openai: refactor to split compat layer and middleware

This makes the core openai compat layer independent of the middleware
that adapts it to our particular gin routes

middleware/openai.go

+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"math/rand"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/openai"
+)
+
+type BaseWriter struct {
+	gin.ResponseWriter
+}
+
+type ChatWriter struct {
+	stream        bool
+	streamOptions *openai.StreamOptions
+	id            string
+	toolCallSent  bool
+	BaseWriter
+}
+
+type CompleteWriter struct {
+	stream        bool
+	streamOptions *openai.StreamOptions
+	id            string
+	BaseWriter
+}
+
+type ListWriter struct {
+	BaseWriter
+}
+
+type RetrieveWriter struct {
+	BaseWriter
+	model string
+}
+
+type EmbedWriter struct {
+	BaseWriter
+	model string
+}
+
+func (w *BaseWriter) writeError(data []byte) (int, error) {
+	var serr api.StatusError
+	err := json.Unmarshal(data, &serr)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.NewError(http.StatusInternalServerError, serr.Error()))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ChatWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// chat chunk
+	if w.stream {
+		c := openai.ToChunk(w.id, chatResponse, w.toolCallSent)
+		d, err := json.Marshal(c)
+		if err != nil {
+			return 0, err
+		}
+		if !w.toolCallSent && len(c.Choices) > 0 && len(c.Choices[0].Delta.ToolCalls) > 0 {
+			w.toolCallSent = true
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if chatResponse.Done {
+			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+				u := openai.ToUsage(chatResponse)
+				c.Usage = &u
+				c.Choices = []openai.ChunkChoice{}
+				d, err := json.Marshal(c)
+				if err != nil {
+					return 0, err
+				}
+				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+				if err != nil {
+					return 0, err
+				}
+			}
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// chat completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToChatCompletion(w.id, chatResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ChatWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	err := json.Unmarshal(data, &generateResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// completion chunk
+	if w.stream {
+		c := openai.ToCompleteChunk(w.id, generateResponse)
+		if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+			c.Usage = &openai.Usage{}
+		}
+		d, err := json.Marshal(c)
+		if err != nil {
+			return 0, err
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if generateResponse.Done {
+			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+				u := openai.ToUsageGenerate(generateResponse)
+				c.Usage = &u
+				c.Choices = []openai.CompleteChunkChoice{}
+				d, err := json.Marshal(c)
+				if err != nil {
+					return 0, err
+				}
+				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+				if err != nil {
+					return 0, err
+				}
+			}
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToCompletion(w.id, generateResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *CompleteWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *ListWriter) writeResponse(data []byte) (int, error) {
+	var listResponse api.ListResponse
+	err := json.Unmarshal(data, &listResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToListCompletion(listResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ListWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
+	var showResponse api.ShowResponse
+	err := json.Unmarshal(data, &showResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// retrieve completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToModel(showResponse, w.model))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *RetrieveWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
+	var embedResponse api.EmbedResponse
+	err := json.Unmarshal(data, &embedResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToEmbeddingList(w.model, embedResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *EmbedWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ListMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		w := &ListWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func RetrieveMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &RetrieveWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      c.Param("model"),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func CompletionsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.CompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		genReq, err := openai.FromCompleteRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &CompleteWriter{
+			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
+			stream:        req.Stream,
+			id:            fmt.Sprintf("cmpl-%d", rand.Intn(999)),
+			streamOptions: req.StreamOptions,
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}
+
+func EmbeddingsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.EmbedRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Input == "" {
+			req.Input = []string{""}
+		}
+
+		if req.Input == nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		if v, ok := req.Input.([]any); ok && len(v) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input, Dimensions: req.Dimensions}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &EmbedWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      req.Model,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func ChatMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ChatCompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "[] is too short - 'messages'"))
+			return
+		}
+
+		var b bytes.Buffer
+
+		chatReq, err := openai.FromChatRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ChatWriter{
+			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
+			stream:        req.Stream,
+			id:            fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
+			streamOptions: req.StreamOptions,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

openai/openai.go

-// openai package provides middleware for partial compatibility with the OpenAI REST API
+// openai package provides core transformation logic for partial compatibility with the OpenAI REST API
 package openai
 
 import (
-	"bytes"
 	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"math/rand"
 	"net/http"
 	"strings"
 	"time"
 
-	"github.com/gin-gonic/gin"
-
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/types/model"
 )
@@ -220,11 +216,12 @@ func NewError(code int, message string) ErrorResponse {
 	return ErrorResponse{Error{Type: etype, Message: message}}
 }
 
-func toUsage(r api.ChatResponse) Usage {
+// ToUsage converts an api.ChatResponse to Usage
+func ToUsage(r api.ChatResponse) Usage {
 	return Usage{
-		PromptTokens:     r.PromptEvalCount,
-		CompletionTokens: r.EvalCount,
-		TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		PromptTokens:     r.Metrics.PromptEvalCount,
+		CompletionTokens: r.Metrics.EvalCount,
+		TotalTokens:      r.Metrics.PromptEvalCount + r.Metrics.EvalCount,
 	}
 }
 
@@ -256,7 +253,8 @@ func toToolCalls(tc []api.ToolCall) []ToolCall {
 	return toolCalls
 }
 
-func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
+// ToChatCompletion converts an api.ChatResponse to ChatCompletion
+func ToChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 	toolCalls := toToolCalls(r.Message.ToolCalls)
 	return ChatCompletion{
 		Id:                id,
@@ -276,12 +274,13 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 				}
 				return nil
 			}(r.DoneReason),
-		}}, Usage: toUsage(r),
+		}}, Usage: ToUsage(r),
 		DebugInfo: r.DebugInfo,
 	}
 }
 
-func toChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChunk {
+// ToChunk converts an api.ChatResponse to ChatCompletionChunk
+func ToChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChunk {
 	toolCalls := toToolCalls(r.Message.ToolCalls)
 	return ChatCompletionChunk{
 		Id:                id,
@@ -305,15 +304,17 @@ func toChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChu
 	}
 }
 
-func toUsageGenerate(r api.GenerateResponse) Usage {
+// ToUsageGenerate converts an api.GenerateResponse to Usage
+func ToUsageGenerate(r api.GenerateResponse) Usage {
 	return Usage{
-		PromptTokens:     r.PromptEvalCount,
-		CompletionTokens: r.EvalCount,
-		TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		PromptTokens:     r.Metrics.PromptEvalCount,
+		CompletionTokens: r.Metrics.EvalCount,
+		TotalTokens:      r.Metrics.PromptEvalCount + r.Metrics.EvalCount,
 	}
 }
 
-func toCompletion(id string, r api.GenerateResponse) Completion {
+// ToCompletion converts an api.GenerateResponse to Completion
+func ToCompletion(id string, r api.GenerateResponse) Completion {
 	return Completion{
 		Id:                id,
 		Object:            "text_completion",
@@ -330,11 +331,12 @@ func toCompletion(id string, r api.GenerateResponse) Completion {
 				return nil
 			}(r.DoneReason),
 		}},
-		Usage: toUsageGenerate(r),
+		Usage: ToUsageGenerate(r),
 	}
 }
 
-func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
+// ToCompleteChunk converts an api.GenerateResponse to CompletionChunk
+func ToCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
 	return CompletionChunk{
 		Id:                id,
 		Object:            "text_completion",
@@ -354,7 +356,8 @@ func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
 	}
 }
 
-func toListCompletion(r api.ListResponse) ListCompletion {
+// ToListCompletion converts an api.ListResponse to ListCompletion
+func ToListCompletion(r api.ListResponse) ListCompletion {
 	var data []Model
 	for _, m := range r.Models {
 		data = append(data, Model{
@@ -371,7 +374,8 @@ func toListCompletion(r api.ListResponse) ListCompletion {
 	}
 }
 
-func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
+// ToEmbeddingList converts an api.EmbedResponse to EmbeddingList
+func ToEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
 	if r.Embeddings != nil {
 		var data []Embedding
 		for i, e := range r.Embeddings {
@@ -396,7 +400,8 @@ func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
 	return EmbeddingList{}
 }
 
-func toModel(r api.ShowResponse, m string) Model {
+// ToModel converts an api.ShowResponse to Model
+func ToModel(r api.ShowResponse, m string) Model {
 	return Model{
 		Id:      m,
 		Object:  "model",
@@ -405,7 +410,8 @@ func toModel(r api.ShowResponse, m string) Model {
 	}
 }
 
-func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
+// FromChatRequest converts a ChatCompletionRequest to api.ChatRequest
+func FromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	var messages []api.Message
 	for _, msg := range r.Messages {
 		toolName := ""
@@ -609,7 +615,8 @@ func fromCompletionToolCall(toolCalls []ToolCall) ([]api.ToolCall, error) {
 	return apiToolCalls, nil
 }
 
-func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
+// FromCompleteRequest converts a CompletionRequest to api.GenerateRequest
+func FromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 	options := make(map[string]any)
 
 	switch stop := r.Stop.(type) {
@@ -660,413 +667,3 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		DebugRenderOnly: r.DebugRenderOnly,
 	}, nil
 }
-
-type BaseWriter struct {
-	gin.ResponseWriter
-}
-
-type ChatWriter struct {
-	stream        bool
-	streamOptions *StreamOptions
-	id            string
-	toolCallSent  bool
-	BaseWriter
-}
-
-type CompleteWriter struct {
-	stream        bool
-	streamOptions *StreamOptions
-	id            string
-	BaseWriter
-}
-
-type ListWriter struct {
-	BaseWriter
-}
-
-type RetrieveWriter struct {
-	BaseWriter
-	model string
-}
-
-type EmbedWriter struct {
-	BaseWriter
-	model string
-}
-
-func (w *BaseWriter) writeError(data []byte) (int, error) {
-	var serr api.StatusError
-	err := json.Unmarshal(data, &serr)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(NewError(http.StatusInternalServerError, serr.Error()))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ChatWriter) writeResponse(data []byte) (int, error) {
-	var chatResponse api.ChatResponse
-	err := json.Unmarshal(data, &chatResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// chat chunk
-	if w.stream {
-		c := toChunk(w.id, chatResponse, w.toolCallSent)
-		d, err := json.Marshal(c)
-		if err != nil {
-			return 0, err
-		}
-		if !w.toolCallSent && len(c.Choices) > 0 && len(c.Choices[0].Delta.ToolCalls) > 0 {
-			w.toolCallSent = true
-		}
-
-		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
-		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-		if err != nil {
-			return 0, err
-		}
-
-		if chatResponse.Done {
-			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-				u := toUsage(chatResponse)
-				c.Usage = &u
-				c.Choices = []ChunkChoice{}
-				d, err := json.Marshal(c)
-				if err != nil {
-					return 0, err
-				}
-				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-				if err != nil {
-					return 0, err
-				}
-			}
-			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
-			if err != nil {
-				return 0, err
-			}
-		}
-
-		return len(data), nil
-	}
-
-	// chat completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toChatCompletion(w.id, chatResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ChatWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
-	var generateResponse api.GenerateResponse
-	err := json.Unmarshal(data, &generateResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// completion chunk
-	if w.stream {
-		c := toCompleteChunk(w.id, generateResponse)
-		if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-			c.Usage = &Usage{}
-		}
-		d, err := json.Marshal(c)
-		if err != nil {
-			return 0, err
-		}
-
-		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
-		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-		if err != nil {
-			return 0, err
-		}
-
-		if generateResponse.Done {
-			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-				u := toUsageGenerate(generateResponse)
-				c.Usage = &u
-				c.Choices = []CompleteChunkChoice{}
-				d, err := json.Marshal(c)
-				if err != nil {
-					return 0, err
-				}
-				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-				if err != nil {
-					return 0, err
-				}
-			}
-			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
-			if err != nil {
-				return 0, err
-			}
-		}
-
-		return len(data), nil
-	}
-
-	// completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toCompletion(w.id, generateResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *CompleteWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *ListWriter) writeResponse(data []byte) (int, error) {
-	var listResponse api.ListResponse
-	err := json.Unmarshal(data, &listResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ListWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
-	var showResponse api.ShowResponse
-	err := json.Unmarshal(data, &showResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// retrieve completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *RetrieveWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
-	var embedResponse api.EmbedResponse
-	err := json.Unmarshal(data, &embedResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toEmbeddingList(w.model, embedResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *EmbedWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func ListMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		w := &ListWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func RetrieveMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		// response writer
-		w := &RetrieveWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			model:      c.Param("model"),
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func CompletionsMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req CompletionRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		var b bytes.Buffer
-		genReq, err := fromCompleteRequest(req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &CompleteWriter{
-			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
-			stream:        req.Stream,
-			id:            fmt.Sprintf("cmpl-%d", rand.Intn(999)),
-			streamOptions: req.StreamOptions,
-		}
-
-		c.Writer = w
-		c.Next()
-	}
-}
-
-func EmbeddingsMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req EmbedRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if req.Input == "" {
-			req.Input = []string{""}
-		}
-
-		if req.Input == nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
-			return
-		}
-
-		if v, ok := req.Input.([]any); ok && len(v) == 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
-			return
-		}
-
-		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input, Dimensions: req.Dimensions}); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &EmbedWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			model:      req.Model,
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func ChatMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req ChatCompletionRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if len(req.Messages) == 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "[] is too short - 'messages'"))
-			return
-		}
-
-		var b bytes.Buffer
-
-		chatReq, err := fromChatRequest(req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &ChatWriter{
-			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
-			stream:        req.Stream,
-			id:            fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
-			streamOptions: req.StreamOptions,
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}

server/routes.go

@@ -37,8 +37,8 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/middleware"
 	"github.com/ollama/ollama/model/parsers"
-	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
@@ -1449,11 +1449,11 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
 
 	// Inference (OpenAI compatibility)
-	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
-	r.POST("/v1/completions", openai.CompletionsMiddleware(), s.GenerateHandler)
-	r.POST("/v1/embeddings", openai.EmbeddingsMiddleware(), s.EmbedHandler)
-	r.GET("/v1/models", openai.ListMiddleware(), s.ListHandler)
-	r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowHandler)
+	r.POST("/v1/chat/completions", middleware.ChatMiddleware(), s.ChatHandler)
+	r.POST("/v1/completions", middleware.CompletionsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/embeddings", middleware.EmbeddingsMiddleware(), s.EmbedHandler)
+	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
+	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 
 	if rc != nil {
 		// wrap old with new