Context Graph

The Context Graph is a native workspace knowledge graph — it complements RAG with structural context that embeddings alone can’t capture (who calls whom, what depends on what, which concepts relate to each other).

How it’s built

1. AST parsing via tree-sitter, with grammars for Python, JavaScript, TypeScript, Go, Rust, Java, C, C++, and JSON.
2. Semantic extraction by LLM on top of the parsing result — identifies concepts, relationships, and the relative importance of nodes.
3. Community and “god node” detection — excessively coupled files/symbols, which tend to be risk points during refactors.
4. Suggested questions — the graph suggests what’s worth asking about the workspace, based on its topology.

Indexing modes

Configurable in the workbench’s Context Graph tab:

• By file type — code, documents, papers — you choose what enters the graph (e.g., markdown only, leaving code to pure RAG).
• Semantic vs. AST — AST is faster and structural; semantic leans more on the LLM to capture relationships syntax alone doesn’t show.

Pausable build

Building the graph for a large workspace consumes LLM calls. The build is pausable and resumable by quota — if you hit a rate/cost limit mid-process, it picks up where it left off instead of starting over.

Where artifacts live

.vectora/context-graph/ inside the workspace itself — the same files .vectoraignore hides from the rest of Vectora also stay out of the graph.

Visual stages (workbench)

While the graph builds, each file shows a stage indicator:

See also

• Hybrid RAG — similarity retrieval, complementary to the graph
• Using the workbench — the Context Graph tab in detail