Turn sheet ranges, queries, tables and folders of PDF, Word and e-mail into a local, searchable knowledge base that remembers where every piece came from and when — and lets Ask Greg cite it. Most of what a desk knows is not in a database. It is in a limits sheet updated every month, a policy PDF from last year, the Word file that replaced it and an e-mail that changed one rule in between. A knowledge base gathers that material into one store you can search by meaning and by exact words, and it never loses track of where a sentence came from: every piece carries its project, its sheet and range or its query, or its document and page, when it was added and by whom, and the business date it describes. Each base is one local DuckDB file — kept in the project folder so it travels with the project, or shared across projects on the machine or server — and its text is turned into vectors by an embedding model the installation configures: an on-premises server such as Ollama or vLLM, or OpenAI, Azure OpenAI or Gemini where the classification allows. Because the same fact often arrives twice, a base also finds contradictions: the same key with different values, found exactly with no model; near-identical statements that disagree, such as one hour in the 2025 policy and four hours in the 2026 one; and plain duplicates. Keep newer retires the old value without deleting it, so the as-of search still shows what was valid on any past date. Ask Greg can search the bases a user selects and cites each piece it uses. Every screen in this chapter comes from the Knowledge-Test project: a Limits sheet with an October and a November block, a policy table and a folder of five documents, one of them a scan with no text.
This site needs JavaScript for the full experience. Every page above is reachable without it.