation checks and abstention. Local Ollama and OpenAI-compatible adapters; separate scripted test backend.
Result
Built a document-to-answer pipeline with chunking, hybrid retrieval, source citations and abstention, supporting local and OpenAI-compatible models.
Evaluation setup and selected results
Evaluated with Llama 3.2 3B and Nomic on 420 synthetic documents and 62 questions, including 15 unanswerable questions. The pipeline records answer grades, citation checks and abstention outcomes. Citation-tag validation checks format; semantic support requires separate evaluation.
Compare the evaluation evidence →
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.