Working prototype
Small AI-Ready Knowledge System
A document ingestion and retrieval prototype with an evaluation workflow for measuring supporting-evidence retrieval.
01 / Problem and approach
The problem
Convert documents into searchable knowledge and check whether retrieval finds the correct supporting information.
My contribution
Built a Flask application with Docling ingestion, structure-aware chunking, semantic retrieval and an evaluation workflow. The current repository also documents grounded generation with source attribution and abstention.
02 / Evaluation and evidence
- Earlier TF-IDF source-retrieval baseline: 72% Top-1 (18/25) and 88% Top-3 (22/25) over the 25 answerable questions in a 30-question set. These are retrieval results, not general answer accuracy or the current system’s overall score.
03 / Scope and context
Scope and limitations
A small synthetic document collection and authored question set limit generalisation. Later repository evaluations use different protocols and should not be conflated with the earlier 25-question retrieval results.