Find related resources using vector representations alongside graph knowledge.
Vector representations provide a way to compare resources by similarity. Use them to find related material even when the resources do not share an exact phrase.
Keep the representation connected to the resource and the workspace context it describes. A similarity score is a retrieval aid, not a statement that two resources mean the same thing.
For example, find reports resembling a selected incident, then use graph relationships to restrict the results to the same equipment family. Similarity suggests candidates; structured knowledge supplies the boundary.
Choose the representation and comparison appropriate to the task. Inspect the original resources before relying on their apparent similarity.
Full-text search matches words. Deductive AI derives conclusions from explicit rules. They complement similarity search without turning its scores into proof.
Developer reference
An embedding binds resource ID, input version, model ID, dimensionality and metric. k-nearest-neighbor execution compares compatible vectors; cosine similarity across unrelated model spaces has no defined retrieval meaning.
Candidate resource IDs and scores join with graph patterns under the query's workspace/access context. Semantic similarity supplies candidates, not deductive entailment or an access grant.
Embeddings and vector indexes are derived state. Source edits or model changes require freshness handling; dimensions alone cannot establish model compatibility. Persist model/input provenance so a result can be reproduced against the intended representation. Canonical graph facts retain their independent identity and version history.