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