From Messy Files to Living Knowledge

AI knowledge systems are moving beyond “search over documents.” The harder ambition is to turn PDFs, spreadsheets, databases, code, and web pages into a shared substrate that can be queried, updated, audited, and trusted. Recent systems increasingly combine document parsing, schema metadata, retrieval, structured extraction, graph construction, tool use, and evaluation rather than treating knowledge as a pile of text chunks.

That shift matters because the most damaging failures often occur before generation. A PDF parser can destroy table structure; a schema mapper can confuse similar fields; an entity resolver can merge two people or split one organization; a retrieval system can surface a stale policy while presenting it as current. Better language models help, but they do not eliminate these data and systems problems.

The emerging designs also disagree about what the “knowledge base” should be. Some favor contextualized text and hybrid retrieval; others build explicit graphs, ontologies, or executable metadata layers. Some preserve uncertainty and conflicting claims, while others normalize toward a single operational representation. The central question is not whether AI can extract knowledge, but how much structure, evidence, and revision machinery a system needs before its answers deserve trust.

The five sessions follow that transformation from ingestion to maintenance: preserving structure, resolving meaning, choosing representations, tracing claims through transformations, and evaluating whether the resulting system knows what it does not know.