pr comments

- default to embeddings enabled
- move embedding logic for loaded model to request
- allow embedding full directory
- close llm on reload

server/routes.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/gin-contrib/cors"
 	"github.com/gin-gonic/gin"
+	"gonum.org/v1/gonum/mat"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llama"
@@ -114,7 +115,22 @@ func GenerateHandler(c *gin.Context) {
 
 	checkpointLoaded := time.Now()
 
-	prompt, err := model.Prompt(req)
+	embedding := ""
+	if model.Embeddings != nil && len(model.Embeddings) > 0 {
+		promptEmbed, err := loaded.llm.Embedding(req.Prompt)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+		// TODO: set embed_top from specified parameters in modelfile
+		embed_top := 3
+		topK := vector.TopK(embed_top, mat.NewVecDense(len(promptEmbed), promptEmbed), loaded.Embeddings)
+		for _, e := range topK {
+			embedding = fmt.Sprintf("%s %s", embedding, e.Embedding.Data)
+		}
+	}
+
+	prompt, err := model.Prompt(req, embedding)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return