transl8-srt 1.0.1 → 1.0.2

package/dist/index.d.cts

@@ -39,8 +39,20 @@ interface TranslateOptions {
     batchSize?: number;
     /** Context overlap: N sentences from the prev batch prepended (default 4) */
     contextOverlap?: number;
-    /** OpenAI model override. Defaults to gpt-4o for non-Latin scripts, gpt-4o-mini otherwise */
-    model?: "gpt-4o" | "gpt-4o-mini";
+    /**
+     * Model for translation batches.
+     * OpenAI users: defaults to "gpt-4o" for complex scripts, "gpt-4o-mini" otherwise.
+     * Custom client users (DeepSeek, Groq, etc.): set this to your provider's model name,
+     * e.g. "deepseek-chat". Auto-selection only works for OpenAI model names.
+     */
+    model?: "gpt-4o" | "gpt-4o-mini" | (string & {});
+    /**
+     * Fallback model used when auto-selection would normally pick gpt-4o-mini
+     * but you're on a custom provider. Only needed if you want different models
+     * for different language complexity tiers on a non-OpenAI backend.
+     * Most custom client users should just set `model` and leave this unset.
+     */
+    modelFallback?: string;
     /** Max concurrent batch requests (default 8) */
     concurrency?: number;
     /** Enable AI primer call — for general/unknown content only */
@@ -89,7 +101,12 @@ declare class SubtitleTranslationError extends Error {
 declare function translate(
 /** Raw SRT or VTT file content */
 fileContent: string, opts: TranslateOptions, 
-/** Pass your own OpenAI client (allows custom baseURL, auth, etc.) */
+/**
+ * Pass your own OpenAI-compatible client.
+ * For DeepSeek: new OpenAI({ baseURL: "https://api.deepseek.com", apiKey: "..." })
+ * For Groq:     new OpenAI({ baseURL: "https://api.groq.com/openai/v1", apiKey: "..." })
+ * Remember to also set opts.model to your provider's model name.
+ */
 client?: OpenAI): Promise<TranslationResult>;
 declare function translateFile(inputPath: string, outputPath: string, opts: TranslateOptions, client?: OpenAI): Promise<TranslationResult>;
 

package/dist/index.d.ts

@@ -39,8 +39,20 @@ interface TranslateOptions {
     batchSize?: number;
     /** Context overlap: N sentences from the prev batch prepended (default 4) */
     contextOverlap?: number;
-    /** OpenAI model override. Defaults to gpt-4o for non-Latin scripts, gpt-4o-mini otherwise */
-    model?: "gpt-4o" | "gpt-4o-mini";
+    /**
+     * Model for translation batches.
+     * OpenAI users: defaults to "gpt-4o" for complex scripts, "gpt-4o-mini" otherwise.
+     * Custom client users (DeepSeek, Groq, etc.): set this to your provider's model name,
+     * e.g. "deepseek-chat". Auto-selection only works for OpenAI model names.
+     */
+    model?: "gpt-4o" | "gpt-4o-mini" | (string & {});
+    /**
+     * Fallback model used when auto-selection would normally pick gpt-4o-mini
+     * but you're on a custom provider. Only needed if you want different models
+     * for different language complexity tiers on a non-OpenAI backend.
+     * Most custom client users should just set `model` and leave this unset.
+     */
+    modelFallback?: string;
     /** Max concurrent batch requests (default 8) */
     concurrency?: number;
     /** Enable AI primer call — for general/unknown content only */
@@ -89,7 +101,12 @@ declare class SubtitleTranslationError extends Error {
 declare function translate(
 /** Raw SRT or VTT file content */
 fileContent: string, opts: TranslateOptions, 
-/** Pass your own OpenAI client (allows custom baseURL, auth, etc.) */
+/**
+ * Pass your own OpenAI-compatible client.
+ * For DeepSeek: new OpenAI({ baseURL: "https://api.deepseek.com", apiKey: "..." })
+ * For Groq:     new OpenAI({ baseURL: "https://api.groq.com/openai/v1", apiKey: "..." })
+ * Remember to also set opts.model to your provider's model name.
+ */
 client?: OpenAI): Promise<TranslationResult>;
 declare function translateFile(inputPath: string, outputPath: string, opts: TranslateOptions, client?: OpenAI): Promise<TranslationResult>;
 

package/dist/index.js

@@ -7126,11 +7126,27 @@ Do not explain metaphors or idioms \u2014 translate them faithfully.`,
 Preserve the speaker's natural register and sentence rhythm.
 Do not add formality that isn't in the source.`
 };
-var NON_LATIN_SCRIPTS = /* @__PURE__ */ new Set(["ar", "am", "zh", "ja", "ko", "hi", "ru", "uk", "el"]);
-function pickModel(targetLang, override) {
-  if (override) return override;
-  const needsFullModel = NON_LATIN_SCRIPTS.has(targetLang) || ["yo", "ig", "ha", "sw", "zu"].includes(targetLang);
-  return needsFullModel ? "gpt-4o" : "gpt-4o-mini";
+var NEEDS_FULL_MODEL = /* @__PURE__ */ new Set([
+  "ar",
+  "am",
+  "zh",
+  "ja",
+  "ko",
+  "hi",
+  "ru",
+  "uk",
+  "el",
+  "yo",
+  "ig",
+  "ha",
+  "sw",
+  "zu"
+]);
+function pickModel(opts) {
+  if (opts.model) return opts.model;
+  const needsFull = NEEDS_FULL_MODEL.has(opts.targetLang);
+  if (needsFull) return "gpt-4o";
+  return opts.modelFallback ?? "gpt-4o-mini";
 }
 function buildGlossaryBlock(meta) {
   const lines = [];
@@ -7145,7 +7161,7 @@ function buildGlossaryBlock(meta) {
   }
   if (meta.glossary) {
     for (const [src, tgt] of Object.entries(meta.glossary)) {
-      lines.push(`"${src}" \u2192 "${tgt}"`);
+      lines.push(`"${src}" -> "${tgt}"`);
     }
   }
   return lines.length ? `GLOSSARY (must follow exactly):
@@ -7161,7 +7177,7 @@ function buildSystemPrompt(opts) {
     domainBlock,
     glossaryBlock,
     `RULES:
-- Preserve ALL-CAPS emphasis, ellipses (...), dashes (\u2014), and line-ending punctuation exactly.
+- Preserve ALL-CAPS emphasis, ellipses (...), dashes, and line-ending punctuation exactly.
 - Never merge or split sentences \u2014 one input item = one output item.
 - Never add explanatory text, parentheticals, or translator notes.
 - Output ONLY valid JSON. No markdown, no code fences, no preamble.`,
@@ -7276,7 +7292,7 @@ async function executeTranslation(client, sentences, opts) {
   const batchSize = opts.batchSize ?? 25;
   const overlap = opts.contextOverlap ?? 4;
   const concurrency = opts.concurrency ?? 8;
-  const model = pickModel(opts.targetLang, opts.model);
+  const model = pickModel(opts);
   const systemPrompt = buildSystemPrompt(opts);
   const payloads = buildBatches(sentences, batchSize, overlap);
   const tasks = payloads.map(
@@ -7389,12 +7405,13 @@ function emitVTT(cues) {
 // src/index.ts
 import { readFile, writeFile } from "fs/promises";
 import { extname } from "path";
-var COST_PER_M_TOKENS = {
+var KNOWN_COSTS = {
   "gpt-4o": { input: 2.5, output: 10 },
   "gpt-4o-mini": { input: 0.15, output: 0.6 }
 };
 function estimateCost(model, inputTokens, outputTokens) {
-  const rates = COST_PER_M_TOKENS[model];
+  const rates = KNOWN_COSTS[model];
+  if (!rates) return 0;
   return inputTokens / 1e6 * rates.input + outputTokens / 1e6 * rates.output;
 }
 async function translate(fileContent, opts, client) {
@@ -7405,14 +7422,14 @@ async function translate(fileContent, opts, client) {
   const sentences = reconstructSentences(cues);
   const { translations, totalInputTokens, totalOutputTokens, totalBatches } = await executeTranslation(oi, sentences, opts);
   const translatedCues = reanchorCues(cues, sentences, translations);
-  const model = opts.model ?? (["yo", "ig", "ha", "sw", "zu", "ar", "am"].includes(opts.targetLang) ? "gpt-4o" : "gpt-4o-mini");
+  const resolvedModel = pickModel(opts);
   const stats = {
     totalCues: cues.length,
     totalSentences: sentences.length,
     totalBatches,
     inputTokens: totalInputTokens,
     outputTokens: totalOutputTokens,
-    estimatedCostUsd: estimateCost(model, totalInputTokens, totalOutputTokens),
+    estimatedCostUsd: estimateCost(resolvedModel, totalInputTokens, totalOutputTokens),
     durationMs: Date.now() - startTime
   };
   return {