deprecate modelfile embed command (#759)

cmd/cmd.go

@@ -78,18 +78,12 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				spinner.Stop()
 			}
 			currentDigest = resp.Digest
-			switch {
-			case strings.Contains(resp.Status, "embeddings"):
-				bar = progressbar.Default(resp.Total, resp.Status)
-				bar.Set64(resp.Completed)
-			default:
-				// pulling
-				bar = progressbar.DefaultBytes(
-					resp.Total,
-					resp.Status,
-				)
-				bar.Set64(resp.Completed)
-			}
+			// pulling
+			bar = progressbar.DefaultBytes(
+				resp.Total,
+				resp.Status,
+			)
+			bar.Set64(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
 			bar.Set64(resp.Completed)
 		} else {

docs/modelfile.md

@@ -12,7 +12,6 @@ A model file is the blueprint to create and share models with Ollama.
   - [FROM (Required)](#from-required)
     - [Build from llama2](#build-from-llama2)
     - [Build from a bin file](#build-from-a-bin-file)
-  - [EMBED](#embed)
   - [PARAMETER](#parameter)
     - [Valid Parameters and Values](#valid-parameters-and-values)
   - [TEMPLATE](#template)
@@ -91,17 +90,6 @@ FROM ./ollama-model.bin
 
 This bin file location should be specified as an absolute path or relative to the `Modelfile` location.
 
-### EMBED
-
-The `EMBED` instruction is used to add embeddings of files to a model. This is useful for adding custom data that the model can reference when generating an answer. Note that currently only text files are supported, formatted with each line as one embedding.
-
-```modelfile
-FROM <model name>:<tag>
-EMBED <file path>.txt
-EMBED <different file path>.txt
-EMBED <path to directory>/*.txt
-```
-
 ### PARAMETER
 
 The `PARAMETER` instruction defines a parameter that can be set when the model is run.

go.mod

@@ -45,7 +45,6 @@ require (
 	golang.org/x/sys v0.11.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
-	gonum.org/v1/gonum v0.13.0
 	google.golang.org/protobuf v1.30.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -145,8 +145,6 @@ golang.org/x/text v0.10.0 h1:UpjohKhiEgNc0CSauXmwYftY1+LlaC75SJwh0SgCX58=
 golang.org/x/text v0.10.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
-gonum.org/v1/gonum v0.13.0 h1:a0T3bh+7fhRyqeNbiC3qVHYmkiQgit3wnNan/2c0HMM=
-gonum.org/v1/gonum v0.13.0/go.mod h1:/WPYRckkfWrhWefxyYTfrTtQR0KH4iyHNuzxqXAKyAU=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
 google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=

parser/parser.go

@@ -40,7 +40,7 @@ func Parse(reader io.Reader) ([]Command, error) {
 			command.Args = string(fields[1])
 			// copy command for validation
 			modelCommand = command
-		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "EMBED", "ADAPTER":
+		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "ADAPTER":
 			command.Name = string(bytes.ToLower(fields[0]))
 			command.Args = string(fields[1])
 		case "PARAMETER":
@@ -51,6 +51,8 @@ func Parse(reader io.Reader) ([]Command, error) {
 
 			command.Name = string(fields[0])
 			command.Args = string(fields[1])
+		case "EMBED":
+			return nil, fmt.Errorf("deprecated command: EMBED is no longer supported, use the /embed API endpoint instead")
 		default:
 			if !bytes.HasPrefix(fields[0], []byte("#")) {
 				// log a warning for unknown commands

server/images.go

@@ -1,7 +1,6 @@
 package server
 
 import (
-	"bufio"
 	"bytes"
 	"context"
 	"crypto/sha256"
@@ -26,7 +25,6 @@ import (
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llm"
 	"github.com/jmorganca/ollama/parser"
-	"github.com/jmorganca/ollama/vector"
 	"github.com/jmorganca/ollama/version"
 )
 
@@ -49,10 +47,9 @@ type Model struct {
 	Digest        string
 	ConfigDigest  string
 	Options       map[string]interface{}
-	Embeddings    []vector.Embedding
 }
 
-func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, error) {
+func (m *Model) Prompt(request api.GenerateRequest) (string, error) {
 	t := m.Template
 	if request.Template != "" {
 		t = request.Template
@@ -67,7 +64,6 @@ func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, e
 		First  bool
 		System string
 		Prompt string
-		Embed  string
 
 		// deprecated: versions <= 0.0.7 used this to omit the system prompt
 		Context []int
@@ -77,7 +73,6 @@ func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, e
 	vars.System = m.System
 	vars.Prompt = request.Prompt
 	vars.Context = request.Context
-	vars.Embed = embedding
 
 	if request.System != "" {
 		vars.System = request.System
@@ -190,15 +185,9 @@ func GetModel(name string) (*Model, error) {
 			model.ModelPath = filename
 			model.OriginalModel = layer.From
 		case "application/vnd.ollama.image.embed":
-			file, err := os.Open(filename)
-			if err != nil {
-				return nil, fmt.Errorf("failed to open file: %s", filename)
-			}
-			defer file.Close()
-
-			if err = json.NewDecoder(file).Decode(&model.Embeddings); err != nil {
-				return nil, err
-			}
+			// Deprecated in versions  > 0.1.2
+			// TODO: remove this warning in a future version
+			log.Print("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
 			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.template":
@@ -310,13 +299,11 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 	var layers []*LayerReader
 	params := make(map[string][]string)
 	var sourceParams map[string]any
-	embed := EmbeddingParams{fn: fn}
 	for _, c := range commands {
 		log.Printf("[%s] - %s\n", c.Name, c.Args)
 		switch c.Name {
 		case "model":
 			fn(api.ProgressResponse{Status: "looking for model"})
-			embed.model = c.Args
 
 			mp := ParseModelPath(c.Args)
 			mf, _, err := GetManifest(mp)
@@ -340,7 +327,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 						return err
 					}
 				} else {
-					embed.model = modelFile
 					// create a model from this specified file
 					fn(api.ProgressResponse{Status: "creating model layer"})
 					file, err := os.Open(modelFile)
@@ -421,12 +407,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 					layers = append(layers, newLayer)
 				}
 			}
-		case "embed":
-			embedFilePath, err := filenameWithPath(path, c.Args)
-			if err != nil {
-				return err
-			}
-			embed.files = append(embed.files, embedFilePath)
 		case "adapter":
 			fn(api.ProgressResponse{Status: fmt.Sprintf("creating model %s layer", c.Name)})
 
@@ -517,18 +497,8 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 		}
 		l.MediaType = "application/vnd.ollama.image.params"
 		layers = append(layers, l)
-
-		// apply these parameters to the embedding options, in case embeddings need to be generated using this model
-		embed.opts = formattedParams
 	}
 
-	// generate the embedding layers
-	embeddingLayers, err := embeddingLayers(workDir, embed)
-	if err != nil {
-		return err
-	}
-	layers = append(layers, embeddingLayers...)
-
 	digests, err := getLayerDigests(layers)
 	if err != nil {
 		return err
@@ -572,146 +542,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 	return nil
 }
 
-type EmbeddingParams struct {
-	model string
-	opts  map[string]interface{}
-	files []string // paths to files to embed
-	fn    func(resp api.ProgressResponse)
-}
-
-// embeddingLayers loads the associated LLM and generates the embeddings to be stored from an input file
-func embeddingLayers(workDir string, e EmbeddingParams) ([]*LayerReader, error) {
-	layers := []*LayerReader{}
-	if len(e.files) > 0 {
-		// check if the model is a file path or a model name
-		model, err := GetModel(e.model)
-		if err != nil {
-			if !strings.Contains(err.Error(), "couldn't open file") {
-				return nil, fmt.Errorf("unexpected error opening model to generate embeddings: %v", err)
-			}
-			// the model may be a file path, create a model from this file
-			model = &Model{ModelPath: e.model}
-		}
-
-		if err := load(context.Background(), workDir, model, e.opts, defaultSessionDuration); err != nil {
-			return nil, fmt.Errorf("load model to generate embeddings: %v", err)
-		}
-
-		// this will be used to check if we already have embeddings for a file
-		modelInfo, err := os.Stat(model.ModelPath)
-		if err != nil {
-			return nil, fmt.Errorf("failed to get model file info: %v", err)
-		}
-
-		addedFiles := make(map[string]bool) // keep track of files that have already been added
-		for _, filePattern := range e.files {
-			matchingFiles, err := filepath.Glob(filePattern)
-			if err != nil {
-				return nil, fmt.Errorf("could not find files with pattern %s: %w", filePattern, err)
-			}
-
-			for _, filePath := range matchingFiles {
-				if addedFiles[filePath] {
-					continue
-				}
-				addedFiles[filePath] = true
-				// check if we already have embeddings for this file path
-				layerIdentifier := fmt.Sprintf("%s:%s:%s:%d", filePath, e.model, modelInfo.ModTime().Format("2006-01-02 15:04:05"), modelInfo.Size())
-				digest, _ := GetSHA256Digest(strings.NewReader(layerIdentifier))
-				existing, err := existingFileEmbeddings(digest)
-				if err != nil {
-					return nil, fmt.Errorf("failed to check existing embeddings for file %s: %v", filePath, err)
-				}
-
-				// TODO: check file type
-				f, err := os.Open(filePath)
-				if err != nil {
-					return nil, fmt.Errorf("could not open embed file: %w", err)
-				}
-				scanner := bufio.NewScanner(f)
-				scanner.Split(bufio.ScanLines)
-
-				data := []string{}
-				for scanner.Scan() {
-					data = append(data, scanner.Text())
-				}
-				f.Close()
-
-				// the digest of the file is set here so that the client knows a new operation is in progress
-				fileDigest, _ := GetSHA256Digest(bytes.NewReader([]byte(filePath)))
-
-				embeddings := []vector.Embedding{}
-				for i, d := range data {
-					if strings.TrimSpace(d) == "" {
-						continue
-					}
-					e.fn(api.ProgressResponse{
-						Status:    fmt.Sprintf("creating embeddings for file %s", filePath),
-						Digest:    fileDigest,
-						Total:     int64(len(data) - 1),
-						Completed: int64(i),
-					})
-					if len(existing[d]) > 0 {
-						// already have an embedding for this line
-						embeddings = append(embeddings, vector.Embedding{Data: d, Vector: existing[d]})
-						continue
-					}
-					embed, err := loaded.llm.Embedding(context.Background(), d)
-					if err != nil {
-						log.Printf("failed to generate embedding for '%s' line %d: %v", filePath, i+1, err)
-						continue
-					}
-					embeddings = append(embeddings, vector.Embedding{Data: d, Vector: embed})
-				}
-
-				b, err := json.Marshal(embeddings)
-				if err != nil {
-					return nil, fmt.Errorf("failed to encode embeddings: %w", err)
-				}
-				r := bytes.NewReader(b)
-
-				layer := &LayerReader{
-					Layer: Layer{
-						MediaType: "application/vnd.ollama.image.embed",
-						Digest:    digest,
-						Size:      r.Size(),
-					},
-					Reader: r,
-				}
-
-				layers = append(layers, layer)
-			}
-		}
-	}
-	return layers, nil
-}
-
-// existingFileEmbeddings checks if we already have embeddings for a file and loads them into a look-up map
-func existingFileEmbeddings(digest string) (map[string][]float64, error) {
-	path, err := GetBlobsPath(digest)
-	if err != nil {
-		return nil, fmt.Errorf("embeddings blobs path: %w", err)
-	}
-	existingFileEmbeddings := make(map[string][]float64)
-	if _, err := os.Stat(path); err == nil {
-		// already have some embeddings for this file, load embeddings previously generated
-		file, err := os.Open(path)
-		if err != nil {
-			return nil, fmt.Errorf("failed to open existing embedding file: %s", err)
-		}
-		defer file.Close()
-
-		existing := []vector.Embedding{}
-		if err = json.NewDecoder(file).Decode(&existing); err != nil {
-			return nil, err
-		}
-		for _, e := range existing {
-			existingFileEmbeddings[e.Data] = e.Vector
-		}
-	}
-	return existingFileEmbeddings, nil
-}
-
 func removeLayerFromLayers(layers []*LayerReader, mediaType string) []*LayerReader {
 	return slices.DeleteFunc(layers, func(layer *LayerReader) bool {
 		return layer.MediaType == mediaType
@@ -727,8 +557,7 @@ func SaveLayers(layers []*LayerReader, fn func(resp api.ProgressResponse), force
 		}
 
 		_, err = os.Stat(fp)
-		// note: embed layers are always written since their digest doesnt indicate anything about the contents
-		if os.IsNotExist(err) || force || layer.MediaType == "application/vnd.ollama.image.embed" {
+		if os.IsNotExist(err) || force {
 			fn(api.ProgressResponse{Status: fmt.Sprintf("writing layer %s", layer.Digest)})
 
 			out, err := os.Create(fp)

server/images_test.go

@@ -12,7 +12,7 @@ func TestModelPrompt(t *testing.T) {
 		Template: "a{{ .Prompt }}b",
 		Prompt:   "<h1>",
 	}
-	s, err := m.Prompt(req, "")
+	s, err := m.Prompt(req)
 	if err != nil {
 		t.Fatal(err)
 	}

server/routes.go

@@ -23,11 +23,9 @@ 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/llm"
-	"github.com/jmorganca/ollama/vector"
 )
 
 var mode string = gin.DebugMode
@@ -47,8 +45,7 @@ func init() {
 var loaded struct {
 	mu sync.Mutex
 
-	llm        llm.LLM
-	Embeddings []vector.Embedding
+	llm llm.LLM
 
 	expireAt    time.Time
 	expireTimer *time.Timer
@@ -90,11 +87,6 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 			loaded.digest = ""
 		}
 
-		if model.Embeddings != nil && len(model.Embeddings) > 0 {
-			opts.EmbeddingOnly = true // this is requried to generate embeddings, completions will still work
-			loaded.Embeddings = model.Embeddings
-		}
-
 		llmModel, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
 		if err != nil {
 			return err
@@ -106,12 +98,12 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 		loaded.options = opts
 
 		if opts.NumKeep < 0 {
-			promptWithSystem, err := model.Prompt(api.GenerateRequest{}, "")
+			promptWithSystem, err := model.Prompt(api.GenerateRequest{})
 			if err != nil {
 				return err
 			}
 
-			promptNoSystem, err := model.Prompt(api.GenerateRequest{Context: []int{0}}, "")
+			promptNoSystem, err := model.Prompt(api.GenerateRequest{Context: []int{0}})
 			if err != nil {
 				return err
 			}
@@ -195,22 +187,7 @@ func GenerateHandler(c *gin.Context) {
 
 	checkpointLoaded := time.Now()
 
-	embedding := ""
-	if model.Embeddings != nil && len(model.Embeddings) > 0 {
-		promptEmbed, err := loaded.llm.Embedding(c.Request.Context(), 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)
+	prompt, err := model.Prompt(req)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return

vector/store.go

@@ -1,69 +0,0 @@
-package vector
-
-import (
-	"container/heap"
-	"sort"
-
-	"gonum.org/v1/gonum/mat"
-)
-
-type Embedding struct {
-	Vector []float64 // the embedding vector
-	Data   string    // the data represted by the embedding
-}
-
-type EmbeddingSimilarity struct {
-	Embedding  Embedding // the embedding that was used to calculate the similarity
-	Similarity float64   // the similarity between the embedding and the query
-}
-
-type Heap []EmbeddingSimilarity
-
-func (h Heap) Len() int           { return len(h) }
-func (h Heap) Less(i, j int) bool { return h[i].Similarity < h[j].Similarity }
-func (h Heap) Swap(i, j int)      { h[i], h[j] = h[j], h[i] }
-func (h *Heap) Push(e any) {
-	*h = append(*h, e.(EmbeddingSimilarity))
-}
-
-func (h *Heap) Pop() interface{} {
-	old := *h
-	n := len(old)
-	x := old[n-1]
-	*h = old[0 : n-1]
-	return x
-}
-
-// cosineSimilarity is a measure that calculates the cosine of the angle between two vectors.
-// This value will range from -1 to 1, where 1 means the vectors are identical.
-func cosineSimilarity(vec1, vec2 *mat.VecDense) float64 {
-	dotProduct := mat.Dot(vec1, vec2)
-	norms := mat.Norm(vec1, 2) * mat.Norm(vec2, 2)
-
-	if norms == 0 {
-		return 0
-	}
-	return dotProduct / norms
-}
-
-func TopK(k int, query *mat.VecDense, embeddings []Embedding) []EmbeddingSimilarity {
-	h := &Heap{}
-	heap.Init(h)
-	for _, emb := range embeddings {
-		similarity := cosineSimilarity(query, mat.NewVecDense(len(emb.Vector), emb.Vector))
-		heap.Push(h, EmbeddingSimilarity{Embedding: emb, Similarity: similarity})
-		if h.Len() > k {
-			heap.Pop(h)
-		}
-	}
-
-	topK := make([]EmbeddingSimilarity, 0, h.Len())
-	for h.Len() > 0 {
-		topK = append(topK, heap.Pop(h).(EmbeddingSimilarity))
-	}
-	sort.Slice(topK, func(i, j int) bool {
-		return topK[i].Similarity > topK[j].Similarity
-	})
-
-	return topK
-}