tools

server/images.go

@@ -38,7 +38,10 @@ var errCapabilityCompletion = errors.New("completion")
 
 type Capability string
 
-const CapabilityCompletion = Capability("completion")
+const (
+	CapabilityCompletion = Capability("completion")
+	CapabilityTools      = Capability("tools")
+)
 
 type registryOptions struct {
 	Insecure bool
@@ -88,6 +91,10 @@ func (m *Model) CheckCapabilities(caps ...Capability) error {
 			if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
 				errs = append(errs, errCapabilityCompletion)
 			}
+		case CapabilityTools:
+			if !slices.Contains(m.Template.Vars(), "tools") {
+				errs = append(errs, errors.New("tools"))
+			}
 		default:
 			slog.Error("unknown capability", "capability", cap)
 			return fmt.Errorf("unknown capability: %s", cap)
@@ -95,7 +102,7 @@ func (m *Model) CheckCapabilities(caps ...Capability) error {
 	}
 
 	if err := errors.Join(errs...); err != nil {
-		return fmt.Errorf("missing capabilities: %w", errors.Join(errs...))
+		return fmt.Errorf("does not support %w", errors.Join(errs...))
 	}
 
 	return nil

server/model.go

@@ -4,6 +4,7 @@ import (
 	"archive/zip"
 	"bytes"
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -11,7 +12,11 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"slices"
+	"strings"
+	"text/template/parse"
 
+	"github.com/google/uuid"
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/llm"
@@ -289,3 +294,103 @@ func detectContentType(r io.Reader) (string, error) {
 
 	return "unknown", nil
 }
+
+// parseToolCalls attempts to parse a JSON string into a slice of ToolCalls.
+// mxyng: this only really works if the input contains tool calls in some JSON format
+func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
+	// create a subtree from the node that ranges over .ToolCalls
+	tmpl := m.Template.Subtree(func(n parse.Node) bool {
+		if t, ok := n.(*parse.RangeNode); ok {
+			return slices.Contains(template.Identifiers(t.Pipe), "ToolCalls")
+		}
+
+		return false
+	})
+
+	if tmpl == nil {
+		return nil, false
+	}
+
+	var b bytes.Buffer
+	if err := tmpl.Execute(&b, map[string][]map[string]any{
+		"ToolCalls": {
+			{
+				"Function": map[string]any{
+					"Name":      "@@name@@",
+					"Arguments": "@@arguments@@",
+				},
+			},
+		},
+	}); err != nil {
+		return nil, false
+	}
+
+	var kv map[string]string
+	// execute the subtree with placeholders to identify the keys
+	if err := json.Unmarshal(b.Bytes(), &kv); err != nil {
+		return nil, false
+	}
+
+	// find the keys that correspond to the name and arguments fields
+	var name, arguments string
+	for k, v := range kv {
+		switch v {
+		case "@@name@@":
+			name = k
+		case "@@arguments@@":
+			arguments = k
+		}
+	}
+
+	var sm []map[string]any
+	decoder := json.NewDecoder(strings.NewReader(s))
+	for {
+		// incrementally decode the JSON into a list of JSON objects
+		// skipping over any invalid tokens
+		if err := decoder.Decode(&sm); err != nil {
+			if errors.Is(err, io.EOF) {
+				break
+			}
+
+			if errors.As(err, new(*json.SyntaxError)) {
+				r := decoder.Buffered()
+				if _, err := r.Read(make([]byte, decoder.InputOffset()+1)); err != nil {
+					break
+				}
+
+				decoder = json.NewDecoder(r)
+				continue
+			}
+
+			return nil, false
+		}
+
+		// break as soon as a valid object is decoded
+		break
+	}
+
+	var toolCalls []api.ToolCall
+	for _, kv := range sm {
+		call := api.ToolCall{
+			ID:   uuid.New().String(),
+			Type: "function",
+		}
+
+		for k, v := range kv {
+			switch k {
+			case name:
+				call.Function.Name = v.(string)
+			case arguments:
+				call.Function.Arguments = v.(map[string]any)
+			}
+		}
+
+		toolCalls = append(toolCalls, call)
+	}
+
+	if len(toolCalls) > 0 {
+		return toolCalls, true
+	}
+
+	return nil, false
+}

server/prompt.go

@@ -15,7 +15,7 @@ type tokenizeFunc func(context.Context, string) ([]int, error)
 // chatPrompt accepts a list of messages and returns the prompt and images that should be used for the next chat turn.
 // chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
 // latest message and 2) system messages
-func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message) (prompt string, images []llm.ImageData, _ error) {
+func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool) (prompt string, images []llm.ImageData, _ error) {
 	var system []api.Message
 	// always include the last message
 	n := len(msgs) - 1
@@ -29,7 +29,7 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 		}
 
 		var b bytes.Buffer
-		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...)}); err != nil {
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools}); err != nil {
 			return "", nil, err
 		}
 
@@ -57,7 +57,7 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 
 	// truncate any messages that do not fit into the context window
 	var b bytes.Buffer
-	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[n:]...)}); err != nil {
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[n:]...), Tools: tools}); err != nil {
 		return "", nil, err
 	}
 

server/prompt_test.go

@@ -192,7 +192,7 @@ func TestChatPrompt(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			model := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
 			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
-			prompt, images, err := chatPrompt(context.TODO(), &model, tokenize, &opts, tt.msgs)
+			prompt, images, err := chatPrompt(context.TODO(), &model, tokenize, &opts, tt.msgs, nil)
 			if err != nil {
 				t.Fatal(err)
 			}

server/routes.go

@@ -265,6 +265,11 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		}
 
 		r.Response = sb.String()
+		if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
+			r.ToolCalls = toolCalls
+			r.Response = ""
+		}
+
 		c.JSON(http.StatusOK, r)
 		return
 	}
@@ -1279,6 +1284,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 
 	caps := []Capability{CapabilityCompletion}
+	if req.Tools != nil {
+		caps = append(caps, CapabilityTools)
+	}
+
 	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, caps, req.Options, req.KeepAlive)
 	if errors.Is(err, errCapabilityCompletion) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support chat", req.Model)})
@@ -1305,7 +1314,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		req.Messages = append([]api.Message{{Role: "system", Content: m.System}}, req.Messages...)
 	}
 
-	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, req.Messages)
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, req.Messages, req.Tools)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1348,13 +1357,13 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		var r api.ChatResponse
+		var resp api.ChatResponse
 		var sb strings.Builder
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.ChatResponse:
 				sb.WriteString(t.Message.Content)
-				r = t
+				resp = t
 			case gin.H:
 				msg, ok := t["error"].(string)
 				if !ok {
@@ -1369,8 +1378,13 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 		}
 
-		r.Message.Content = sb.String()
-		c.JSON(http.StatusOK, r)
+		resp.Message.Content = sb.String()
+		if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
+			resp.Message.ToolCalls = toolCalls
+			resp.Message.Content = ""
+		}
+
+		c.JSON(http.StatusOK, resp)
 		return
 	}