Real numbers from the index and the query log. No projections.

Targets for the knowledge engine, scored live from the metrics above. Green on track, amber behind, red at risk.

Cumulative chunks embedded as documents are ingested. This is the library growing.

Root to categories to documents. Each question asked appears as a node linked to the documents it retrieved, so files that answer the same questions cluster together. New questions appear here in real time. Hover any node.

The system learns from usage: it remembers grounded answers (served instantly next time) and learns query expansions from weak questions so vague queries hit the right clause. Model weights are unchanged; the retrieval around them improves.

A fixed gold set of questions scored with learning OFF vs ON. If the learned bars beat the baseline, the system is genuinely improving retrieval, not just staying busy. Re-run to track the trend.

A clean, quality-gated export built from the classified corpus and grounded answers: taxonomy, per-document records, category-tagged passages, and grounded QA pairs. Junk test files and duplicate passages are dropped. Ready for future fine-tuning.

A timestamped record of every change to the knowledge base: documents learned, categories and subcategories created, corpus growth. Grounded only in the documents. Filter by period.

How one document flows through the pipeline: source file, to markdown, to chunks, to embeddings, to the vector index. Pick a document.

Map-reduce over the whole library on the local model: each document is summarised, then combined into one overview. This is the feature the chat could not do.

Live status of the retrieval services and the model.

What the build agents constructed. Historical, not live activity.