docs(router): recommend nomic-embed-text-v1.5 over Longformer

llama.cpp doesn't support Longformer's sliding-window + global
attention pattern — confirmed by grepping convert_hf_to_gguf.py for
LongformerModel (not present; supported encoder archs are Bert,
DistilBert, Roberta, XLMRoberta, NomicBert, JinaBert, ModernBert,
NeoBERT, EuroBert).

For routing the dataset schema is encoder-agnostic; we just need
SOME long-context sentence encoder. nomic-embed-text-v1.5
(NomicBert arch, 8192 native context, GGUF available, already
in gallery/index.yaml) fits the bill and runs on the existing
llama-cpp embedding path.

Updates the model-editor description for router.embedding_model
to surface nomic-embed-text-v1.5 as the default suggestion, with
modernbert-embed-base / jina-embeddings-v3 as alternatives.

Also corrects an inaccurate comment in routing_data.go that
conflated Longformer's context length (4096 tokens) with
embedding dimensionality (768) when justifying the 8MB scanner
buffer.

Assisted-by: claude-code:claude-opus-4-7 [Edit] [Bash]
Signed-off-by: Richard Palethorpe <io@richiejp.com>

Author: richiejp

core/config/meta/registry.go

@@ -416,7 +416,7 @@ func DefaultRegistry() map[string]FieldMetaOverride {
 		"router.embedding_model": {
 			Section:              "other",
 			Label:                "Router Embedding Model",
-			Description:          "Embedding model for the KNN classifier — embeds both probe prompts and candidate exemplars. Required when classifier is 'knn'.",
+			Description:          "Embedding model for the KNN classifier — embeds both probe prompts and candidate exemplars. Required when classifier is 'knn'. A long-context sentence encoder works best; nomic-embed-text-v1.5 (NomicBert, 8192 ctx, GGUF via llama-cpp) is a good default. Other supported encoders include modernbert-embed-base and jina-embeddings-v3.",
 			Component:            "model-select",
 			AutocompleteProvider: ProviderModels,
 			Order:                232,

core/services/routing/router/routing_data.go

@@ -100,9 +100,12 @@ func LoadRoutingDataset(path string) (*RoutingDataset, error) {
 
 	ds := &RoutingDataset{}
 	scanner := bufio.NewScanner(f)
-	// Default scanner buffer maxes at 64KB; routing rows with
-	// inline embeddings can exceed that easily (a 4096-D Longformer
-	// row is ~50KB of float JSON alone). Bump to 8MB per line.
+	// Default scanner buffer maxes at 64KB. A row with an inline
+	// pre-computed embedding can blow past that — a 1024-D
+	// jina-v3 vector serialised as JSON is already ~10KB; some
+	// models output 4096-D. Bump to 8MB per line with the same
+	// 64KB initial allocation, so small rows don't pay the
+	// memory cost up front.
 	scanner.Buffer(make([]byte, 0, 64*1024), 8*1024*1024)
 
 	lineNum := 0