Confidence tags
Graphify glossary
Graphify's provenance labels — EXTRACTED, INFERRED, AMBIGUOUS — attached to every edge so you always know what was found versus what was guessed.
Confidence tags are Graphify's provenance labels, attached to every edge in the graph. They answer a question most retrieval systems can't: how do we know this? Not how relevant it scored — how the relationship was actually established.
There are three tags. EXTRACTED means the edge came straight out of the parsed AST: tree-sitter found the call, the import, the class definition. It's deterministic, it runs locally, and it doesn't hallucinate. INFERRED means your model connected the dots — a doc page linked to the service it describes, a Terraform resource matched to the code that consumes it. These are usually right, but they're judgment calls, and they're labeled as judgment calls. AMBIGUOUS means Graphify found evidence it couldn't fully resolve — dynamic dispatch, imports built from strings, reflection. Rather than pick a guess and present it as fact, it keeps the edge and tells you it isn't sure.
The design principle is that trust comes from visible seams, not flattened confidence. When an assistant tells you the payment worker depends on the sessions table, you should be able to see at a glance whether that came from the parser or from an inference, and lean on it accordingly. A cosine similarity score of 0.83 can't tell you this — it measures resemblance, not provenance. Tags travel with every answer: graphify explain shows how each connection on a node was established, and GRAPH_REPORT.md separates the architecture the parser extracted from the connections the model proposed.
Related terms
- Edge — A typed, directed relationship between two nodes — calls, imports, defines, references — each carrying a provenance tag.
- Tree-sitter / AST — An AST is the parsed structure of source code; tree-sitter is the open-source parser library Graphify runs entirely locally to extract it.
- RAG (retrieval-augmented generation) — A pipeline that chunks documents, embeds them as vectors, and retrieves the most similar chunks at query time — lookup by resemblance, not structure.