How Matrx vector data stores work

Write path
  1. 1Upload a file (PDF, DOCX, MD)
  2. 2Extract raw text from each page
  3. 3Clean it into structured markdown
  4. 4Chunk into semantic windows
  5. 5Embed each chunk into a vector
  6. 6Index into the data store
Read path
  1. 1User asks a question
  2. 2Classifier decides if Knowledge is helpful
  3. 3Agent calls the search tool
  4. 4Query is turned into a vector
  5. 5Closest chunks come back
  6. 6Agent answers grounded in them
The data store
  1. 1A scoped, named vector index
  2. 2Multiple documents per store
  3. 3Multiple stores per agent
  4. 4Hybrid (vector + lexical) ranking
  5. 5Optional reranker for precision
  6. 6Filtered by metadata at query time