reduce api calls for text extraction
This commit is contained in:
@@ -554,6 +554,24 @@ export async function generateItemDescription(itemKey: string): Promise<string>
|
||||
* If the extracted text is not in the user's preferred language, auto-translates it.
|
||||
* Returns { extractedText, translatedText }.
|
||||
*/
|
||||
/**
|
||||
* Parse a structured extraction response from the AI.
|
||||
* Returns null if the response cannot be parsed as valid JSON with the expected shape.
|
||||
*/
|
||||
function parseStructuredExtraction(raw: string): { text: string; needsTranslation: boolean } | null {
|
||||
const jsonMatch = raw.match(/\{[\s\S]*\}/)
|
||||
if (!jsonMatch) return null
|
||||
try {
|
||||
const parsed = JSON.parse(jsonMatch[0])
|
||||
if (typeof parsed.text === 'string' && typeof parsed.needsTranslation === 'boolean') {
|
||||
return { text: parsed.text, needsTranslation: parsed.needsTranslation }
|
||||
}
|
||||
} catch {
|
||||
// fall through
|
||||
}
|
||||
return null
|
||||
}
|
||||
|
||||
export async function extractItemText(itemKey: string): Promise<{ extractedText: string; translatedText: string | null }> {
|
||||
const libraryId = itemKey.split(':')[0]
|
||||
const config = getEffectiveAiConfig(libraryId)
|
||||
@@ -590,9 +608,46 @@ export async function extractItemText(itemKey: string): Promise<{ extractedText:
|
||||
const thumbnailPath = await getAiImagePath(resolvedMedia.path, libraryId)
|
||||
const base64Images = [fs.readFileSync(thumbnailPath, 'base64')]
|
||||
|
||||
const systemPrompt = `You are an OCR assistant. Extract ALL text visible in the image exactly as it appears. Preserve line breaks and formatting.${config.promptExtract ? ' ' + config.promptExtract : ''} If there is no text in the image, respond with exactly: [NO TEXT]`
|
||||
const preferredLanguage = getPreferredLanguage()
|
||||
const customInstruction = config.promptExtract ? ' ' + config.promptExtract : ''
|
||||
|
||||
const extractedText = await callVisionApiText(config.endpoint, extractModel, base64Images, systemPrompt)
|
||||
// When a preferred language is configured, ask the AI to also flag whether translation is needed.
|
||||
// This avoids a separate translation API call for text already in the target language.
|
||||
let systemPrompt: string
|
||||
if (preferredLanguage) {
|
||||
systemPrompt = `You are an OCR assistant. Extract ALL text visible in the image exactly as it appears. Preserve line breaks and formatting.${customInstruction}
|
||||
|
||||
Respond ONLY with a valid JSON object — no markdown, no explanation:
|
||||
{"needsTranslation": boolean, "text": "extracted text"}
|
||||
|
||||
Rules:
|
||||
- Set needsTranslation to true if the text is NOT already written in ${preferredLanguage}.
|
||||
- Set needsTranslation to false if the text IS in ${preferredLanguage}, or if there is no text.
|
||||
- If there is no text in the image, use exactly: {"needsTranslation": false, "text": "[NO TEXT]"}`
|
||||
} else {
|
||||
systemPrompt = `You are an OCR assistant. Extract ALL text visible in the image exactly as it appears. Preserve line breaks and formatting.${customInstruction} If there is no text in the image, respond with exactly: [NO TEXT]`
|
||||
}
|
||||
|
||||
const rawResponse = await callVisionApiText(config.endpoint, extractModel, base64Images, systemPrompt)
|
||||
|
||||
// Parse the response — structured JSON when a preferred language is set, plain text otherwise
|
||||
let extractedText: string
|
||||
let needsTranslation: boolean
|
||||
|
||||
if (preferredLanguage) {
|
||||
const parsed = parseStructuredExtraction(rawResponse)
|
||||
if (parsed) {
|
||||
extractedText = parsed.text
|
||||
needsTranslation = parsed.needsTranslation
|
||||
} else {
|
||||
// Malformed JSON fallback: treat raw response as plain text and attempt translation
|
||||
extractedText = rawResponse
|
||||
needsTranslation = true
|
||||
}
|
||||
} else {
|
||||
extractedText = rawResponse
|
||||
needsTranslation = false
|
||||
}
|
||||
|
||||
if (!extractedText || extractedText === '[NO TEXT]') {
|
||||
db.prepare('UPDATE media_items SET extracted_text = NULL, extracted_text_translated = NULL WHERE item_key = ?').run(itemKey)
|
||||
@@ -601,10 +656,9 @@ export async function extractItemText(itemKey: string): Promise<{ extractedText:
|
||||
|
||||
db.prepare('UPDATE media_items SET extracted_text = ? WHERE item_key = ?').run(extractedText, itemKey)
|
||||
|
||||
// Auto-translate if preferred language is set
|
||||
const preferredLanguage = getPreferredLanguage()
|
||||
// Only translate if the extraction step determined the text is not already in the preferred language
|
||||
let translatedText: string | null = null
|
||||
if (preferredLanguage) {
|
||||
if (preferredLanguage && needsTranslation) {
|
||||
const translateModel = config.modelTranslate || config.model
|
||||
try {
|
||||
translatedText = await translateText(config.endpoint, translateModel, extractedText, preferredLanguage, config.promptTranslate)
|
||||
|
||||
Reference in New Issue
Block a user