Console: Retrieval
The Retrieval pillar is where you run and tune hybrid search. It has three views: Search & eval, Embeddings, and Indexes.
Search & eval
Section titled “Search & eval”An interactive search surface over the selected graph. Enter a natural-language query and see fused hybrid results, with controls for the same knobs the SDK exposes:
- source (
persisted/ephemeral) and consistency (strong/eventual) - targets (entities, assertions, observations, …)
- structured filters and facets
Crucially, it renders the explain breakdown so you can see how a result was
scored — the per-signal contribution across snapshot, ontology, target prep, bm25, vector, graph scoring, and facets. This is where you confirm
the vector leg is using the embedding model you configured (vector_model_id).
The eval side runs accuracy/latency checks so you can compare configurations.
Embeddings
Section titled “Embeddings”Manage the graph’s embedding strategy:
- Choose the active model (e.g.
bge-small-en-v1.5,all-MiniLM-L6-v2,e5-small-v2,Qwen3-Embedding-0.6B) and see its dimensions. - Toggle query auto-embed so search embeds queries with the active model.
- Backfill corpus to embed entities written before the model was enabled, and watch progress to “Backfill complete”.
- Promote a freshly built embedding run so search starts using it.
This is the visual front-end for managed embeddings.
Indexes
Section titled “Indexes”Inspect persisted index coverage per family (BM25, vector/ANN, adjacency). Each family shows coverage chips and a one-line summary of how much of the graph the current runs cover and how long the unindexed WAL tail is. From here you can trigger a build/refresh and watch it complete.
In the hosted product, base indexes build automatically at a doubling cadence and search never 404s — this view makes that otherwise-invisible lifecycle legible, showing index status as a first-class signal.