vetrag/openai_client.go

325 lines
9.9 KiB
Go

package main
import (
"bytes"
"context"
"encoding/json"
"fmt"
"io"
"net/http"
"strings"
"time"
"github.com/sirupsen/logrus"
)
// --- OpenAIClient implementation ---
type OpenAIClient struct {
APIKey string
BaseURL string
Model string
Repo ChatRepositoryAPI
}
func NewOpenAIClient(apiKey, baseURL, model string, repo ChatRepositoryAPI) *OpenAIClient {
return &OpenAIClient{APIKey: apiKey, BaseURL: baseURL, Model: model, Repo: repo}
}
func (llm *OpenAIClient) ExtractKeywords(ctx context.Context, message string) (map[string]interface{}, error) {
_, parsed, err := llm.ExtractKeywordsRaw(ctx, message)
return parsed, err
}
func (llm *OpenAIClient) ExtractKeywordsRaw(ctx context.Context, message string) (string, map[string]interface{}, error) {
prompt, err := renderPrompt(appConfig.LLM.ExtractKeywordsPrompt, map[string]string{"Message": message})
if err != nil {
logrus.WithError(err).Error("[CONFIG] Failed to render ExtractKeywords prompt")
return "", nil, err
}
logrus.WithField("prompt", prompt).Info("[LLM] ExtractKeywords prompt")
// Use the utility function instead of inline format definition
format := GetExtractKeywordsFormat()
resp, err := llm.openAICompletion(ctx, prompt, format)
logrus.WithFields(logrus.Fields{"response": resp, "err": err}).Info("[LLM] ExtractKeywords response")
if err != nil {
return resp, nil, err
}
var result map[string]interface{}
if err := json.Unmarshal([]byte(resp), &result); err != nil {
return resp, nil, err
}
return resp, result, nil
}
func (llm *OpenAIClient) DisambiguateBestMatch(ctx context.Context, message string, candidates []Visit) (string, error) {
_, vr, err := llm.DisambiguateBestMatchRaw(ctx, message, candidates)
return vr, err
}
func (llm *OpenAIClient) DisambiguateBestMatchRaw(ctx context.Context, message string, candidates []Visit) (string, string, error) {
// Use the utility function instead of inline format definition
format := GetDisambiguateFormat()
entries, _ := json.Marshal(candidates)
prompt, err := renderPrompt(appConfig.LLM.DisambiguatePrompt, map[string]string{"Entries": string(entries), "Message": message})
if err != nil {
logrus.WithError(err).Error("[CONFIG] Failed to render Disambiguate prompt")
return "", "", err
}
logrus.WithField("prompt", prompt).Info("[LLM] DisambiguateBestMatch prompt")
resp, err := llm.openAICompletion(ctx, prompt, format)
logrus.WithFields(logrus.Fields{"response": resp, "err": err}).Info("[LLM] DisambiguateBestMatch response")
if err != nil {
return resp, "", err
}
var parsed map[string]string
if err := json.Unmarshal([]byte(resp), &parsed); err != nil {
return resp, "", fmt.Errorf("failed to unmarshal disambiguation response: %w", err)
}
visitReason := strings.TrimSpace(parsed["visitReason"])
if visitReason == "" {
return resp, "", fmt.Errorf("visitReason not found in response")
}
return resp, visitReason, nil
}
func (llm *OpenAIClient) openAICompletion(ctx context.Context, prompt string, format map[string]interface{}) (string, error) {
apiURL := llm.BaseURL
if apiURL == "" {
apiURL = "https://api.openai.com/v1/chat/completions"
}
isOpenAIStyle := strings.Contains(apiURL, "openrouter.ai") || strings.Contains(apiURL, "/v1/")
// Helper to stringify the expected JSON schema for instructions
schemaDesc := func() string {
b, _ := json.MarshalIndent(format, "", " ")
return string(b)
}
truncate := func(s string, n int) string {
if len(s) <= n {
return s
}
return s[:n] + "...<truncated>"
}
buildBody := func() map[string]interface{} {
if isOpenAIStyle {
return map[string]interface{}{
"model": llm.Model,
"messages": []map[string]string{
{"role": "system", "content": "You are a strict JSON generator. ONLY output valid JSON matching this schema: " + schemaDesc() + " Do not add explanations."},
{"role": "user", "content": prompt},
},
"response_format": map[string]interface{}{"type": "json_object"},
}
}
// This should never be reached in OpenAI client but keeping for safety
return map[string]interface{}{
"model": llm.Model,
"messages": []map[string]string{{"role": "user", "content": prompt}},
"stream": false,
"format": format,
}
}
body := buildBody()
// Enhanced logging similar to the unified client
jsonBody, _ := json.Marshal(body)
bodySize := len(jsonBody)
logrus.WithFields(logrus.Fields{
"event": "llm_request",
"api_url": apiURL,
"model": llm.Model,
"is_openai_style": isOpenAIStyle,
"prompt_len": len(prompt),
"body_size": bodySize,
}).Info("[LLM] sending request")
req, _ := http.NewRequestWithContext(ctx, http.MethodPost, apiURL, bytes.NewBuffer(jsonBody))
if llm.APIKey != "" {
req.Header.Set("Authorization", "Bearer "+llm.APIKey)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Accept", "application/json")
if strings.Contains(apiURL, "openrouter.ai") {
req.Header.Set("Referer", "https://github.com/")
req.Header.Set("X-Title", "vetrag-app")
}
start := time.Now()
client := &http.Client{}
resp, err := client.Do(req)
dur := time.Since(start)
if err != nil {
logrus.WithFields(logrus.Fields{
"event": "llm_response",
"status": 0,
"latency_ms": dur.Milliseconds(),
"error": err,
}).Error("[LLM] request failed")
return "", err
}
defer resp.Body.Close()
raw, err := io.ReadAll(resp.Body)
if err != nil {
return "", err
}
logrus.WithFields(logrus.Fields{
"event": "llm_raw_response",
"status": resp.StatusCode,
"latency_ms": dur.Milliseconds(),
"raw_trunc": truncate(string(raw), 600),
"raw_len": len(raw),
}).Debug("[LLM] raw response body")
parseVariant := "unknown"
// Attempt OpenAI/OpenRouter style parse first
var openAI struct {
Choices []struct {
Message struct {
Content string `json:"content"`
} `json:"message"`
} `json:"choices"`
Error *struct {
Message string `json:"message"`
Type string `json:"type"`
} `json:"error"`
}
if err := json.Unmarshal(raw, &openAI); err == nil {
if openAI.Error != nil || resp.StatusCode >= 400 {
parseVariant = "openai"
var msg string
if openAI.Error != nil {
msg = openAI.Error.Message
} else {
msg = string(raw)
}
logrus.WithFields(logrus.Fields{
"event": "llm_response",
"status": resp.StatusCode,
"latency_ms": dur.Milliseconds(),
"parse_variant": parseVariant,
"error": msg,
}).Error("[LLM] provider error")
return "", fmt.Errorf("provider error: %s", msg)
}
if len(openAI.Choices) > 0 && openAI.Choices[0].Message.Content != "" {
parseVariant = "openai"
content := openAI.Choices[0].Message.Content
logrus.WithFields(logrus.Fields{
"event": "llm_response",
"status": resp.StatusCode,
"latency_ms": dur.Milliseconds(),
"parse_variant": parseVariant,
"content_len": len(content),
"content_snip": truncate(content, 300),
}).Info("[LLM] parsed response")
return content, nil
}
}
// As a fallback, attempt Ollama format parse
var ollama struct {
Message struct {
Content string `json:"content"`
} `json:"message"`
Error string `json:"error"`
}
if err := json.Unmarshal(raw, &ollama); err == nil && ollama.Message.Content != "" {
parseVariant = "ollama"
content := ollama.Message.Content
logrus.WithFields(logrus.Fields{
"event": "llm_response",
"status": resp.StatusCode,
"latency_ms": dur.Milliseconds(),
"parse_variant": parseVariant,
"content_len": len(content),
"content_snip": truncate(content, 300),
}).Info("[LLM] parsed response")
return content, nil
}
logrus.WithFields(logrus.Fields{
"event": "llm_response",
"status": resp.StatusCode,
"latency_ms": dur.Milliseconds(),
"parse_variant": parseVariant,
"raw_snip": truncate(string(raw), 300),
}).Error("[LLM] unrecognized response format")
return "", fmt.Errorf("unrecognized LLM response format: %.200s", string(raw))
}
func (llm *OpenAIClient) GetEmbeddings(ctx context.Context, input string) ([]float64, error) {
apiURL := llm.BaseURL
if apiURL == "" {
apiURL = "https://api.openai.com/v1/embeddings"
}
body := map[string]interface{}{
"model": llm.Model,
"input": input,
}
jsonBody, _ := json.Marshal(body)
req, _ := http.NewRequestWithContext(ctx, http.MethodPost, apiURL, bytes.NewBuffer(jsonBody))
if llm.APIKey != "" {
req.Header.Set("Authorization", "Bearer "+llm.APIKey)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Accept", "application/json")
if strings.Contains(apiURL, "openrouter.ai") {
req.Header.Set("Referer", "https://github.com/")
req.Header.Set("X-Title", "vetrag-app")
}
client := &http.Client{}
resp, err := client.Do(req)
if err != nil {
return nil, err
}
defer resp.Body.Close()
raw, err := io.ReadAll(resp.Body)
if err != nil {
return nil, err
}
var openAI struct {
Data []struct {
Embedding []float64 `json:"embedding"`
} `json:"data"`
Error *struct {
Message string `json:"message"`
} `json:"error"`
}
if err := json.Unmarshal(raw, &openAI); err == nil && len(openAI.Data) > 0 {
return openAI.Data[0].Embedding, nil
}
if openAI.Error != nil {
return nil, fmt.Errorf("embedding error: %s", openAI.Error.Message)
}
return nil, fmt.Errorf("unrecognized embedding response: %.200s", string(raw))
}
func (llm *OpenAIClient) TranslateToEnglish(ctx context.Context, message string) (string, error) {
prompt, err := renderPrompt(appConfig.LLM.TranslatePrompt, map[string]string{"Message": message})
if err != nil {
logrus.WithError(err).Error("[CONFIG] Failed to render Translate prompt")
return "", err
}
logrus.WithField("prompt", prompt).Info("[LLM] TranslateToEnglish prompt")
resp, err := llm.openAICompletion(ctx, prompt, nil)
logrus.WithFields(logrus.Fields{"response": resp, "err": err}).Info("[LLM] TranslateToEnglish response")
if err != nil {
return resp, err
}
return strings.TrimSpace(resp), nil
}