Knowledge with Guaardvark (MCP tools)

Everything here is the guaardvark MCP server. All retrieval is local (Ollama embeddings, hybrid search, cross-encoder reranking on the user's GPU).

Find and read

tool	use
search_knowledge_base	the question in plain words; returns passages with sources. Cite them.
summarize_corpus	what the whole knowledge base covers, before searching blind
list_documents	which documents are indexed and how many passages each has
get_document_outline	sections or pages of one document, in order
read_document_section	the actual text of one section or page, no search
process_file	extract content from a PDF, DOCX, CSV, Excel, image or other file the user names
index_document	add a file in Guaardvark's uploads, outputs or install folder to the knowledge base, so the tools above find it
fetch_url	title, description and main text of one web page; pass query (the user's question) to get the passage about it instead of the top of the page
web_search	web results (titles, snippets, URLs) when the answer is not local; source names the search engine that answered, and no_results says when it found nothing

Pattern: summarize_corpus or list_documents → search_knowledge_base → read the exact section with get_document_outline + read_document_section before quoting.

Remember

• save_memory: a fact, preference or instruction to keep across sessions. Use when the user says "remember", or states a durable preference. One fact per call.
• search_memory: check before assuming; recalled memories are what was true when saved.
• update_memory: correct a saved fact, preference or note in place, by the ID search_memory prints, instead of saving a second version next to it.
• delete_memory: forget one entry by its ID when the user asks. It cannot be undone.
• Over MCP these two change only fact, preference and note entries; lessons and Guaardvark's own entries are managed on the Agent Memory page.

Rules

• Quote the source (document name and section) for any claim that came from retrieval.
• If retrieval returns nothing, say so; do not fill the gap from general knowledge without labelling it.
• index_document adds one file at a time and waits up to wait_seconds (default 45) for the indexer; a long PDF may still be indexing when it answers, and list_documents shows it once done. Whole folders are indexed from the Studio's Library page.