1package openai
2
3import (
4 "encoding/base64"
5 "fmt"
6 "strings"
7
8 "charm.land/fantasy"
9 "github.com/openai/openai-go/v2"
10 "github.com/openai/openai-go/v2/packages/param"
11 "github.com/openai/openai-go/v2/shared"
12)
13
14// LanguageModelPrepareCallFunc is a function that prepares the call for the language model.
15type LanguageModelPrepareCallFunc = func(model fantasy.LanguageModel, params *openai.ChatCompletionNewParams, call fantasy.Call) ([]fantasy.CallWarning, error)
16
17// LanguageModelMapFinishReasonFunc is a function that maps the finish reason for the language model.
18type LanguageModelMapFinishReasonFunc = func(finishReason string) fantasy.FinishReason
19
20// LanguageModelUsageFunc is a function that calculates usage for the language model.
21type LanguageModelUsageFunc = func(choice openai.ChatCompletion) (fantasy.Usage, fantasy.ProviderOptionsData)
22
23// LanguageModelExtraContentFunc is a function that adds extra content for the language model.
24type LanguageModelExtraContentFunc = func(choice openai.ChatCompletionChoice) []fantasy.Content
25
26// LanguageModelStreamExtraFunc is a function that handles stream extra functionality for the language model.
27type LanguageModelStreamExtraFunc = func(chunk openai.ChatCompletionChunk, yield func(fantasy.StreamPart) bool, ctx map[string]any) (map[string]any, bool)
28
29// LanguageModelStreamUsageFunc is a function that calculates stream usage for the language model.
30type LanguageModelStreamUsageFunc = func(chunk openai.ChatCompletionChunk, ctx map[string]any, metadata fantasy.ProviderMetadata) (fantasy.Usage, fantasy.ProviderMetadata)
31
32// LanguageModelStreamProviderMetadataFunc is a function that handles stream provider metadata for the language model.
33type LanguageModelStreamProviderMetadataFunc = func(choice openai.ChatCompletionChoice, metadata fantasy.ProviderMetadata) fantasy.ProviderMetadata
34
35// LanguageModelToPromptFunc is a function that handles converting fantasy prompts to openai sdk messages.
36type LanguageModelToPromptFunc = func(prompt fantasy.Prompt, provider, model string) ([]openai.ChatCompletionMessageParamUnion, []fantasy.CallWarning)
37
38// DefaultPrepareCallFunc is the default implementation for preparing a call to the language model.
39func DefaultPrepareCallFunc(model fantasy.LanguageModel, params *openai.ChatCompletionNewParams, call fantasy.Call) ([]fantasy.CallWarning, error) {
40 if call.ProviderOptions == nil {
41 return nil, nil
42 }
43 var warnings []fantasy.CallWarning
44 providerOptions := &ProviderOptions{}
45 if v, ok := call.ProviderOptions[Name]; ok {
46 providerOptions, ok = v.(*ProviderOptions)
47 if !ok {
48 return nil, &fantasy.Error{Title: "invalid argument", Message: "openai provider options should be *openai.ProviderOptions"}
49 }
50 }
51
52 if providerOptions.LogitBias != nil {
53 params.LogitBias = providerOptions.LogitBias
54 }
55 if providerOptions.LogProbs != nil && providerOptions.TopLogProbs != nil {
56 providerOptions.LogProbs = nil
57 }
58 if providerOptions.LogProbs != nil {
59 params.Logprobs = param.NewOpt(*providerOptions.LogProbs)
60 }
61 if providerOptions.TopLogProbs != nil {
62 params.TopLogprobs = param.NewOpt(*providerOptions.TopLogProbs)
63 }
64 if providerOptions.User != nil {
65 params.User = param.NewOpt(*providerOptions.User)
66 }
67 if providerOptions.ParallelToolCalls != nil {
68 params.ParallelToolCalls = param.NewOpt(*providerOptions.ParallelToolCalls)
69 }
70 if providerOptions.MaxCompletionTokens != nil {
71 params.MaxCompletionTokens = param.NewOpt(*providerOptions.MaxCompletionTokens)
72 }
73
74 if providerOptions.TextVerbosity != nil {
75 params.Verbosity = openai.ChatCompletionNewParamsVerbosity(*providerOptions.TextVerbosity)
76 }
77 if providerOptions.Prediction != nil {
78 // Convert map[string]any to ChatCompletionPredictionContentParam
79 if content, ok := providerOptions.Prediction["content"]; ok {
80 if contentStr, ok := content.(string); ok {
81 params.Prediction = openai.ChatCompletionPredictionContentParam{
82 Content: openai.ChatCompletionPredictionContentContentUnionParam{
83 OfString: param.NewOpt(contentStr),
84 },
85 }
86 }
87 }
88 }
89 if providerOptions.Store != nil {
90 params.Store = param.NewOpt(*providerOptions.Store)
91 }
92 if providerOptions.Metadata != nil {
93 // Convert map[string]any to map[string]string
94 metadata := make(map[string]string)
95 for k, v := range providerOptions.Metadata {
96 if str, ok := v.(string); ok {
97 metadata[k] = str
98 }
99 }
100 params.Metadata = metadata
101 }
102 if providerOptions.PromptCacheKey != nil {
103 params.PromptCacheKey = param.NewOpt(*providerOptions.PromptCacheKey)
104 }
105 if providerOptions.SafetyIdentifier != nil {
106 params.SafetyIdentifier = param.NewOpt(*providerOptions.SafetyIdentifier)
107 }
108 if providerOptions.ServiceTier != nil {
109 params.ServiceTier = openai.ChatCompletionNewParamsServiceTier(*providerOptions.ServiceTier)
110 }
111
112 if providerOptions.ReasoningEffort != nil {
113 switch *providerOptions.ReasoningEffort {
114 case ReasoningEffortMinimal:
115 params.ReasoningEffort = shared.ReasoningEffortMinimal
116 case ReasoningEffortLow:
117 params.ReasoningEffort = shared.ReasoningEffortLow
118 case ReasoningEffortMedium:
119 params.ReasoningEffort = shared.ReasoningEffortMedium
120 case ReasoningEffortHigh:
121 params.ReasoningEffort = shared.ReasoningEffortHigh
122 default:
123 return nil, fmt.Errorf("reasoning model `%s` not supported", *providerOptions.ReasoningEffort)
124 }
125 }
126
127 if isReasoningModel(model.Model()) {
128 if providerOptions.LogitBias != nil {
129 params.LogitBias = nil
130 warnings = append(warnings, fantasy.CallWarning{
131 Type: fantasy.CallWarningTypeUnsupportedSetting,
132 Setting: "LogitBias",
133 Message: "LogitBias is not supported for reasoning models",
134 })
135 }
136 if providerOptions.LogProbs != nil {
137 params.Logprobs = param.Opt[bool]{}
138 warnings = append(warnings, fantasy.CallWarning{
139 Type: fantasy.CallWarningTypeUnsupportedSetting,
140 Setting: "Logprobs",
141 Message: "Logprobs is not supported for reasoning models",
142 })
143 }
144 if providerOptions.TopLogProbs != nil {
145 params.TopLogprobs = param.Opt[int64]{}
146 warnings = append(warnings, fantasy.CallWarning{
147 Type: fantasy.CallWarningTypeUnsupportedSetting,
148 Setting: "TopLogprobs",
149 Message: "TopLogprobs is not supported for reasoning models",
150 })
151 }
152 }
153
154 // Handle service tier validation
155 if providerOptions.ServiceTier != nil {
156 serviceTier := *providerOptions.ServiceTier
157 if serviceTier == "flex" && !supportsFlexProcessing(model.Model()) {
158 params.ServiceTier = ""
159 warnings = append(warnings, fantasy.CallWarning{
160 Type: fantasy.CallWarningTypeUnsupportedSetting,
161 Setting: "ServiceTier",
162 Details: "flex processing is only available for o3, o4-mini, and gpt-5 models",
163 })
164 } else if serviceTier == "priority" && !supportsPriorityProcessing(model.Model()) {
165 params.ServiceTier = ""
166 warnings = append(warnings, fantasy.CallWarning{
167 Type: fantasy.CallWarningTypeUnsupportedSetting,
168 Setting: "ServiceTier",
169 Details: "priority processing is only available for supported models (gpt-4, gpt-5, gpt-5-mini, o3, o4-mini) and requires Enterprise access. gpt-5-nano is not supported",
170 })
171 }
172 }
173 return warnings, nil
174}
175
176// DefaultMapFinishReasonFunc is the default implementation for mapping finish reasons.
177func DefaultMapFinishReasonFunc(finishReason string) fantasy.FinishReason {
178 switch finishReason {
179 case "stop":
180 return fantasy.FinishReasonStop
181 case "length":
182 return fantasy.FinishReasonLength
183 case "content_filter":
184 return fantasy.FinishReasonContentFilter
185 case "function_call", "tool_calls":
186 return fantasy.FinishReasonToolCalls
187 default:
188 return fantasy.FinishReasonUnknown
189 }
190}
191
192// DefaultUsageFunc is the default implementation for calculating usage.
193func DefaultUsageFunc(response openai.ChatCompletion) (fantasy.Usage, fantasy.ProviderOptionsData) {
194 completionTokenDetails := response.Usage.CompletionTokensDetails
195 promptTokenDetails := response.Usage.PromptTokensDetails
196
197 // Build provider metadata
198 providerMetadata := &ProviderMetadata{}
199
200 // Add logprobs if available
201 if len(response.Choices) > 0 && len(response.Choices[0].Logprobs.Content) > 0 {
202 providerMetadata.Logprobs = response.Choices[0].Logprobs.Content
203 }
204
205 // Add prediction tokens if available
206 if completionTokenDetails.AcceptedPredictionTokens > 0 || completionTokenDetails.RejectedPredictionTokens > 0 {
207 if completionTokenDetails.AcceptedPredictionTokens > 0 {
208 providerMetadata.AcceptedPredictionTokens = completionTokenDetails.AcceptedPredictionTokens
209 }
210 if completionTokenDetails.RejectedPredictionTokens > 0 {
211 providerMetadata.RejectedPredictionTokens = completionTokenDetails.RejectedPredictionTokens
212 }
213 }
214 return fantasy.Usage{
215 InputTokens: response.Usage.PromptTokens,
216 OutputTokens: response.Usage.CompletionTokens,
217 TotalTokens: response.Usage.TotalTokens,
218 ReasoningTokens: completionTokenDetails.ReasoningTokens,
219 CacheReadTokens: promptTokenDetails.CachedTokens,
220 }, providerMetadata
221}
222
223// DefaultStreamUsageFunc is the default implementation for calculating stream usage.
224func DefaultStreamUsageFunc(chunk openai.ChatCompletionChunk, _ map[string]any, metadata fantasy.ProviderMetadata) (fantasy.Usage, fantasy.ProviderMetadata) {
225 if chunk.Usage.TotalTokens == 0 {
226 return fantasy.Usage{}, nil
227 }
228 streamProviderMetadata := &ProviderMetadata{}
229 if metadata != nil {
230 if providerMetadata, ok := metadata[Name]; ok {
231 converted, ok := providerMetadata.(*ProviderMetadata)
232 if ok {
233 streamProviderMetadata = converted
234 }
235 }
236 }
237 // we do this here because the acc does not add prompt details
238 completionTokenDetails := chunk.Usage.CompletionTokensDetails
239 promptTokenDetails := chunk.Usage.PromptTokensDetails
240 usage := fantasy.Usage{
241 InputTokens: chunk.Usage.PromptTokens,
242 OutputTokens: chunk.Usage.CompletionTokens,
243 TotalTokens: chunk.Usage.TotalTokens,
244 ReasoningTokens: completionTokenDetails.ReasoningTokens,
245 CacheReadTokens: promptTokenDetails.CachedTokens,
246 }
247
248 // Add prediction tokens if available
249 if completionTokenDetails.AcceptedPredictionTokens > 0 || completionTokenDetails.RejectedPredictionTokens > 0 {
250 if completionTokenDetails.AcceptedPredictionTokens > 0 {
251 streamProviderMetadata.AcceptedPredictionTokens = completionTokenDetails.AcceptedPredictionTokens
252 }
253 if completionTokenDetails.RejectedPredictionTokens > 0 {
254 streamProviderMetadata.RejectedPredictionTokens = completionTokenDetails.RejectedPredictionTokens
255 }
256 }
257
258 return usage, fantasy.ProviderMetadata{
259 Name: streamProviderMetadata,
260 }
261}
262
263// DefaultStreamProviderMetadataFunc is the default implementation for handling stream provider metadata.
264func DefaultStreamProviderMetadataFunc(choice openai.ChatCompletionChoice, metadata fantasy.ProviderMetadata) fantasy.ProviderMetadata {
265 if metadata == nil {
266 metadata = fantasy.ProviderMetadata{}
267 }
268 streamProviderMetadata, ok := metadata[Name]
269 if !ok {
270 streamProviderMetadata = &ProviderMetadata{}
271 }
272 if converted, ok := streamProviderMetadata.(*ProviderMetadata); ok {
273 converted.Logprobs = choice.Logprobs.Content
274 metadata[Name] = converted
275 }
276 return metadata
277}
278
279// DefaultToPrompt converts a fantasy prompt to OpenAI format with default handling.
280func DefaultToPrompt(prompt fantasy.Prompt, _, _ string) ([]openai.ChatCompletionMessageParamUnion, []fantasy.CallWarning) {
281 var messages []openai.ChatCompletionMessageParamUnion
282 var warnings []fantasy.CallWarning
283 for _, msg := range prompt {
284 switch msg.Role {
285 case fantasy.MessageRoleSystem:
286 var systemPromptParts []string
287 for _, c := range msg.Content {
288 if c.GetType() != fantasy.ContentTypeText {
289 warnings = append(warnings, fantasy.CallWarning{
290 Type: fantasy.CallWarningTypeOther,
291 Message: "system prompt can only have text content",
292 })
293 continue
294 }
295 textPart, ok := fantasy.AsContentType[fantasy.TextPart](c)
296 if !ok {
297 warnings = append(warnings, fantasy.CallWarning{
298 Type: fantasy.CallWarningTypeOther,
299 Message: "system prompt text part does not have the right type",
300 })
301 continue
302 }
303 text := textPart.Text
304 if strings.TrimSpace(text) != "" {
305 systemPromptParts = append(systemPromptParts, textPart.Text)
306 }
307 }
308 if len(systemPromptParts) == 0 {
309 warnings = append(warnings, fantasy.CallWarning{
310 Type: fantasy.CallWarningTypeOther,
311 Message: "system prompt has no text parts",
312 })
313 continue
314 }
315 messages = append(messages, openai.SystemMessage(strings.Join(systemPromptParts, "\n")))
316 case fantasy.MessageRoleUser:
317 // simple user message just text content
318 if len(msg.Content) == 1 && msg.Content[0].GetType() == fantasy.ContentTypeText {
319 textPart, ok := fantasy.AsContentType[fantasy.TextPart](msg.Content[0])
320 if !ok {
321 warnings = append(warnings, fantasy.CallWarning{
322 Type: fantasy.CallWarningTypeOther,
323 Message: "user message text part does not have the right type",
324 })
325 continue
326 }
327 messages = append(messages, openai.UserMessage(textPart.Text))
328 continue
329 }
330 // text content and attachments
331 // for now we only support image content later we need to check
332 // TODO: add the supported media types to the language model so we
333 // can use that to validate the data here.
334 var content []openai.ChatCompletionContentPartUnionParam
335 for _, c := range msg.Content {
336 switch c.GetType() {
337 case fantasy.ContentTypeText:
338 textPart, ok := fantasy.AsContentType[fantasy.TextPart](c)
339 if !ok {
340 warnings = append(warnings, fantasy.CallWarning{
341 Type: fantasy.CallWarningTypeOther,
342 Message: "user message text part does not have the right type",
343 })
344 continue
345 }
346 content = append(content, openai.ChatCompletionContentPartUnionParam{
347 OfText: &openai.ChatCompletionContentPartTextParam{
348 Text: textPart.Text,
349 },
350 })
351 case fantasy.ContentTypeFile:
352 filePart, ok := fantasy.AsContentType[fantasy.FilePart](c)
353 if !ok {
354 warnings = append(warnings, fantasy.CallWarning{
355 Type: fantasy.CallWarningTypeOther,
356 Message: "user message file part does not have the right type",
357 })
358 continue
359 }
360
361 switch {
362 case strings.HasPrefix(filePart.MediaType, "text/"):
363 base64Encoded := base64.StdEncoding.EncodeToString(filePart.Data)
364 content = append(content, openai.FileContentPart(openai.ChatCompletionContentPartFileFileParam{
365 FileData: param.NewOpt(base64Encoded),
366 }))
367 case strings.HasPrefix(filePart.MediaType, "image/"):
368 // Handle image files
369 base64Encoded := base64.StdEncoding.EncodeToString(filePart.Data)
370 data := "data:" + filePart.MediaType + ";base64," + base64Encoded
371 imageURL := openai.ChatCompletionContentPartImageImageURLParam{URL: data}
372
373 // Check for provider-specific options like image detail
374 if providerOptions, ok := filePart.ProviderOptions[Name]; ok {
375 if detail, ok := providerOptions.(*ProviderFileOptions); ok {
376 imageURL.Detail = detail.ImageDetail
377 }
378 }
379
380 imageBlock := openai.ChatCompletionContentPartImageParam{ImageURL: imageURL}
381 content = append(content, openai.ChatCompletionContentPartUnionParam{OfImageURL: &imageBlock})
382
383 case filePart.MediaType == "audio/wav":
384 // Handle WAV audio files
385 base64Encoded := base64.StdEncoding.EncodeToString(filePart.Data)
386 audioBlock := openai.ChatCompletionContentPartInputAudioParam{
387 InputAudio: openai.ChatCompletionContentPartInputAudioInputAudioParam{
388 Data: base64Encoded,
389 Format: "wav",
390 },
391 }
392 content = append(content, openai.ChatCompletionContentPartUnionParam{OfInputAudio: &audioBlock})
393
394 case filePart.MediaType == "audio/mpeg" || filePart.MediaType == "audio/mp3":
395 // Handle MP3 audio files
396 base64Encoded := base64.StdEncoding.EncodeToString(filePart.Data)
397 audioBlock := openai.ChatCompletionContentPartInputAudioParam{
398 InputAudio: openai.ChatCompletionContentPartInputAudioInputAudioParam{
399 Data: base64Encoded,
400 Format: "mp3",
401 },
402 }
403 content = append(content, openai.ChatCompletionContentPartUnionParam{OfInputAudio: &audioBlock})
404
405 case filePart.MediaType == "application/pdf":
406 // Handle PDF files
407 dataStr := string(filePart.Data)
408
409 // Check if data looks like a file ID (starts with "file-")
410 if strings.HasPrefix(dataStr, "file-") {
411 fileBlock := openai.ChatCompletionContentPartFileParam{
412 File: openai.ChatCompletionContentPartFileFileParam{
413 FileID: param.NewOpt(dataStr),
414 },
415 }
416 content = append(content, openai.ChatCompletionContentPartUnionParam{OfFile: &fileBlock})
417 } else {
418 // Handle as base64 data
419 base64Encoded := base64.StdEncoding.EncodeToString(filePart.Data)
420 data := "data:application/pdf;base64," + base64Encoded
421
422 filename := filePart.Filename
423 if filename == "" {
424 // Generate default filename based on content index
425 filename = fmt.Sprintf("part-%d.pdf", len(content))
426 }
427
428 fileBlock := openai.ChatCompletionContentPartFileParam{
429 File: openai.ChatCompletionContentPartFileFileParam{
430 Filename: param.NewOpt(filename),
431 FileData: param.NewOpt(data),
432 },
433 }
434 content = append(content, openai.ChatCompletionContentPartUnionParam{OfFile: &fileBlock})
435 }
436
437 default:
438 warnings = append(warnings, fantasy.CallWarning{
439 Type: fantasy.CallWarningTypeOther,
440 Message: fmt.Sprintf("file part media type %s not supported", filePart.MediaType),
441 })
442 }
443 }
444 }
445 messages = append(messages, openai.UserMessage(content))
446 case fantasy.MessageRoleAssistant:
447 // simple assistant message just text content
448 if len(msg.Content) == 1 && msg.Content[0].GetType() == fantasy.ContentTypeText {
449 textPart, ok := fantasy.AsContentType[fantasy.TextPart](msg.Content[0])
450 if !ok {
451 warnings = append(warnings, fantasy.CallWarning{
452 Type: fantasy.CallWarningTypeOther,
453 Message: "assistant message text part does not have the right type",
454 })
455 continue
456 }
457 messages = append(messages, openai.AssistantMessage(textPart.Text))
458 continue
459 }
460 assistantMsg := openai.ChatCompletionAssistantMessageParam{
461 Role: "assistant",
462 }
463 for _, c := range msg.Content {
464 switch c.GetType() {
465 case fantasy.ContentTypeText:
466 textPart, ok := fantasy.AsContentType[fantasy.TextPart](c)
467 if !ok {
468 warnings = append(warnings, fantasy.CallWarning{
469 Type: fantasy.CallWarningTypeOther,
470 Message: "assistant message text part does not have the right type",
471 })
472 continue
473 }
474 assistantMsg.Content = openai.ChatCompletionAssistantMessageParamContentUnion{
475 OfString: param.NewOpt(textPart.Text),
476 }
477 case fantasy.ContentTypeToolCall:
478 toolCallPart, ok := fantasy.AsContentType[fantasy.ToolCallPart](c)
479 if !ok {
480 warnings = append(warnings, fantasy.CallWarning{
481 Type: fantasy.CallWarningTypeOther,
482 Message: "assistant message tool part does not have the right type",
483 })
484 continue
485 }
486 assistantMsg.ToolCalls = append(assistantMsg.ToolCalls,
487 openai.ChatCompletionMessageToolCallUnionParam{
488 OfFunction: &openai.ChatCompletionMessageFunctionToolCallParam{
489 ID: toolCallPart.ToolCallID,
490 Type: "function",
491 Function: openai.ChatCompletionMessageFunctionToolCallFunctionParam{
492 Name: toolCallPart.ToolName,
493 Arguments: toolCallPart.Input,
494 },
495 },
496 })
497 }
498 }
499 messages = append(messages, openai.ChatCompletionMessageParamUnion{
500 OfAssistant: &assistantMsg,
501 })
502 case fantasy.MessageRoleTool:
503 for _, c := range msg.Content {
504 if c.GetType() != fantasy.ContentTypeToolResult {
505 warnings = append(warnings, fantasy.CallWarning{
506 Type: fantasy.CallWarningTypeOther,
507 Message: "tool message can only have tool result content",
508 })
509 continue
510 }
511
512 toolResultPart, ok := fantasy.AsContentType[fantasy.ToolResultPart](c)
513 if !ok {
514 warnings = append(warnings, fantasy.CallWarning{
515 Type: fantasy.CallWarningTypeOther,
516 Message: "tool message result part does not have the right type",
517 })
518 continue
519 }
520
521 switch toolResultPart.Output.GetType() {
522 case fantasy.ToolResultContentTypeText:
523 output, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentText](toolResultPart.Output)
524 if !ok {
525 warnings = append(warnings, fantasy.CallWarning{
526 Type: fantasy.CallWarningTypeOther,
527 Message: "tool result output does not have the right type",
528 })
529 continue
530 }
531 messages = append(messages, openai.ToolMessage(output.Text, toolResultPart.ToolCallID))
532 case fantasy.ToolResultContentTypeError:
533 // TODO: check if better handling is needed
534 output, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentError](toolResultPart.Output)
535 if !ok {
536 warnings = append(warnings, fantasy.CallWarning{
537 Type: fantasy.CallWarningTypeOther,
538 Message: "tool result output does not have the right type",
539 })
540 continue
541 }
542 messages = append(messages, openai.ToolMessage(output.Error.Error(), toolResultPart.ToolCallID))
543 }
544 }
545 }
546 }
547 return messages, warnings
548}