tools: refactor tool call parsing and enable streaming (#10415)

server/routes.go

@@ -38,6 +38,7 @@ import (
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/tools"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -1482,11 +1483,20 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
+	var toolParser *tools.Parser
+	if len(req.Tools) > 0 {
+		toolParser, err = tools.NewParser(m.Template.Template)
+		if err != nil {
+			slog.Error("failed to create tool parser", "error", err)
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+	}
+
 	ch := make(chan any)
 	go func() {
 		defer close(ch)
-		var sb strings.Builder
-		var toolCallIndex int = 0
+
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
 			Prompt:  prompt,
 			Images:  images,
@@ -1512,37 +1522,21 @@ func (s *Server) ChatHandler(c *gin.Context) {
 				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 			}
 
-			// TODO: tool call checking and filtering should be moved outside of this callback once streaming
-			// however this was a simple change for now without reworking streaming logic of this (and other)
-			// handlers
-			if req.Stream != nil && !*req.Stream || len(req.Tools) == 0 {
-				ch <- res
-				return
-			}
-
-			// Streaming tool calls:
-			// If tools are recognized, use a flag to track the sending of a tool downstream
-			// This ensures that content is cleared from the message on the last chunk sent
-			sb.WriteString(r.Content)
-			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
-				res.Message.ToolCalls = toolCalls
-				for i := range toolCalls {
-					toolCalls[i].Function.Index = toolCallIndex
-					toolCallIndex++
+			if len(req.Tools) > 0 {
+				toolCalls, content := toolParser.Add(r.Content)
+				if len(content) > 0 {
+					res.Message.Content = content
+				} else if len(toolCalls) > 0 {
+					res.Message.ToolCalls = toolCalls
+					res.Message.Content = ""
+				} else {
+					if r.Done {
+						ch <- res
+					}
+					return
 				}
-				res.Message.Content = ""
-				sb.Reset()
-				ch <- res
-				return
-			}
-
-			if r.Done {
-				// Send any remaining content if no tool calls were detected
-				if toolCallIndex == 0 {
-					res.Message.Content = sb.String()
-				}
-				ch <- res
 			}
+			ch <- res
 		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
@@ -1551,11 +1545,15 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	if req.Stream != nil && !*req.Stream {
 		var resp api.ChatResponse
 		var sb strings.Builder
+		var toolCalls []api.ToolCall
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.ChatResponse:
 				sb.WriteString(t.Message.Content)
 				resp = t
+				if len(req.Tools) > 0 {
+					toolCalls = append(toolCalls, t.Message.ToolCalls...)
+				}
 			case gin.H:
 				msg, ok := t["error"].(string)
 				if !ok {
@@ -1571,12 +1569,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		}
 
 		resp.Message.Content = sb.String()
-
-		if len(req.Tools) > 0 {
-			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
-				resp.Message.ToolCalls = toolCalls
-				resp.Message.Content = ""
-			}
+		if len(toolCalls) > 0 {
+			resp.Message.ToolCalls = toolCalls
 		}
 
 		c.JSON(http.StatusOK, resp)