Jānis Kacēns
120 lines
3.3 KiB
Go
120 lines
3.3 KiB
Go
package llm
|
|
|
|
import (
|
|
"context"
|
|
"crypto/sha256"
|
|
"encoding/json"
|
|
"fmt"
|
|
|
|
openai "github.com/sashabaranov/go-openai"
|
|
)
|
|
|
|
// ParsedQuestion is a question extracted from a document chunk by the LLM.
|
|
type ParsedQuestion struct {
|
|
Question string
|
|
Answers []ParsedAnswer
|
|
}
|
|
|
|
// ParsedAnswer is one answer choice for a ParsedQuestion.
|
|
type ParsedAnswer struct {
|
|
Text string
|
|
Correct bool
|
|
}
|
|
|
|
// ChatClient is the interface for creating chat completions.
|
|
// The concrete *openai.Client satisfies this interface.
|
|
type ChatClient interface {
|
|
CreateChatCompletion(ctx context.Context, req openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error)
|
|
}
|
|
|
|
// Client wraps a ChatClient with question-extraction logic.
|
|
type Client struct {
|
|
cc ChatClient
|
|
model string
|
|
}
|
|
|
|
// New creates a Client backed by the real OpenAI API.
|
|
func New(apiKey, model string) *Client {
|
|
if model == "" {
|
|
model = "gpt-4o-mini"
|
|
}
|
|
return &Client{cc: openai.NewClient(apiKey), model: model}
|
|
}
|
|
|
|
// NewWithClient creates a Client with an injected ChatClient (useful for tests).
|
|
func NewWithClient(cc ChatClient, model string) *Client {
|
|
return &Client{cc: cc, model: model}
|
|
}
|
|
|
|
const systemPrompt = `You extract multiple-choice questions from study material. Return every question found. Exactly one answer per question must be marked correct. If the source doesn't clearly mark a correct answer, omit that question entirely. Do not invent questions not present in the text.
|
|
|
|
Respond with JSON matching this schema exactly:
|
|
{"questions":[{"question":"<text>","answers":[{"text":"<text>","correct":false},{"text":"<text>","correct":true}]}]}`
|
|
|
|
type llmResponse struct {
|
|
Questions []struct {
|
|
Question string `json:"question"`
|
|
Answers []struct {
|
|
Text string `json:"text"`
|
|
Correct bool `json:"correct"`
|
|
} `json:"answers"`
|
|
} `json:"questions"`
|
|
}
|
|
|
|
// ExtractQuestions sends chunk to the LLM and returns validated, deduplicated questions.
|
|
// Questions that do not have exactly one correct answer are silently dropped.
|
|
func (c *Client) ExtractQuestions(ctx context.Context, chunk string) ([]ParsedQuestion, error) {
|
|
resp, err := c.cc.CreateChatCompletion(ctx, openai.ChatCompletionRequest{
|
|
Model: c.model,
|
|
Messages: []openai.ChatCompletionMessage{
|
|
{Role: openai.ChatMessageRoleSystem, Content: systemPrompt},
|
|
{Role: openai.ChatMessageRoleUser, Content: chunk},
|
|
},
|
|
ResponseFormat: &openai.ChatCompletionResponseFormat{
|
|
Type: openai.ChatCompletionResponseFormatTypeJSONObject,
|
|
},
|
|
})
|
|
if err != nil {
|
|
return nil, fmt.Errorf("openai: %w", err)
|
|
}
|
|
if len(resp.Choices) == 0 {
|
|
return nil, fmt.Errorf("openai: empty response")
|
|
}
|
|
|
|
var raw llmResponse
|
|
if err := json.Unmarshal([]byte(resp.Choices[0].Message.Content), &raw); err != nil {
|
|
return nil, fmt.Errorf("parse llm response: %w", err)
|
|
}
|
|
|
|
seen := make(map[string]bool)
|
|
var out []ParsedQuestion
|
|
for _, q := range raw.Questions {
|
|
var nCorrect int
|
|
for _, a := range q.Answers {
|
|
if a.Correct {
|
|
nCorrect++
|
|
}
|
|
}
|
|
if nCorrect != 1 {
|
|
continue
|
|
}
|
|
key := textHash(q.Question)
|
|
if seen[key] {
|
|
continue
|
|
}
|
|
seen[key] = true
|
|
|
|
pq := ParsedQuestion{Question: q.Question}
|
|
for _, a := range q.Answers {
|
|
pq.Answers = append(pq.Answers, ParsedAnswer{Text: a.Text, Correct: a.Correct})
|
|
}
|
|
out = append(out, pq)
|
|
}
|
|
return out, nil
|
|
}
|
|
|
|
func textHash(s string) string {
|
|
h := sha256.Sum256([]byte(s))
|
|
return fmt.Sprintf("%x", h[:8])
|
|
}
|