server: add logprobs and top_logprobs support to Ollama's API (#12899)

Adds logprobs support to Ollama's API including support for Ollama's
OpenAI-compatible API. By specifying the new 'logprobs' boolean parameter
in the API, Ollama will return the log probabilities for each token generated.
'top_logprobs', an integer value can also be specified up to the value 20.
When specified, the API will also provide the number of most likely tokens to
return at each token position

Co-authored-by: Baptiste Jamin <baptiste@crisp.chat>

api/types.go

@@ -117,6 +117,14 @@ type GenerateRequest struct {
 	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
 	// template instead of calling the model.
 	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
+
+	// Logprobs specifies whether to return log probabilities of the output tokens.
+	Logprobs bool `json:"logprobs,omitempty"`
+
+	// TopLogprobs is the number of most likely tokens to return at each token position,
+	// each with an associated log probability. Only applies when Logprobs is true.
+	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
+	TopLogprobs int `json:"top_logprobs,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -159,6 +167,14 @@ type ChatRequest struct {
 	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
 	// template instead of calling the model.
 	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
+
+	// Logprobs specifies whether to return log probabilities of the output tokens.
+	Logprobs bool `json:"logprobs,omitempty"`
+
+	// TopLogprobs is the number of most likely tokens to return at each token position,
+	// each with an associated log probability. Only applies when Logprobs is true.
+	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
+	TopLogprobs int `json:"top_logprobs,omitempty"`
 }
 
 type Tools []Tool
@@ -343,6 +359,24 @@ func (t *ToolFunction) String() string {
 	return string(bts)
 }
 
+// TokenLogprob represents log probability information for a single token alternative.
+type TokenLogprob struct {
+	// Token is the text representation of the token.
+	Token string `json:"token"`
+
+	// Logprob is the log probability of this token.
+	Logprob float64 `json:"logprob"`
+}
+
+// Logprob contains log probability information for a generated token.
+type Logprob struct {
+	TokenLogprob
+
+	// TopLogprobs contains the most likely tokens and their log probabilities
+	// at this position, if requested via TopLogprobs parameter.
+	TopLogprobs []TokenLogprob `json:"top_logprobs,omitempty"`
+}
+
 // ChatResponse is the response returned by [Client.Chat]. Its fields are
 // similar to [GenerateResponse].
 type ChatResponse struct {
@@ -369,6 +403,10 @@ type ChatResponse struct {
 
 	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
 
+	// Logprobs contains log probability information for the generated tokens,
+	// if requested via the Logprobs parameter.
+	Logprobs []Logprob `json:"logprobs,omitempty"`
+
 	Metrics
 }
 
@@ -677,6 +715,10 @@ type GenerateResponse struct {
 	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
 
 	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
+
+	// Logprobs contains log probability information for the generated tokens,
+	// if requested via the Logprobs parameter.
+	Logprobs []Logprob `json:"logprobs,omitempty"`
 }
 
 // ModelDetails provides details about a model.