Skip to main content

GraphLocalSearchDeps

@forge/monorepo

@forge/monorepo / backend/src / GraphLocalSearchDeps

Type Alias: GraphLocalSearchDeps

GraphLocalSearchDeps = object

Defined in: backend/src/knowledge/graph-retrieval.ts:133

Properties​

depth?​

readonly optional depth?: number

Defined in: backend/src/knowledge/graph-retrieval.ts:136

embeddings?​

readonly optional embeddings?: EmbeddingProvider

Defined in: backend/src/knowledge/graph-retrieval.ts:154

Ranks the chunks the traversal gathered — #277.

Optional, and the mode is materially worse without it. Measured on this repository's corpus, connectivity-only ranking scored 20.8% success@5 against semantic search's 75%. The gap between its own columns was the tell: it found the right document 45.8% of the time and the right chunk half as often again, because an entity's provenance is every chunk that mentioned it and counting entity touches cannot tell two chunks of one document apart.

So the graph is used for selection and embeddings for ranking, which is what Microsoft's local search does and what ours did not.

Left optional rather than required because a deployment that has a graph and no embedder still gets a working traversal, and docs/29 says plainly what it costs. rankedBy in the result reports which happened, so a measurement can never silently be of the other one.

graph​

readonly graph: GraphStore

Defined in: backend/src/knowledge/graph-retrieval.ts:134

knowledge​

readonly knowledge: KnowledgeStore

Defined in: backend/src/knowledge/graph-retrieval.ts:135

neighbourLimit?​

readonly optional neighbourLimit?: number

Defined in: backend/src/knowledge/graph-retrieval.ts:137

poolFactor?​

readonly optional poolFactor?: number

Defined in: backend/src/knowledge/graph-retrieval.ts:162

How many candidates to gather before ranking, as a multiple of the caller's limit.

Ranking can only reorder what selection handed it, so a pool of exactly limit makes the re-rank a no-op — which is precisely the bug: the old code stopped gathering at limit, so the top-k by connectivity were the only chunks that could ever be returned.