RapidCloud Visio Parser extracts content from Microsoft Visio .vsdx files by reading the Visio XML directly, not by flattening diagrams through a LibreOffice-to-PDF-to-OCR path. For each tab it produces a high-quality cropped image, faithful text with exact IPs, hostnames and VLANs preserved, and a topology graph of nodes and connections. It is a microservice in the RapidCloud document pipeline, so a parsed diagram flows straight into synthesis, knowledge-graph extraction and semantic search.
Stop screenshotting your architecture
Every Visio diagram is a decision recorded as a picture. The old way of ingesting them flattened that picture to a caption and threw the connections away — yielding zero relationships and paraphrased labels.
RapidCloud Visio Parser reads the .vsdx XML itself. Exact IPs, hostnames and VLANs survive intact. It infers the real topology from connector geometry, recovering hundreds of edges where naive extraction finds none. Every tab gets a crisp 220dpi cropped image and its own node-and-edge graph.
Then it plugs into the RapidCloud pipeline — synthesis, knowledge graph, semantic search — so your network and process diagrams become answerable knowledge, not static attachments.
Upload a .vsdx and watch a diagram turn into a queryable graph.
Under the hood▾
Visio Parser treats a .vsdx as what it is: a zip of XML. It parses visio/pages/*.xml per tab rather than rasterising through LibreOffice, so text extraction is faithful — exact IPs, hostnames, VLANs and AD groups, no VLM paraphrase.
Topology is recovered by geometric inference. Visio’s <Connect> glue table is typically empty (authors draw free connectors), so each 1D connector’s Begin/End point is snapped to the smallest containing node — the tiebreak that stops a zone box swallowing both endpoints. On a real network file this recovered 307 edges where the old path produced 0. An optional VLM pass adds arrow direction and edge labels, merged and deduped by endpoint pair with per-edge origin tags (geometric | vlm | both).
Per-tab imaging probes at low DPI, computes the ink bbox with Pillow, then re-renders just that crop at 220dpi via poppler. Pages are matched to tabs by token-Jaccard overlap, since LibreOffice scrambles PDF page order. Artifacts (document.json, graph.json, content.md, per-tab PNG/JSON) land in MinIO and flow into synthesis, KG and rc-indexer over parse-phase dispatch.
